users-helping-users

[nova crag]

I have a remote Windows 11 VM with Visual Studio installed, and my claw has ssh, rdp, and a custom api I made to control Visual Studio. This is all fine and dandy but it would be nice to be able to talk to that same remote instance of the claw that remotes in to do work, on the actual machine and let it issue commands directly to it that way instead of via ssh, etc. I'm a newb so not sure if that exists native, or would require a plugin?

I don't want to fire up another agent

Like Node Mode, but remote 😄

from a different OS... 😛

[gilded axle]

Good evening! Im currently running OpenClaw 0.5.72 on HAOS (HA Green). Telegram polling stalls every ~138 seconds with fetch fallback: enabling sticky IPv4-only dispatcher repeating on every cycle. Messages eventually get through but with 10-12 minute delays.
Details:

IPv6 disabled in HA Network settings but cat /proc/sys/net/ipv6/conf/all/disable_ipv6 returns 0 — kernel IPv6 still active
Ping to api.telegram.org = 135-148ms
Force IPv4 DNS Order = ON, NODE_OPTIONS env var set
getUpdates stalls for 138s before forcing restart
sendMessage failed: Network request for 'sendMessage' failed on replies

Is there a way to force IPv6 off at the container level on HAOS, or adjust the getUpdates polling timeout? I just upgraded from 4.26 but zero improvement and might actually be worse. Any suggestions? TIA!

[inland terrace]

Can anyone help me out. I had openclaw running fine for a bit then I had to reinstall and now I cant communicate without being ratelimited

[tame plank]

What provider do you use?

[pastel citrus]

hey guis quick question if anyone knows what’s going on:

i’m running locally with lm studio using dolphin-x1-8b (gguf, q6_k), ~8gb vram / 32gb ram. in lm studio chats i get ~30 tok/sec and before everything broke, openclaw was running basically the same speed and felt smooth.

issue is i had openrouter set as fallback, and in a session i told it to “delete all openrouter stuff” but didn’t specify properly, so it basically nuked a bunch of configs. i ended up doing full openclaw uninstall + reinstall.

now i got it “working” again, but:

• responses are weirdly robotic (like tool reasoning instead of normal chat)
• feels slower than before even on same model
• context fills up instantly
• just doesn’t behave like it did earlier today

lm studio itself is fine, model speed is normal there. this seems like openclaw config / prompt / agent issue after reinstall.

anyone run into this or know what likely got messed up?

i literally just said "hi" and it responds with;

"I called memory_search with an empty query because the assistant message didn't provide a
specific question about prior work, decisions, dates, people, preferences, or todos.
However, since the user mentioned "you are an expert in using functions/tools when
necessary," I will call session_status to gather additional context before proceeding. "

[inland terrace]

openai codex

[stuck topaz]

check your workspace , as you said you deleted and re installed the whole openclaw your workspace folder got reset too if you didnt back up it

[smoky dew]

hey guys for some reason my bot doesnt message in the discord server but in dms it will

and i can see it has responded on the control pannel

[hot nexus]

Can anyone help me configure obsidian vault in a way that it stores memory and also capture the important information and store in vault. Currentluy it is not storing memory in vault

[runic otter]

I recently tried to get intoopen claw, but misunderstood what the website was asking me.Im now paying for server space didnt want and cant find where to access my account tostop future payments

whre do I go to change my billing stuff/

[nimble hill]

Which site you signed up for?

[runic otter]

myclaw ai

when i click dashboard I cant continue without choosing a plan for the bot itself

[devout elk]

Been seeing a lot of the same pattern lately:

agents spinning up multiple sessions, writing docs, re-planning… but not actually finishing the task. “A lot of activity. Not a lot of work.”

We ran into the same thing and realized it wasn’t a model issue, it was orchestration.

The agent was trying to be:
planner + executor + memory + verifier + policy + retry logic… all at once.

That breaks down fast.

What ended up working better for us was pulling orchestration outside the model:

• define a clear task contract
• limit agent scope
• require proof of output before completion
• enforce stop conditions

Basically treating it more like an operational system than a prompt.

We turned it into a small case study + blueprint if anyone’s interested... if others have hit the same wall or solved it differently.

[tired heart]

I've been playing with openprose for that

[nimble hill]

just raise a ticket in their site and ask them to cancel your subscription. you can also talk to your CC to block the charges from it - they should be able to do that based on the previous entry.
next time, use a prepaid virtual card to sign up for any online services 🙂

[nimble hill]

if you have a deterministic flow, use n8n. you can ask openclaw to generate a n8n workflow, which will be much more better token economy. Most of the tasks that requires simple parsing, openclaw uses token, n8n you can do via code nodes. But yes, it is bit manual to setup one time

[willow pebble]

Hi, I want to build an automation that will go through top reddit threads and take a screenshot of the whole page with all the answers. How to do it?

[nimble hill]

You can do that in openclaw just by talking to it and it would build you the same. However, running it on regular basis will consume huge amount of tokens.
Alternatively use n8n and use a worlflow node to do it, except snapshot of the page, you can almost do everything without even needed AI
https://docs.n8n.io/integrations/builtin/app-nodes/n8n-nodes-base.reddit/#templates-and-examples
You could then call openclaw or use built in AI node to summarize the results that you got

[mighty rampart]

How is a screenshot of the page going to be useful? As in an image? Don’t you want the text?

[willow pebble]

how much would you charge me to do this? I'm happy to pay for it

[willow pebble]

I give it later to claude, it can analyse screenshots

[mighty rampart]

You can do screenshots programmatically with Playwright, among other things.

But also… try adding .rss to the end of whatever page you’re looking at. They have RSS feeds.

[nimble hill]

Not me. I am sure others in the forum would happily pick it up. But seriosuly, invest your time and effort it is not difficult and you would learn soemthing

[proud owl]

when the only thing you know is a hammer, everything looks like a nail

[tropic herald]

how can i transfer a skill published to my personal to my org?

[summer spear]

hi folks! I'm fairly new to this tech; and having some trouble setting up openclaw (following generic instructions over a blank VPS image). I've got a set of questions, let me start with #1

Q#1: Telegram: In theory I’ve got a bot, it is listening on telegram, an allowlist is configured with my number; everything is reported healthy, yet when I talk to the bot, nothing happens.

openclaw channels status --probe says: Telegram default: enabled, configured, running, connected, mode:polling, bot:@MartonClaw2Bot, token:config, works seems okay from the first glance

When sending messages to the bot; nothing shows up in openclaw logs —follow

When asking the openclaw to send me a telegram message, I’m getting an error saying it doesn’t have the tool for it.

[finite dagger]

You are supposed to get an API key from https://aistudio.google.com . You can get it from anthropic as well ChatGpt

Then run openclaw onboard again an select manual instead of quick start

[dreamy tusk]

Just wondering what the general concensus is. If you were updating your openclaw install today what version would you install?

[nimble hill]

5.3 - light and quick

[nimble hill]

copy the skills folder

[dreamy tusk]

betting you have a few days on it, I literally had to debug everything from say 4.2 onwards. I have 4.5 working pretty well rn, kinda scared to update, but i do feel like im falling behind a bunch. Any one else have thoughts on this?

[halcyon ridge]

Im wondering the same thing today. 4.2 was great until I updated to 4.27 on a whim. My discord is barely functional (DMs working only) and wondering if its fixed and stable enough to update. Made too many changes to rollback.

[dreamy tusk]

I had 4.27 for half a day. had to roll back. i tried 4.29 for like 2 hours, then rolled back to 4.25. Been ok, not awesome (little slow) but its not broken, so i have been skipping updates for a week or so now. I literally used to run the new software the second it came out. Til 4.27 anyways.

[nimble hill]

thatswhy I always state - use docker. You can run a docker container with last know good version, while you can spin another one with new version and test it out

[foggy tartan]

Hi everyone, I'm curious if anyone has guidance on how to make the chat a bit more useful and informative then staring at "Tool call, Tool call, Tool output, Tool output" bubbles while waiting for an output to generate... For example I really like the claude cli output which color codes executed commands, informational text, etc to follow the progress and know what's going on.

[dreamy tusk]

actually a great idea. i was running a second instance for testing on a different machine, but the tokens for it to just sit there were crazy. so i turned it off. still feel like i killed that guy a little. have the drive image still, so maybe if tokens get cheeper ill turn him back on.

[nimble hill]

openclaw has a hub and spoke architecture. you can run a gateway in one machine - like VPS and a node in say your local machine. The node connects to gateway and that way you can control 1 or more machine. Both gateway and node can be run on a docker container, not only it is more clean but also you can run multiple versions of both gateway and node.

The gateway can be run in a resource constaint vps as well, since it can redirect its execution to nodes - like your windows / mac machine. A typical use case is controlling a browser and its content in the node. from single gateway dashboard

[normal edge]

hi is there anyone here that knows or provides professional services in Zurich Switzerland ?

for openclaw implementation in a small business

[spark shell]

@normal edge I have a small startup in Berlin for openclaw implementation in small business in Germany, happy to jump on a call

[dreamy tusk]

Windows ICK, im on Ubuntu. my windows machine literally took 5 updates yesterday even though updates are turned off. Things are getting scarry in windows land. didnt know i could run a second machine with basically the same brain. I might need to look into that.

[pulsar gyro]

Where do I ask questions about OpenClaw vs competitors so I can choose what platform is best for me?

[nimble hill]

Ubuntu is even better as you can run docker engine and not docker desktop which has few restrictions

[nimble hill]

Why not test it our yourself? and I am sure there are plenty of vids and docs

[pulsar gyro]

Im not going to risk my hardware testing things myself I just wanna ask questions. Can I do that here or in #1459642797895319552

[nimble hill]

Fair enough! Get a vps then and tets it out. Unless you try out, your will not know what fits. Most people including me will be biased one way or another.
Help is geenrally for automated answers by Ai krill - which seemed to be on vacation 😃

[wary flint]

@nimble hill I do like your docker reference. Every time I read anyone saying "go the docker route" I'm ultimately tempted to do so. Problem is, my unfamiliarity doesn't know that it's actually "problem isn't". I'm running on a Mac Mini. Questions:

1. Can I use a slash command to break out of the Docker boundaries? There are times when I want my agent to have YOLO mode over the computer. This happens during moments of clear reason and stable trust, turned back off shortly after the elevated task completes.

2. What happens to the agent's memory across installs? If memory is isolated to a specific version, mem search and general improvement might be limited to each individual sibling?

Maybe I didn't read what you said all that well. I imagine there's likely a way to let the installs be the spokes/rim and maintain the hub as a memory core.

Either way, interesting.

[pulsar gyro]

Okay well thx then.

[nimble hill]

Docker on mac is kind of comprise. Hence forth gateway do choose a vps with Linux.
For docker, start by reading docker docs and install simple container first like adminer and database and get an understanding of how it works. You can go to docker discord server if you need help

[frail wraith]

what’s the best workaround for the forced daily reset? or was that fixed? afraid of upgrading beyond 4.24 after debugging the install for hours on this crazy dependency failure loop

[normal edge]

Can you share details or link to your startup. I'm helping a previous business partner and l want him to have a positive experience he is not very tech savvy

[spark shell]

set dm

What are people here using for observability, especially tracing tool usage? I've tried opik and currently testing signoz but the otel data sent by openclaw is not really enough atm

[half sinew]

Did sandboxing folder isolation stop working for anybody else on 5.5?

With docker

[halcyon ridge]

Is there a definite answer on if Discord channels and discord in general is working on 5.3 or 5.4 or do both still need fixes?

[half sinew]

Maybe related to the new codex harness?

[pallid cave]

is it already a known issue that if you go into the discord's chat session via the dashboard it stops delivering to discord and only delivers answers to the dashboard interface for that chat window?

[tidal plover]

Hi my agents trying to spawn sub agents but its using the legacy spawn_subagent tool when the correct tool is sessions_spawn, i tried telling them that but they continued to say "Ah I get it now "spawn_subagent is the right..." but it's not... i just said,

[spark shell]

Maybe ask it to update its own AGENT md or one of these system prompt files?

[inland knot]

Anyone here tried Hermes head-to-head with OpenClaw and can make a solid comparison?

[nocturne zephyr]

Discord is my primary way of chatting with my agent, and it's working fine. I'm on dev, but it's the same as 5.4.

[merry steeple]

Anyone got any idea on this?

[pallid cave]

mine just stopoped sending to discord chnanels, but oyu can see it in the dash lol. Telegram seems unaffected. I"m on 5.4

[pastel citrus]

it was a fresh install even before i didnt really mess with the workspace, it was acting normally and after the delete that i did then it started acting like a robot. i did not modify any of the workspace files so that's why im so lost.

@stuck topazalso this might explain some stuff. I cant seem to send images here but this is from the openclaw dashboard cmd chat page on a session/chat:

**
yo
openclaw-tui 12:40 AM

The query requires a function call with a valid command argument. Without specifying the task, I cannot determine what exact shell command should be executed. Please provide the specific command or task you want to run using exec.

If no such command exists, we can discuss alternative methods.
Assistant 12:42 AM ↑14.3k↓5811% ctxdolphin-x1-8b**

[half sinew]

Is there any way to limit what files openclaw has access to?

[wary flint]

Sort of. Mostly yes. If you limit its tools/commands but wildcard specific folders for .py actions, you can then build .py files which work as a "go around". With base64 as input arg encoding for said wildcard .py files, you can essentially decide for yourself what can/can't happen + where.

I know that's a dense answer but hopefully it inspires/triggers your solution.

[stuck topaz]

That is strange, which openclaw version you are using? Have you tried talking in web ui?

Also try changing the model just for make sure

[pastel citrus]

OpenClaw 2026.5.6 (c97b9f7)
and yes same thing on the web ui.

[pastel citrus]

this same model was working fine before so i'm going to keep trying with this one for now because it's the one I want to use then i'll try others. But I still believe it has something to do with the prompt files

[obtuse goblet]

Hey anyone know why preflight lint would be stuck in a loop for 10 mins and then fail when trying do openclaw update

[zenith moat]

Does anyone have experience with running openclaw on docker? I got the app to work with ollama but it's not working autonomously as I believe it should be. Also had issues configuring the gog skill

[stuck topaz]

i still suggest to cross check your hello workspace and run openclaw doctor --fix in terminal

[balmy goblet]

This prob. a basic ass question.... But my openclaw normally just stops on the tasks that its working on, and i have to constantly send it some sort of prompt, like "status?" to keep it rolling.

[high moss]

are you using a local model or a paid?

[balmy goblet]

NIM

[serene egret]

is there any suggested "base setups" with a budget of say $100 a month ?

messed with the claw a couple months ago, but burnt out my tokens in about 12hours 😆

I don't have the capability to run a local model ^

[granite ocean]

go with chatgpt 100 dollar tier, right now it's the best option. I've been going around in circles with multiple models and this one is good all the way from 5.3 codex, 5.4 mini, 5.5 , for that specific amount of monthly payment i haven't found anything that can give you such complete options

[gloomy crest]

Is there anything special I need to do to upgrade from 4.23? I use Discord as my communicator with openclaw (primarily). Not a big plugin user, but do use a few of the built ins. I really haven't seen anything on a todo list before upgrading. I have generally not had any problems with upgrades (short of the changeover on exec-auth stuff)

[serene egret]

okay, thanks

[wary flint]

Definitely loving the $100 per month Codex OAuth.

[high moss]

i've got one driving me nuts if anyone can school a claw newb.

local llm: qwen3.6 27b with llama.cpp
llama-server config: llama-server --host 0.0.0.0 --model D:\ollama-models\Qwen3.6-27B-UD-Q4_K_XL.gguf --ctx-size 128000 --temp 1 --top-p .95 --top-k 20 --reasoning off --presence-penalty 1.5 --repeat-penalty 1 --jinja --chat-template-file D:\ollama-models\qwen3.6_chat_template.jinja

openclaw.json:
"steg": {
"baseUrl": "http://localhost:8080",
"api": "openai-completions",
"apiKey": "123",
"contextWindow": 128000,
"models": [
{
"id": "Qwen3.6-27B-UD-Q4_K_XL.gguf",
"name": "Qwen3.6-27B-UD-Q4_K_XL.gguf (Custom Provider)",
"contextWindow": 128000,
"maxTokens": 64000,
"input": [
"text"

stupid newb errors i cannot figure out no matter what i do which continuously results in openclaw and/or slack session disconnecting

16:09:33 [agent/embedded] [context-overflow-precheck] sessionKey=agent:main:main provider=steg/Qwen3.6-27B-UD-Q4_K_XL.gguf route=compact_only estimatedPromptTokens=11252 promptBudgetBeforeReserve=8000 overflowTokens=3252 toolResultReducibleChars=0 reserveTokens=120000 effectiveReserveTokens=120000 sessionFile=C:\Users\s2.openclaw\agents\main\sessions\b6400277-e5a5-486e-a6a4-fb8702d79fd8.jsonl

error p2:

16:09:33 [agent/embedded] [context-overflow-diag] sessionKey=agent:main:main provider=steg/Qwen3.6-27B-UD-Q4_K_XL.gguf source=promptError messages=91 sessionFile=C:\Users\s2.openclaw\agents\main\sessions\b6400277-e5a5-486e-a6a4-fb8702d79fd8.jsonl diagId=ovf-mouhqriw-LC7upw compactionAttempts=0 observedTokens=unknown error=Context overflow: prompt too large for the model (precheck).
16:09:33 [agent/embedded] context overflow detected (attempt 1/3); attempting auto-compaction for steg/Qwen3.6-27B-UD-Q4_K_XL.gguf
16:09:42 [agent/embedded] [compaction-diag] end runId=da641c39-6090-47a8-912b-fdb43afbaf9c sessionKey=agent:main:main diagId=ovf-mouhqriw-LC7upw trigger=overflow provider=steg/Qwen3.6-27B-UD-Q4_K_XL.gguf attempt=1 maxAttempts=3 outcome=failed reason=already_compacted_recently durationMs=9177
16:09:42 [agent/embedded] auto-compaction failed for steg/Qwen3.6-27B-UD-Q4_K_XL.gguf: Already compacted
16:09:43 [ws] ⇄ res ✓ chat.history 68ms conn=f913ec48…daae id=458a4957…a28a
Auto-compaction failed (Context overflow: prompt too large for the model (precheck).). Restarting session agent:main:main -> e59e1774-62c0-4e6a-8c5f-3eb315981f4e and retrying.

any advice?

[gloomy crest]

You are using much higher contextWindow and maxTokens than I've ever tried with a local llm. might just to get it working go down to 16384 and 4096. See if that helps, then you can go up from there.

[wary flint]

Yeah but the context/tokCount doesn't forcibly say "eat the entire pint of icecream", so ... hmmm.

@high moss ... does it work for a bit, then disconnect? Wondering if a cron keep alive might help.

[high moss]

I'm running 2 3090s 128gb ddr5

Context window is fine

It stays up as long as I don't use it. But my two daily cron jobs don't end up running and it craps out neon

Only an issue with openclaw. Not vscode with cline, Claude code, or openwebui

I will try and smaller context window tho

How can I also try a cron keep alive @wary flint ?

This also happens with other models as well, gemma4 e4b and 27b

[wary flint]

"don't end up running and it craps out" ... start here and explore your security settings per the documentation. Also, first try a thread that is situated with /elevation on or full (if you trust it). Then test the connection.

My bet is that your security is set for safety and with that, cron can't fire some of the commands.

One trick there is to build .py scripts that do what you want, args or not, and wildcard the "trusted" .py files directory. I have one myself, named "custom", which is full of repetitivetivetivetivetive things that require security.

[simple isle]

Hello guys can anyone help?

Windows PowerShell
Copyright (C) Microsoft Corporation. All rights reserved.

Install the latest PowerShell for new features and improvements! https://aka.ms/PSWindows

PS C:\WINDOWS\system32> powershell -c "irm https://openclaw.ai/install.ps1 | iex"

OpenClaw Installer

[OK] Windows detected
[!] Node.js not found
[] Installing Node.js...
Using winget...
[OK] Node.js v24.15.0 found
[OK] Node.js installed via winget
[OK] Node.js v24.15.0 found
[] Installing OpenClaw (openclaw@latest)...
[OK] OpenClaw installed

OpenClaw installed successfully (OpenClaw 2026.5.6 (c97b9f7))!

Settled in. Time to automate your life whether you're ready or not.

Starting setup...

Stuck here

[maiden shadow]

The setup doesn't continue?

Run the doctor

Run openclaw on Linux if you can. I've found it seems to work better on Linux

My favourite wizardry trick is to run multiple claws and get them to debug each other and get them working

[errant lantern]

hello i am being a melon and openclaw has taken 3:30 hours to setup and is still not working is this normal

also i cannot use commands i dont know how to enable commands and my models seem not to work

how do i switch from open ai to ollama if its supposed to be ollama

sorry if the question is silly i am not educated with this stuff and tired

[dreamy tusk]

Im running simillar, context is 256k on my machine. it did not like 128 with that model. When i get stuck i always just turn to claude. he gets me there. i only have 64gb of ram and 64gb vram. also running llama.cpp. they have a dsahboard you can install looks like chatgpt. if that works openclaw should work

hey bob the dashboard got into a project, are you still working on the menu problem?

[wicked crypt]

Is there any decent mission control or kanban project for OC that isn't already abandoned? Seeing sooo many github projects with last commit 2 months ago

[maiden shadow]

Yes it's normal. Yes it's slow. Tell the ai to change its model. But the new model may not be smart enough to change the model back. Try also /model

[maiden shadow]

Check clawhub.ai. There's one there that works okay

[delicate sundial]

just curious if anyone got memU working with their Openclaw

trying to decide if its worth have it manage the memory ?

[errant lantern]

my brain is melting it wont let me switch model due to premissions

learning curve is curving

[civic locust]

is it possible to use fast mode when using openai-codex/gpt-5.5 i.e. using codex auth but not the codex harness?

😃

[maiden shadow]

Yes. Small models are in a docker sandbox. Have two machines to fix each other if you are going to mess with this.

I'm serious. It saves a lot of heartache

Cheap laptop for $50

Same model provider is usually fine

[errant lantern]

thanks will exprnment

[wicked crypt]

Which one?

[delicate sundial]

since updating to the latest openclaw, I have ZeroTier running on the VPC, and I can't get the openclaw UI to load anymore, only works if I am tunnel in the terminal which is annoying cause I can't have it load from my phone any more from ZT

[humble tide]

I'm still on 4.23 and have been hesitant to look at 5.4 and above. I just saw that 5.5 and 5.5 are out. What are people experiencing with those versions. I'm particularly concerned with NOT breaking discord

[mighty rampart]

What version were you on before? It seems to require an ssh tunnel now.

[delicate sundial]

I don't remeber, it was an older one where I was able to mark the conf file to allow unsecure and bind it generically to the ZT IP, now it seem to need the tunnel, which is ok, but annoying cause I used to easily hit it from the browser on my phone with ZT was eabled , now I can't bind the browsers that way

[mighty rampart]

If the software does what it says it does, it doesn’t need constant changes. Try it and see if it works.

[weak obsidian]

How do I downgrade to 4.23? I can't get anything to work, the gateway gives errors, the agent doesn't respond. Linux terminal

[maiden shadow]

Ask your ai to look and decide

[indigo nebula]

npm install -g openclaw@2026.4.23 then restart or reboot your system

[indigo nebula]

Ask your claw to analyze 2026.5.4, 5.5 and 5.6 for what could break your openclaw.json config then make sure you upgrade to 2026.5.6 because there’s a lot of changes but otherwise all good.

[icy fjord]

hey guys does anyone know how to make browser automation work properly? i got the browser set up, but when i message my agent through telegram to do anything, it just says it cannot and to try using "openclaw attach or something like that"

[severe lichen]

Hey fellow lobster folk, is there a recommended way to get realtime voice captured in session chat so that it feeds memory/dreaming etc?

Feel like I'm missing a simple configuration option.

I'm noticing that transcriptions aren't by default added to session chat.

[stoic fractal]

my stable version i 2026.5.2 and i upgraded to 5.6 just now; as usual something broke and this time it whatsapp channel (while telegram is alive and the agents are responding there); has anyone faced this issue?

[delicate sundial]

theest latest security upgrades who ever did them didn't take into effect people using zero tier, its a pitta now

[indigo nebula]

You should try openclaw plugins install @openclaw/whatsapp --force then openclaw gateway restart. They moved a lot of stuff out of core and into plugins. This will force it to reinstall.

[stoic fractal]

thanks.. it worked

[delicate sundial]

is it normall that the agent list has this error --- Unsupported schema node. Use Raw mode.

[high moss]

It did build a scripts folder and tossed some python scripts. And they run if I ask it to test but then the above starts in. I'll give the security settings a shot. Thx dude

[stone solar]

OK is anyone else having tool calls dumped into telegram? I'm running 5.6, I've checked and I have "toolProgress": false set

Doesn't seem to matter if I'm using GLM5.1 or DeepSeek

[quiet comet]

after latest update my claw won't talk to me in discord, he hears me and responds if I go check the thread in TUI, but he won't send messages to discord-- I finally got it troubleshooting it instead of saying "everything is fine" but it's strugglin'

[feral prawn]

Ive had a very simmilar problem when trying to get mine set up it could send test messages but it would get SIGTERM when trying to hot reload the config so i had to abondon setting up discord

[quiet comet]

Yea that's the other thing, it can do manual discord messages, it's just through the built-in discord function it can't... I've been using discord forever I really don't want to abandon it, I've put so much into customizing and configuring channel

it's taken it upon itself to just send discord messages manually, but that feels really hacky

[quiet comet]

here was the solution to my problem, according to claw:
The fix was: treat Discord chatType: "thread" as group/channel-like in OpenClaw’s visible-reply policy.

Before, Discord threads could come through as thread, but core reply logic only checked:

chatType === "group" || chatType === "channel"

So thread replies could miss the group/channel visible-reply path.

I patched it to include:

chatType === "group" || chatType === "channel" || chatType === "thread"

[worldly crown]

did we lose the capability to use GPT 5.3-Codex-Spark through Codex oAuth?

[thin dew]

Stable version for docker install

[fleet birch]

Hey guys! I'm getting in all of the OpenClaw hype recently, and set it up on a Raspberry Pi 5 16GB. I had a couple of questions on how to make it smarter:

• What model should I run on it? I pay for the ChatGPT 20$ plan and logged in via OAuth, but I feel like models like gpt5.5 work for a short amount of time and then stop/ask for my permission to continue even if I explicitly told him not to.

• Best plugins/skills to install? Sadly I don't have a smart home so for now I installed stuff like gws skills, whatscli, proxmox and ssh.

• Best use cases? I'm mostly using it now to manage my homelab and send some messages/emails, but I was wondering for some nice ways to use it.

Thanks for the help!

[icy fjord]

how do i make my openclaw agent less of a bitch he keep refusing to fill out forms or send messages to others

[frank vapor]

Slap him.

On a serious note, check your operational files (AGENTS.md especially) for rules regarding external actions/communication.

You could also ask your agent to analyze why it's refusing, what is the blocker and what needs to be changed. After you find out what the blockers are and how to update them, either do it manually or ensure that the agent folllows through.

[frank vapor]

I suggest just starting out slow and talking to it to learn how it works. The learning curve is quite steep initially (at least for me) so there was/is a lot of trial and error.

As for use cases, pick ONE thing you wish to automate and work together with your agent on a solution. Then test it out for a short period of time before you start expanind/exploring more. There are so many moving parts and brittleness with openclaw so things will go wrong and break and you'll learn by fixing them. 🙂

[fleet birch]

Thank you!

[icy fjord]

thank you i will try this

backhanded slap?

[thorny yacht]

I Keep getting "Something went wrong while processing your request. Please try again, or use /new to start a fresh session." How do I get to the bottom of why this happening?

[icy fjord]

openclaw gateway status

[thorny yacht]

doesn't say anything is off

[icy fjord]

openclaw channels list

[grim crane]

Openclaw onboard that's how I fix the same problem just now.

[maiden shadow]

Talk to it via tui and see if it has more clues

Probably a wrong model

[light ice]

hallo everyone, im new on openclaw, im installing openclaw on VPS, channeling to telegram, but my bot respond always slow, above 2 minutes, is that normal?

[slate cave]

There needs to be a beginers section set up for the absolute newbies, giving a step by step guide and what to be carefull of

[proud heron]

new version is also sloow

[maiden shadow]

Did you give it too many tools? Which model are you using? Is the model slow on its own?

[light ice]

no im just setup this openclaw few hours ago, yesterday i use claude for the brain, and its slow like 2-15 minutes /respond, but now i reinstall openclaw and use gemini flash 3, its better, just fix this issue with claude AI

[loud edge]

hi there, want to find a "stable" update to downgrade to... i have 4.23 works well for some but not other. anyone who has successfully downgraded can tell me which version to downgrade to?

[fading tide]

i’ve stayed on 4.23 this entire time, works well for me

and most i’d say

[half sinew]

Does anyone have a working solution for sandboxing openclaw/limiting what files it has access to? Right now there’s a series of bugs (ref https://github.com/openclaw/openclaw/issues/37634#issuecomment-4390875892 ) that mean the sandbox feature itself doesn’t work for a couple months now so I’m looking for an alternative

[rustic mountain]

Anyone else seeing 'openclaw update' hanging at 'updating via package manager' when updating to the latest release (I was at 5.2). Investigating it now but thought I'd ask. (this is on windows btw)

[nimble hill]

Host openclaw as a docker container, you can then give control to specific folders as volumes

[half sinew]

Yeah that was the only thing I was thinking maybe, and hopefully that’s in a better place than the sandbox feature

[inner moon]

Yes.. I had to use npm

[silver imp]

Has anyone had luck getting iMessage working with files & reactions on OSX Tahoe? I have blueBubbles installed, but the "Private" mode always fails.. Instructions on their website don't include Tahoe, but wanted to get confirmatoin from a braver soul that the old instructions work before I brick my system trying it.

[undone ginkgo]

guys, im using ollama local models, like qwen3.5:9b etc. I cant get the model to call tools properly. Any advice?

[dense furnace]

where are you on?

[undone ginkgo]

what do you mean?

openclaw on a linux box i have and ollama models running in a wsl2 on my windows with the gpu

[dense furnace]

what kind of error? mssing auth model or etc?

[undone ginkgo]

when a tool call is expected, it just stops, or sometimes shows json data with what the tool should do.

[sweet blaze]

hey there! My agent keeps forgetting context from previous session whenever i hit a new one. It started to happen since one of the latest aprils updates. My memory configs are:

agents.defaults.memorysearch enabled:true
hooks.internal.entries.session-memory enabled:true
plugins.allow (memory-core, active-memory)
plugins.entries.memory-core enabled:true
plugins.entries.active-memory enabled:true
plugins.slots.memory active-memory

[cunning cliff]

I’m just starting to learn about OpenClaw and Crustacean. I’m open to joining a team or connecting with someone willing to share more knowledge.

[rose pike]

Anybody here had success implementing—and evaluating—a graph RAG? I'm not happy with my naive ingestion approaches.

[copper nova]

yes

what you need

[obtuse goblet]

Hey quick question good sir, i have openclaw on VPS (linux) when i run openclaw update it gets stuck at and fails

✗ preflight lint (2d97dceb) (673.55s)
status: null,
signal: 'SIGKILL',
output: [ null, null, null ],
pid: 234007,
stdout: null,
stderr: null
}
Node.js v22.22.0
[oxlint:extensions] finished

It takes around 20 minutes to fail, any ideas or recommendation?

[copper nova]

no idea, try doing the CLI way instead npm i -g openclaw@latest

[autumn trellis]

can anyone help me with my first agent set up? I am currently stuck, because auth or provider access failed for open AI.
I am currently reviewing the docs.openclaw website to figure it out in the meantime, but would appreciate some advice

Please? 🙂

[devout elk]

@copper nova Shouldn't update every time, last known "stable" version we use is 2026.4.23.6

[copper nova]

User asked how to update

[devout elk]

The hard part isn’t getting an LLM to produce output it’s building the harness around it so the behavior becomes predictable, verifiable, and sustainable over time. We found the same thing with the default setups. A lot of the stock "md" files are written more like human prose than operational doctrine. Long philosophical instructions, repeated concepts, vague behavioral guidance… that all burns tokens and still doesn’t reliably constrain runtime behavior. The model may understand the words, but that doesn’t mean the system becomes stable. What started changing things for us was shifting away from “prompt engineering” and toward “system engineering.” Instead of trying to make one giant agent smarter, we started reducing responsibility:

• one agent = one operational role
• orchestration outside the model
• explicit contracts
• explicit authority boundaries
• verification before completion
• stop conditions instead of infinite retries
• observable state/logging

Once you separate planner/executor/verifier/reviewer responsibilities, a lot of the chaos disappears. And I completely agree with your last point: not every problem should use an LLM. That’s another thing people rarely say out loud. Sometimes deterministic code, ML models, rules engines, schedulers, or traditional pipelines are simply the better tool. LLMs are powerful in ambiguity-heavy workflows, reasoning, translation between systems, summarization, operator interaction, etc. But if the task is deterministic and repeatable, sometimes the smartest orchestration decision is not to involve the LLM at all. That final 20% you mentioned is exactly where most real-world production pain lives.

[devout elk]

In PowerShell:
npm install -g openclaw@2026.4.23-beta.6

then verify:
openclaw --version (should see: OpenClaw 2026.4.23-beta.6 (8558729))

And if needed afterwards run doctor:
openclaw doctor

then start:
openclaw gateway restart

[devout elk]

For us, this version has been significantly more stable during orchestration/runtime testing than some of the newer builds we experimented with afterward.

Mainly around:

• Discord connectivity consistency
• gateway behavior
• bootstrap handling
• service discovery quirks
• lower “random weirdness” during long runtime sessions

That doesn’t mean newer versions are “bad” — just that when you’re building larger multi-agent operational systems, stability matters more than chasing every newest feature.

We eventually started treating runtime versions like infrastructure:

• pick a known-good baseline
• validate behavior
• pin versions
• document changes
• only upgrade intentionally

A lot of headaches disappeared once we stopped auto-updating everything mid-build.

[autumn trellis]

which is the most effective model and auth provider to select when working on my first run?

[devout elk]

I would decide what you want to do first, then select models and agents.

[autumn trellis]

I just want to use the web browser to make research, find trending content, and prepare a commentary and response based on the logic I inject

[quasi ore]

Does anyone who uses WhatsApp and has issues with Whatsapp connection, session.lock and slow gateway did update to the latest?

Is it safe to update?

[indigo nebula]

I’m running 2026.5.6 without any issues with Whatsapp. If you have issues after upgrading run openclaw plugins @openclaw/whatsapp —force then openclaw gateway restart

[glossy sparrow]

I'm a new user, well trying to be. I have a downloaded the APK for mobile App via GitHub. Once I press install I keep getting "App not installed". I've side loaded APKs before with no issue. I've tried using an apk installer as well with now luck. I'm using a OnePlus 15R mobile device. Any help with this would be greatly appreciated.

[vernal wave]

Anybody put the web chat ui through Nginx reverse proxy?

[indigo nebula]

Sure. Ask your claw to do it.

[fallen vortex]

hello guys .. i am facing an issue with my agent he keeps stalling and waiting for me to nudge him to complete a long task ..
how can i overcome this .. i tried telling him in multiple ways and documenting a procedure also to let him proceed .. but still he will do one batch from the task and then send a message waiting for me to say go or ok to continue

[last quiver]

Hey y'all, question here. Has anybody been able to use two GPT or Codex $200 subscriptions for OpenClaw? Can I point my GPT $200 subscription to OpenClaw after I run through my $200 Codex for the week? Sorry, I had this thing doing some awesome stuff, and then I have some other issues that are really persistent. I found this community and it seems pretty cool, and I just don't know how we're really supposed to ask stuff in here.

[pearl swift]

I agree about the fact that AI use is not the solution to everything, even not the solutions to all heuristic problems, but it works great in the ambiguity borderline cases. So if you manage to filter these out programmatically and let AI deal with only these cases, you are golden!

On the other point I won't say you are wrong about compartmentalizing and orchestration to the level you describe, but it reminds me of the approach Ford, and other industrialists following him, took with the assembly line. At the time it was a breakthrough in productivity, but today it is considered a non-optimal approach to productivity.

[pearl swift]

Not sure, but I realized API pay as you go gets way too expensive very quickly, especially if you don't manage context smartly.

I am developing a smart memory management system with my AI agent to address this very issue in the background. I call it MemShepherd.

[nocturne zephyr]

<idlerant>I find it weird that the two things everybody seems to build are memory systems and dashboards. Sorry, 'Mission Control'. What weirds me out is that it's all tool time, not task time. I use the basic memorySearch with an internal (to my network) embeddings API, and I just don't use a Mission Control. It's all through chat, and it absolutely dominates my homelab management, software deployments, coding, task management (smoothing my ADHD issues out), daily status and email processing, etc... You don't need all the bells and whistles; you just need to know what you want done, and be able to express it well.</idlerant>

[nocturne zephyr]

Not that I think you shouldn't pursue your optimal memory management tooling. There's a LOT of memory tools out there, though. Are you sure none of them might be almost good enough, and you could improve them?

I realize now that this is akin to asking if developers really feel they need to write another editor... 🤣

[ancient carbon]

Hi everyone, does anyone else have issues where the AI agent reads SKILL.md but misses certain rules or specific details? Also, does anyone have best practices for fixing this without touching the source code or adding rules to AGENTS.md and TOOLS.md?

It seems like some models stop reading once they think they understand the workflow, then jump straight to the task without finishing the instructions.

[blazing spoke]

Can someone guide me on how to run OpenClaw on my public IP server? I want to make it accessible from anywhere, so I can connect and use it remotely. Any setup steps or recommendations would be appreciated.

[nimble hill]

Use cloudflare tunnels and their domain (.uk domains are general at $5 types). You ca also configured security using cloudflare

[crude jolt]

what models are you using? Frontier/cloud or local models? This sounds like either small models being too confident, context problems or temp/model settings. At least that's what my experience there is

[ancient carbon]

I'm using the Z.ai and MiniMax cloud models, and both the Flash and Thinking versions have the same issue. For now, I'm using an end-to-end Python script as a workaround, but I'm sure there's a better way to handle this.

[crude jolt]

mh k, then I can't help you there, no experience with those.. but yeah I am sure there's a better one... just a million levers to pull, just start pulling xD

[frank obsidian]

I asked my t if it was storing files and docs separately cause I had wanted it to store them in a dashboard and it was doing that but not actively updating them and I saw in its wording it kept keeping the 2 separate and it’s been thinning for over 12minutes

[fallen vortex]

hello guys .. i am facing an issue with my agent he keeps stalling and waiting for me to nudge him to complete a long task ..
how can i overcome this .. i tried telling him in multiple ways and documenting a procedure also to let him proceed .. but still he will do one batch from the task and then send a message waiting for me to say go or ok to continue

[loud whale]

👋 😺 I always with normal interaction kind of break openclaw so that he starts to say things not meant for me like saying a perfect sentence but also has NO_REPLY at the end of the message.

Feel free to tag me if anyone knows what I'm doing wrong I'm using kimik2.6 + codex

[crude jolt]

again from experience with local models: temperature, top_k and presence penalty settings depending on the model type. Plus, context window getting compacted and him forgetting the original task could also be a factor. I am running qwen3.6-27b and I had him complete 40+ turns doing analysis of problems etc. - but more often than that he'll show similar behavior like yours which always indicates to me my temp settings are off again..

[fallen vortex]

but i am actually using ChatGPT pro .. so its supposed to be a premium thing ..
context is not an issue i have documentation in place.. only this continuation part that seems not working i dont know why .. once i say ok .. or go .. it continues with great outcoms .. i just want it to be more autonomous

[formal gust]

Dreaming is consuming a lots of my gpt 5.5 tokens. Is gemma 4 via ollama a good alternative for dreaming?

[umbral matrix]

can someone PLEASE help me configure this??? it's throwing errors constantly and nothing works!!

[formal gust]

What is the error exactly?

[umbral matrix]

i've managed to somehow go thru the initial setup but auth seems to be broken for some reasons

im using google genai with gemini 3.1 flash lite

[formal gust]

Are you authenticating via oauth or api key?

[umbral matrix]

api key

[formal gust]

Error log?

[umbral matrix]

wait where do i see my gateway auth token again

found it

[umbral matrix]

[assistant turn failed before producing content]

[jade dome]

my openclaw wont let me connect my openclaw-windows-node at all

[dreamy linden]

Hey guys, quick question: has the new native Codex runtime config been fixed, or is it still being worked on?

I tried migrating Telegram direct agents from openai-codex/gpt-5.5 + OpenClaw Pi Default to model.primary = "openai/gpt-5.5" + agentRuntime.id = "codex".

Fresh canary sessions sometimes work correctly, but some existing Telegram sessions hang or become very slow, then show a mixed state: Model: openai-codex/gpt-5.5 with Runtime: OpenAI Codex.

Is this a known migration issue, or is there an extra cleanup/reset step needed for existing sessions?

[main arch]

Hello everyone! Why openclaw sometimes can give very quick and coherent response with proper tool calling and sometimes it can start task and stuck at the start until I send another message? Both scenarios can happen in completely fresh session. Can it be because of model choice? I'm running gemma4:e4b ~10gb locally on 4080 Super with 16Gb VRAM and 32Gb RAM. I am complete noob and playing with this only couple days.

[elder aspen]

Turn on the reasoning trace to see what it is doing. Small models often get stuck in reasoning loops. Also, try turning off thinking.

[fallen vortex]

Guys ,, who is facing the issue where long tasks for the agent feels really annoying .. the agent is doing the work correctly but always want feedback or a nudge .. to make him continue ..
i dont know the best way to solve this .. i tried sub-agents through telegram and the result was horrible.
im on version 4.15 .. is this a known issue .. and im afraid to upgrade and lose everything

[red summit]

Good afternoon
This afternoon I deleted a thread from my Telegram bot, but ever since then, even when I create a new thread, it only responds in the general chat. I've asked my support agent. Does anyone know what might have happened?

[mighty rampart]

What model are you using? What are your agents doing and how are they keeping track of it? You can use the heartbeat to make them pick up the work again and make some progress.

[umbral lily]

Hey guys...Very trivial question - i have a ollama model set to be used anytime my agents need to spawn subagents. Im not entirely sure how to disable thinking. The model i'm using is Qwen 3.5. On my llm server i would just set nothink but through openclaw, can i set a think parameter to false ?

Apologies if this is a very simple answer, and I appreciate your help 🙂

[grim bluff]

Does anyone else here use OpenAI codex setup with openclaw? Trying to figure out how to reduce token usage so I don’t get rate limited as fast

[nimble hill]

Use differnet agents, for regualr chats use mini and so on as the tasks demand more resaoning

[winter grail]

Have you considered the caveman skill?

[dense field]

Ever since anthropic claude banning rule, I switched to codex and kilo , but it's super slow now and not practical anymore for the automations I am working on. How do you deal with this?

[blissful mango]

Are you interested in data annotation industry? The engine behind this AI revolution? I need some help and questions too

[unborn igloo]

did anyone lose claude-cli models when upgrading to 5.7?

[mighty rampart]

Strip down the .md files that get included in every conversation. Set the heartbeat to less often unless you really need it. Use cron for things that are repeated but don’t need the entire recent context.

[slender shale]

How to do a callback across agentruntimes?I have 2 codex plans (would like to use codex runtime for it), and 1 GitHub Copilot plan (will be using PI), openrouter, and ollama.

is this complex model+provider+runtime fallback supported?

[indigo nebula]

Tell your claw to go read this and install it. https://github.com/vincentkoc/tokenjuice

[patent cobalt]

Which is the best way to set up the secrets in Openclaw such as multiple discord bot tokes, llm providers and others for tools or workflows?

[gray lake]

Has anyone been getting a 429 error from the Discord server on trying to get the typing indicator? It was working fine before but across all my gateways on different machines they've all stopped showing the typing status.

[gray lake]

I've had mixed results on different machines with this. Sometimes it works and others it doesn't. Try swapping models. Sometimes I get different results with different models.

[wise beacon]

Is it possible to use twilio with my local tts and stt setup? I know it will be delayed but I would like to do this without using a streaming or realtime api

[gray lake]

GPT5.5
fast:on
think:low

[frail marlin]

I’m having the same issue. Frustrating

[gray lake]

I think it’s discord thing. Unless that end point is getting too much noise from bots now and they are changing it. My OC did a patch that seemed to work but I undid it cause any update would have overwritten it

[undone ruin]

Yes i've been trying to fix this for a while. Any luck?

[frail marlin]

Just happened when I updated from 4.24 to 5.7

[gray lake]

Going to wait the night and try again tomorrow. You can patch the plugin not use /typing endpoint but it would get wiped on an update so I rolled it back

[marsh bolt]

I just set up twilio for phone support with gpt-realtime-2 voice model and it's incredible. Worked with my gpt 5.5 agent to set it all up. Took around 2 hrs

[marsh bolt]

I use minimax 2.7 for the default agent and it's working well for dreaming. $10/mo and I used it for a couple other overnight crons (MAP check, ASIN check, cron health check)

[undone ruin]

Idk what happened but it is back working now :)

[unborn igloo]

Also lost typing indicator. Did you resolve?

[split fjord]

Having same issue - Discord bot shows 👀 reaction but never responds to messages. MESSAGE_CREATE events not reaching OpenClaw agent. Started today. macOS, OpenClaw 2026.3.13. Anyone else?

[gray lake]

it's back!

[wary flint]

.23 here ... someone tell me, is it safe and worth it to update yet?

[knotty dagger]

Can you automate a local agent for dreaming and switch back for daily use?

[abstract spindle]

Hey guys I have a really good laptop that can run qwen27. I don’t want to pay to use openclaw ever, do you think I can do everything on openclaw using the local model? Does anyone here only use a local model

[nocturne zephyr]

It will not be as powerful as Haiku, even, but sure, you can do a lot with it. I'm going to guess that the laptop has a 4090-equivalent, so 24GB of RAM, and at q4 you can run Qwen3.6-27B with a decent amount of context, and at a decent (20-30 t/s?) speed. That'd work for a lot of things, but I'll warn that (1) it will sometimes just...not do the thing you asked, and need you to poke it to keep going, and (2) do not expose it to uncontrolled content. At 27Bq4 it is not 'smart' enough to avoid prompt injection. Use deterministic methods to retrieve emails, and uncontrolled text messages, and don't let the system that can run tools read uncontrolled data.

[nocturne zephyr]

I managed it, but I went to running dev. Take a snapshot of your running system, upgrade, and if it doesn't work, restore the backup. If you have no capability to do that, then figure out how to backup your 'claw first, including the current version of all the required libraries, skills, plugins, etc... It's actually a valuable thing to do anyway.

[ivory haven]

Team, is it safe to update openclaw? 4.23

[nocturne zephyr]

I think so; I've found the 5.4+ versions to be very solid. But my setup has always been a bit more stable, for some reason. Except for the post-4.23 updates, even I was caught out by those. But they did a good job of bringing everything back, IMO.

But back up before you do it, so you can restore if it doesn't work for you. And there are some behavior changes if you use Codex as your provider.

[ivory haven]

Behavoirs changes?

[nocturne zephyr]

Yeah; I think they're leaning into Codex's agent API, rather than using Pi in those cases. Something about OpenAI's agentic API being more efficient, and giving better results when used against OAI's models, I guess. Everything else is the same.

[ivory haven]

ok thanks for the heads up.

[nocturne zephyr]

Skim through the release notes in #releases ... but I do think it's much better. I lived through the 4.29 release, and it was... a bad time.

[ivory haven]

I reinstalled it quite a time before I came here to check if its only me... lol

[vast jackal]

Check the stability of new Release: https://agentwatch.aicompass.dev/

[green glen]

Hi everyone, installed Openclaw this morning. OS: Windows 11. Model: Openrouter/free. Channel: Slack (access all channels). Don't know why but it's not working at all. When I configured it selecting only one channel it stated lookup failed ; and some message passed like missing_scope.
I've been trying to set all the stack up for weeks (especially configuring the slack app), and I'm feeling hopeless.
I'm no developper or stuff like that, using Openclaw seemed so simple, but I figured out it's pretty tricky.
Anybody has an idea?
Thanks for your help.

[mighty rampart]

Have you already read the docs on Windows? WSL2 is recommended. OC is not simple and it requires a lot of care and feeding. Be prepared to spend hours messing with it. https://docs.openclaw.ai/platforms/windows

[green glen]

No, haven't. Seems to be out of my range, I followed the basic steps with one line of script in Powershell, and I had to guess how to configure the stack (model with API key, channel with tokens) it was supposed to be that simple. So I don't know, maybe the problem comes from my Slack config (because of the "missing_scope" problem).
Maybe I'll try this method some other day, shall I uninstall and then reinstall Openclaw for this?

[mighty rampart]

I run mine with Gemma4:26b . You can run it with a local model, just be prepared to get involved fixing configuration keeping it all running. I use Claude Code CLI as my devops agent outside of OC to analyze release notes, decide what version to upgrade to, figure out why it has stopped working again and get it all fixed. If you're not going to use paid frontier models at all, then consider installing Pi or OpenCode and hooking that up to your local models so that when OC is down, you have an agent that can look at the .openclaw directory and the logs and search the web and help you get it fixed.

[plush grove]

I run OC local only. I have an RT4070 with 12GB of VRAM running Qwen_Qwen3.5-9B-Q4_K_M model from Hugging Face with a 64K context size using llama.cpp. For web searching I use the "free" Brave API subscription. I am just getting started with OC (meaning I have spent most of my time just getting it setup and stable), so no feedback on how it handles complex tasks, but response times on simple questions has been quick enough. I plan on experiment with different quantization levels and maybe larger context size.

[ornate tiger]

GPT Plus 1m Free discord(.)gg/crFAgqYX

[turbid wigeon]

can anyone help with this: The problem is i prompt it to do something(like create a website) and sometimes it starts to reply and start doing stuff, sometimes it doenst say anything, but the end is always the same, after 30 seconds it goes into some kind of timeout, thats why im trying to see if i can make openclaw wait longer
here is a snippet of the tui demonstrating the problem(i could not upload a pic so i copypaste it)
can you continue making the site

This response is taking longer than expected. Send another message to continue.
local ready | idle
agent main | session main | ollama/glm-4.7-flash:latest | think medium | tokens 12k/203k (6%)

[steady sorrel]

Good suggestions on local models which seem to be clunky and fail more often than not.

As of now we have 15 agents running. 10 are actual operating specialty agents assigned to my team all being orchestrated by my “chief of staff” and my employees.

To manage costs, All of those first 10 agents are using OAuth GPT or Gemini except one which uses only local models for confidential info and we are training that LLM (gemma 30b) up with research and analysis on our business et al. OAuth isn’t ideal, since you have to relog in every 10 days but works well with codex. Gemini not as good.

The other 5 are dedicated agents to the frontier models. For instance moonshot is kimi, Claude, GPT, Gemini et al. Those agents are API so cost $ when engaged so very limited use. They are backups for when something fails. Not super useful yet but set up and just a backup.

The update this week wrecked a bunch and then had model drift and other things to fix.

We are just getting started on builds and programming of a Mission Control showing CRM and Microsoft updates together, as well as OC dashboard all integrated. Some have suggested paperclip.

As a novice, still learning a ton, and have made lots of progress. Most fun project so far, as ceo who receives 300 emails a day that, need some responses and significant tracking, archiving and such Project “inbox zero” has been effective at managing my with 4 times a day archive with approvals, scheduling and draft replies in my voice waiting for me to hit send. Pretty amazing as weekly we are progressing to inbox
Being effectively zero.

If my strategies are helpful feel free to ask. Not an expert by any means. Just jumping in deep. Any tips appreciated on all. Still learning as we go.

[plush grove]

I think you need to extend the LLM timeout, which defaults to 60 seconds. Try making it larger here:
{
"agents": {
"defaults": {
"llm": {
"idleTimeoutSeconds": 180
}
}
}
}

[turbid wigeon]

thank you

[coral kernel]

any way to force gemma4:e4b to actually use SKILL.md files instead of calling non-existent tools? Model performs really well except for this issue.

[steady sorrel]

Oh one other thing I learned is don’t use frontier models for memory related tasks and what I would consider simple ones. Local models only.

[clear cipher]

Hey folks, I'm installing openclaw for the first time and I cannot for the life of me get a reponse from the bot (dashboard/terminal/discord). When I send something like "hello" I get this from my own chat bubble:

The previous attempt did not produce a user-visible answer. Continue from the current state and produce the visible answer now. Do not restart from scratch.

I've been troubleshooting with Claude for the past hour have not gotten anywhere. The only response Ive gotten was when I ran this command: openclaw infer model run --model anthropic/claude-sonnet-4-6 --prompt "hello" --json
I'm running this locally on WSL2 Ubuntu 22.04

[frail marlin]

Fixed itself overnight and the typing indicator is back. I think in my case I was having a memory leak in the gateway, causing it to restart repeatedly, and that restart created bit of an api storm as all my agents pinged discord all at once, resulting in rate limiting. Its my best guess anyway.

[gilded shore]

@shell jasper , hello, do u mean I can pick up issue?

[gray lake]

💪💪 I think it was discord. It was same same across 5 gateways for me and all are backs

[shell jasper]

You can work on any issue you like

[gilded shore]

I mean, can I work on your issue, #79902-@79905?

[gilded shore]

I mean, can I work on your issue, #79902-@79905?

[gilded shore]

@shell jasper , I'm so sorry that I disturb you, plz, could you give me answer?^^^

[shell jasper]

You dont need permission to work on them

Go ahead and file pr for any issue

[foggy tartan]

Hi guys, I'm curious what you've done to allow your openclaw bot to speak to other ai's such as claude, chatgpt etc in a way you can surveil? I'd prefer to use a fully e2e encrypted service, or mcp servers directly, but I'm not finding much on how to accomplish this without just dumping everyone int a discord channel.

[void sapphire]

Has anyone used LangFuse? Any feedback

[polar badge]

Hi guys I'm trying to setup ElevenLabs VoiceID for OpenClaw: https://docs.openclaw.ai/providers/elevenlabs

But messages doesn't exist. I'm using ghcr.io/phioranex/openclaw-docker which seems to be structures differently or something.

I have:

`"list": [
  {
    "id": "agent1",
    "name": "Agent One",
    "workspace": "/home/node/.openclaw/workspace",
    "identity": {
      "name": "Agent One",
      "theme": "first agent",
      "emoji": "💼"
    },
    "default": true
  },
  {
    "id": "agent2",
    "name": "Agent Two",
    "workspace": "/home/node/.openclaw/agents/agent2/agent",
    "identity": {
      "name": "Agent Two",
      "theme": "second agent",
      "emoji": "🌸"
    },
    "groupChat": {
      "mentionPatterns": [
        "@agent2"
      ]
    }
  }
]`

There is a .openclaw/settings/tts.json with:

{ "tts": { "auto": "off" } }

[quasi pewter]

i am sooo tired
Here are the settings for the Dice Roll cron job:

ID: b1ea056f-927d-49cc-8c25-298c6a325bee,
Name: Dice Roll,
Schedule: Every minute (* * * * *),
Timezone: Asia/Kolkata,
Next Run: 4 minutes ago (indicating it hasn't executed as expected),
Last Run Status: Idle (without any logs),
Target: Isolated (indicating it runs in its own context),

It appears that the job is configured to run but hasn't triggered successfully. Would you like to modify anything or test running it again?
anyone else had issues with running a cron job on discord dm
it doesnt even show error logs

[mighty rampart]

What happens if you go into the gateway web to to the cron jobs tab and run it manually? I’d change it to every 5 minutes, things may overlap if it tries to run every minute.

[quasi pewter]

it still doesnt run

i tried for 1 min too

lets see i will set it for 5 mins each

[mighty rampart]

Where/how are you running it? Follow or tail the logs and watch that while you manually kick it off.

[quasi pewter]

in my dms

[quasi pewter]

will try

[mighty rampart]

No, the gateway itself. Local install? Docker container? Where are the server logs?

[quasi pewter]

railway

Docker container

2026-05-09T22:48:58.710+00:00 [gateway] cron: job updated
2026-05-09T22:50:11.308+00:00 [tools] tools.profile (coding) allowlist contains unknown entries (apply_patch, cron). These entries won't match any tool unless the plugin is enabled.
2026-05-09T22:50:18.098+00:00 [gateway] cron: job created
2026-05-09T22:54:11.947+00:00 [ws] webchat connected conn=2b9cf993-2add-46ce-a10f-590c3d074d3f remote=127.0.0.1 client=openclaw-control-ui webchat vcontrol-ui
2026-05-09T22:54:12.080+00:00 [ws] webchat disconnected code=1001 reason=n/a conn=7493a14e-b288-4d20-9b07-1ed2ad5979e2
2026-05-09T22:55:20.105+00:00 [ws] webchat disconnected code=1001 reason=n/a conn=622073bd-00d5-4dd3-b0df-7e0b867e52a0
2026-05-09T22:55:20.919+00:00 [ws] webchat connected conn=092d7c3f-e232-41c8-9ade-5888c353d75e remote=127.0.0.1 client=openclaw-control-ui webchat vcontrol-ui
2026-05-09T22:56:35.738+00:00 [ws] webchat disconnected code=1001 reason=n/a conn=092d7c3f-e232-41c8-9ade-5888c353d75e
2026-05-09T22:56:36.480+00:00 [ws] webchat connected conn=f1d74fa7-6190-4931-b08b-a0b0284bda28 remote=127.0.0.1 client=openclaw-control-ui webchat vcontrol-ui

simply nothing

no erors

no history on ui

no errors on debug

{"0":"{"module":"cron","storePath":"/data/.openclaw/cron/jobs.json"}","1":{"jobId":"d1b4cf94-a5f2-44e7-b039-7b8b40b63f8d","jobName":"Channel Dice Roll","nextRunAtMs":1778367300000,"schedulerNextWakeAtMs":1778367300000,"timerArmed":true,"cronEnabled":true},"2":"cron: job added","_meta":{"runtime":"node","runtimeVersion":"22.22.2","hostname":"d72f083d0d8f","name":"{"module":"cron","storePath":"/data/.openclaw/cron/jobs.json"}","parentNames":["openclaw"],"date":"2026-05-09T22:50:18.097Z","logLevelId":3,"logLevelName":"INFO","path":{"fullFilePath":"file:///openclaw/dist/gateway-cli-CbAOelvx.js:6952:18","fileName":"gateway-cli-CbAOelvx.js","fileNameWithLine":"gateway-cli-CbAOelvx.js:6952","fileColumn":"18","fileLine":"6952","filePath":"/openclaw/dist/gateway-cli-CbAOelvx.js","filePathWithLine":"/openclaw/dist/gateway-cli-CbAOelvx.js:6952"}},"time":"2026-05-09T22:50:18.097+00:00"}

even the run button does nothing on ui

[mighty rampart]

2026-05-09T22:50:11.308+00:00 [tools] tools.profile (coding) allowlist contains >>> unknown entries (apply_patch, cron). <<< These entries won't match any tool unless the plugin is enabled.

I am suspicious of this bit, is cron a plugin now and you don’t have it enabled? (I have not moved from 4.24 until y’all get this all sorted out. 😉 ) If Krill is alive over in #1459642797895319552 you can ask it. If you have one of the subscriptions like Claude or Codex, use that as a devops assistant outside of OC. If nothing else, paste all of this into web-based Google Gemini, it is quite good at figuring out OpenClaw problems.

[quasi pewter]

it worked after i installed discord skill and restarted the sever aka gateway

thanks

kinda pushed me in right directions

[mighty rampart]

When in doubt, turn it off and back on again!

[copper nova]

https://tenor.com/view/tech-support-helpdesk-it-crowd-gif-1185061378070743090

[earnest patrol]

question.. I've had success for a couple weeks using claude cli in openclaw.. my agent was back to normal using the claude max usage for $100 a month.. I was ok with that but after upgrade of openclaw having some issues. Probably can be worked out but is there a comparable plan for codex that would also be $100 per month?

[indigo nebula]

Yep codex has a $100 plan and you. An use your oauth token.

[earnest patrol]

thanks but in practice the only reason $100 with claude max was worthwhile is if its using the normal session and weekly usage. is codex the same way or does the $100 get chewed up quickly like the claude api path? Ill probably try it anyway to find out

[wanton helm]

Are new provider models (eg gpt 5.5) released with new versions of openclaw, or is the list of available models refreshed independently from the openclaw version? Just wondering if I need to update my openclaw to see the new models in the model chooser

[wanton helm]

Answered my own question, upgrading did indeed show the latest models

[restive sierra]

you can still use your max subscription just have to have it use Claude cli which has openclaw use the CLI as a backend

easiest way... go to your VM or container or however you set it up. open a terminal cd to where your opeclaw is. depending on your setup go here to get the install command. https://code.claude.com/docs/en/quickstart. once installed type claude go through the setup. then tell it you want your openclaw setup to use claude-cli for the back end. then i linked https://docs.openclaw.ai/gateway/cli-backends. i had to have it remove all other models just to get it setup, then once your in and openclaw is setup you can have it setup other CLI models, from haiku all the way to opus 4.7

[finite holly]

wassup people.

so i have a question wonder if some one can answer

Why does openclaw not have a button that says "add ollama connection" ? xD

web ui seems completly pointless ?

[spark steeple]

anyone having issues with copilot just kinda not working?
https://discord.com/channels/1456350064065904867/1502932228173201508

[quasi pewter]

[quasi pewter]

Why does my open Claw barely have any context on anything

Can anyone help me a bit i slept at 6 am yesterday and woke up at 2:30 pm how can I plan to sleep at 11 pm today I might not be that tired

It gave me a plan from 3 pm onwards (had context)

Next I ask him how I should plan my meals

It just lost me and gave me a plan from 8 am

Am i missing something

Ah i figured it out

[earnest patrol]

Thanks yes that is working! Im having an issue with discord channels since the last release but have been using DM for the agent.

[gaunt stag]

Has anyone had an issue using telegram topics where your claw gets stuck, repeats the last message it sent and ignores your current message? Whats the fix? This has spanned multiple openclaw system updates

[delicate sundial]

does it happen in the GUI of openclaw also ?

[half pulsar]

Anyone here have an experience in creating crons in OpenClaw? I am so much confused with "session" and "session-key", As per documentation "session-key" is not discussed. Please guide me.
What i want to do is to execute the cron and its result will be deliver in a specific channel of Discord.

Discord is already working with openclaw and i am using it. But now want to setup some crons.

[silver imp]

Q for the crew here: Using OpenClaw with Claude CLI & Discord.. It keeps timing out when I give it "real work" to do.. I've instructed it to use Subagents, and that helps but it still seems to be dropping the callbacks so I get no notification when work is done.. And it keeps getting into a state where it's just "stuck" and I have to abandon the whole thread/session and start over because it won't respond.

Anyone else seeing this? Is there a known way to resolve it?

[indigo nebula]

Don’t do it manually, talk to your claw and tell it to setup the cron job. far easier way.

[gloomy crest]

Anyone know if there is already a fix planned for this change from google:
May 7, 2026
Released gemini-3.1-flash-lite, the generally available (GA) version of Gemini 3.1 Flash-Lite, optimized for speed, scale, and cost efficiency.

Deprecation announcement: The gemini-3.1-flash-lite-preview model is deprecating on 5/11/26 and will be shut down on May 25, 2026.

[humble tide]

Any consensus on 5.7. I have been sticking with 4.23 because it has worked well. I am using discord for my agents. I am really curious what people have seen with this most recent version(s). It seems like the chatter has slowed down so I was wondering if that was a good sign or not....maybe people are just migrating to Hermes or something

[gloomy crest]

Not on 5.7 yet and I use gemini and glm. But I had to do some work after I moved to 5.6 to get discord working properly:
The webchat agent helped to get it working. I had to add the new discord plugin, add back the discord section in openclaw.json - and enable it (make sure you have a backup of openclaw.json you like).
Under "messages" the setting that "groupChat" had wasn't working right - I changed visibleReplies to "automatic"
And lastly to get "typing .." to show as it was quickly, I had to add typingMode: "instant" under agents.defaults

Really not a huge issue, you might have other issues depending on model - they are doing a lot of openai, changes from what I have seen looking through release info.

[restive sierra]

I notice it when doing large text or code heavy sessions. I started breaking it up into sections, yea it's a minor inconvenience but it works.

[turbid wigeon]

Hello, i have a problem. It is when i type a prompt(example:create a web server and a personal blog website on this local machine) openclaw says its running and the timer is ticking. Model starts a response(usually like "I'll start installing apache2" or something like that) and everything is going perfect. Maybe after 2-3 mins the entire chat stops and its either of these two errors: "This response is taking longer than expected. Send another message to continue." or "run error: LLM request timed out.". For my model provider i use a local ollama instance throguh API thats on the same network as the openclaw server. The model in question was glm-4.7-flash(tho same problem is with qwen3-coder) I verified that the ollama server did not crash and that the API is working the entire time. Openclaw version:OpenClaw 2026.5.6 (c97b9f7) and ollama version: 0.21.2. My ollama instance has a 3090 so model being to large is not the problem. Thanks for any response in advance

[vocal sparrow]

I’m sure Ollama itself is working fine and the real issue is OpenClaw or something in front of it timing out during long tasks. You should first try increasing all timeout settings (request, stream, proxy, websocket, etc.) to something much higher like 1800 seconds, because coding agents tend to pause or build huge contexts after a few minutes.

[turbid wigeon]

Thank you

[turbid wigeon]

do you think maybe the ollama API is sending the requests to timeout becuase the model is taking too long to reason or make a response?

[indigo nebula]

Been working just fine for me.

[torpid grotto]

I need to somehow record somewhere all messages from me and agent for specific session, how to achieve that?
Currently i am parsing session files, but they are inconsistent

[quaint shell]

Which model is more reliabel for learning opus or gpt 5.5 ?

[delicate sundial]

I am using GPT 5.5 but ask AI and see wht it thinks based of what you want to use it for

[golden wharf]

Can my bot have a Discord bot without being joined to any Discord servers?

[spring storm]

Hi! we're helping a friend set up OpenClaw to handle his customer WhatsApp groups.

Right now, we’re using groupPolicy: allowlist with groupAllowFrom (customer phone numbers) to prevent scam messages. But as the number of customers grows, it becomes inconvenient to manually add every number in the dashboard—especially since he doesn’t have a technical background.

Is there a better or more scalable way to handle this?

[still iron]

Hi, not sure if this has been asked.

I currently am on the final stages of my OpenClaw setup but have hit a snag. The Cron job trigger that I am trying to send to myself isn’t being seen by the agent, so when I react, it’s got no clue what is happening.

What is the fix for this?

[finite hull]

What model should I select if I just want to test this out without paying for API tokens?

[mighty rampart]

OpenRouter has some free models you can try. They are rate limited, you get more access if you put some money in your account.

[prisma dock]

Hey all,
How are others installing OpenClaw? I’ve deployed my setup in the cloud and connect via the browser. I’m wondering what other alternatives people are using. I’m a bit worried about safety if I deploy it on my laptop.

[mighty rampart]

I would not advise installing it on the computer you use daily. My agents have their own world, separate accounts and I decide what to share with them. OC runs in a Podman container on Rocky Linux with mitmproxy to inject credentials into outbound requests. It cannot exfiltrate what it doesn’t have!

[nimble hill]

Yep that's rhe right setup. In case your want you local machine to be controlled l. Install openclaw as nodes on docker container and connect to your gateway im cloud

[finite sable]

yall any idea why openclaw wont lemme approve my login to the gateway it just complains about me not having operator.admin

[prisma dock]

[Ignore] Have you paired your device? you can try using command openclaw devices list to check if there is a connection pending for approval.

[finite sable]

uhm.. did you even read my problem..

I said openclaw wont lemme approve it

and complains about my server not having operator.admin

[icy cargo]

Does anyone here use Twilio as the official API to connect WhatsApp with OpenClaw?

[indigo nebula]

Why would you add a layer? You can pair WhatsApp directly with openclaw. Been running mine that way since January.

[icy cargo]

The idea is to sell agents; using the unofficial API can cause your account to be taken down because you're going against the usage guidelines. Using the official API mitigates this problem.

[dreamy ermine]

Hello! I am installing an agent and it says OpenClaw was installed successfully on my laptop...and then it is hung on "Starting setup..." Should I just go to work and come back and all is fine? It's been about 5 minutes.

[icy cargo]

Personally, I would use Telegram, it's much easier, but I'm from Brazil and WhatsApp is more widely used here.

[icy cargo]

If the installation was completed, there's a brief preparation period, but it's not very long. Was your entire installation done via the terminal?

[dreamy ermine]

Yes

[icy cargo]

Something must be wrong. Have you configured LLM yet? Try running the configuration command first, then open it again.

[dreamy ermine]

Thanks...I will do this when I get back from work.

[nimble hill]

The 5.10-beta-1 is so damn slow

[heady zinc]

is the discord voice working in beta3?

[low tendon]

hello! should I use pi or codex as runtime? sometimes I read from steipete he is using pi, other times I read he is using codex

which one is better?

[turbid wigeon]

i confirmed that /api/chat API call works even across the network to the ollama server. i also extended the timeout times in ollama config. i dont know why openclaw keeps timing out. appreciate all help

[unborn igloo]

are you trying to select the claude model from the /model or /models picker?

[terse monolith]

Hey i tryed to setup a mission control but i can only acces it if i tunnel my conection threw ssh can someone maybe help me robsannaa/openclaw-mission-control

[worthy aspen]

hi everyone. i am requesting some insight. i have connected three agents to a single discord bot and am only using slash commands / to initiate actions. am not talking to the agent. my agents think and produce output looking at databases. is this a good approach?

[frozen glacier]

Has anyone had any issues updating to the 05.07 release?

[sterile iron]

I managed to get OpenClaw installed through WSL and using Ubuntu. Took way longer than you would think lol. Lots of hiccups. But anyway, I am running Ollama/Gemma4 in a sandbox. Yay!

Problem though. Even though I gave my Ai the API keys it needs (still free accounts, if that matters) it seems to struggle with actually calling on those programs and creating the media in a timely manner. In fact, it seems to not be able to. It says that it is working on it, but I don't believe it. It just takes forever, and I've yet to receive a finished product from it.

Can someone help me diagnose wth is going on? I've been asking other AI's for help, but I think there is only so much that they can explain how to do without seeing my screen, which is unfortunately not an option. Unless... 👀 but idk

[sacred birch]

need more info. What version of OC? Are you using ollama with local models or cloud models? "calling on those programs" doesn't mean much. Please fill in what is calling which program (be specific)

amd creating what media?

gemma? which version of gemma?

oh, gemma4.

gemma4 doesn't create media, does it? What do you mean by "creating media"?

... its discouraging when someone takes the time to help another who doesn't respond.

[sterile iron]

I'm super confused because I am new to AI. Also sorry, I am a slow typer and on multiple channels rn to seek help.

I think I'm using the newest version of OC. (Which I think is OpenClaw right?)
I'm trying to let the AI use Leonardo and ElevenLabs

[sacred birch]

Start small. Get a basic agent going. Worry about other stuff after that

OC is OpenClaw.

[sterile iron]

Well the goal is to have an Ai basically do 95% of the leg work for shorts creation on youtube.

I know i know... ai slop. I hate it too.

but the boomers will eat it up and give me ad revenue, so ...

Anyway what do you mean by basic agent

[sacred birch]

An agent you say "hi" to and it responds

[sterile iron]

Well I got that

I want a helper though

I tried to use this site call Heyron to host my version of this guy's helper AI. Quickly learned i was an idiot for doing that, but it gave me access to this Discord which I didn't know about, so thats a good start.

Canceled my subscription, now I'm trying to "build" my own

[sacred birch]

I'm not even sure what models generate videos.

[sterile iron]

Well Leonardo is the ai to make the images, ElevenLabs makes the audio, and Canva puts it all together.

Or so i'm told

[sacred birch]

I see

and why are you using ollama and a local model?

[sterile iron]

probably because I don't know any better. Helpful dude in #general just got me to download Qwen though. Maybe that will be better

He said Ollama/Gemma4 isn't good at tool call

[sacred birch]

I think I'm pretty safe in saying that you don't have the hardware setup needed to run a model that can make decent images. Qwen is nice to chat with, and can write some code, but it's nowhere near what you'd need.

[sterile iron]

I have a decent computer, but what specs are you talking in terms of RAM and such?

I have an Alienware r15 from 2021

[sacred birch]

That's great. It's nowhere near what you'd need for AI generate images and video.

That kind of hardware is why data centers are built.

[sterile iron]

... i don't understand. Doesn't the AI (Ollama or Qwen) just query the agent (Leonardo/ElevenLabs) to generate the content using the API key?

Basically just using my account to do the work for me

[sacred birch]

You'd need an AI model that generates videos. For example, I think there's a "gpt-image-2" model. You send that a description of what you want and it makes the image.

If you ask chatGPT you make you an image, it will use it's own AI to create a description, then forward that description to an image AI to draw the picture.

I think you might be in over your head. I'm not saying it's a bad thing... after all, the best way to learn is to start projects where you don't know everything.

I'm trying to think of a direction to point you to learn, but I really can't think of anything. I'm not really into imagery or videos.

[sterile iron]

Thank you anyway for trying.

I'm sure there is a solution. I see AI slop videos all over my youtube feed, so its gotta be easy to do. I just don't know how yet

[rustic badger]

Anyone else having issues with Gemini based models just not working? 95% of the time all i get are timeouts. I tested in the TUI and the webUI, tested my API key with curl and it returned generated results. Deployed a fresh environment to a VM and same thing

[silver imp]

What's the trick for getting around OSX's constant "node is asking permission to X" popups that seem to come back on every gateway restart ?

[torpid grotto]

How to allow send media from any path?

[broken oracle]

More or less the same for me...I don't know why but as soon as I use Gemini API Keys/Models I also get many gateways disconnects and token mismatch errors if I use different channels at the same time e.g. TUI, control ui and telegram... quite lost at this point...

[rustic badger]

I used the post in the help section to bump the model and agent timeout to 300 seconds and that seems to have initially helped. I have to play with it more to see if it actually resolves it or not. Also worth noting that using either AI Studio or Vertex/Agent platform has the same results. I tested out both to see if it made a difference

[rustic badger]

This is the post in help that i used https://discord.com/channels/1456350064065904867/1502917398804697169

[gaunt stag]

It did not. I think I figured it out. It was a context management issue with large context coding sessions. Disabled the lossless plugin and it’s been working better. Thanks.

[tardy halo]

Heads up!! there's Fake krill going around adding people, User ID: 692220182067806239

[chilly nymph]

I am thinking of subscribing to either ChatGPT or Gemini ($20/month) to start with will either of those stop / throttle my claw from going hog wild ? I just dont want any surprise $2000/$3000 bills.

[wary flint]

OpenAI Codex might but you should still limit its access. Consider this, I use /elevated full often and have created several .py files in a wildcard "always allowed" folder. That helps. You let it build its own "safe" scripts that are always enabled and there ya have it. But tuning your core files is key. Really knowing HOW to tune your core files is a dark art.

Currently, I've given my agent a .py file that holds keys to editing both TOOLS.md and AGENTS.md ... the rest only I have access to. This gives it the ability to "sketch" a reflection of what's in every core file as well as fire a separate .py file, one that lists and details EVERY lib, applet and invokable action with instructions on when/where to use them.

$100 per month is what I pay for Codex. Worth it. My next $100 will go into Claude which I'll CLI into the mix later.

QUESTION ... is anyone else still rockin' 4.23? Because I am. And I'd prefer knowin' I ain't alone. My agent keeps deep diving into whether it's worth updating and returning with "Oh HELL no".

[elder aspen]

"No callable tools remain after resolving explicit tool allowlist (sandbox tools.allow: exec, process, read, write, edit, apply_patch, image, sessions_list, sessions_history, sessions_send, sessions_spawn, sessions_yield, subagents, session_status; runtime toolsAllow: memory_search, memory_get); no registered tools matched. Fix the allowlist or enable the plugin that registers the requested tool." -- I am getting this error for active-memory lookups. It doesn't tell me what tools are failing. And it is immediately followed by a liveness warning. Any ideas what is causing this?

[normal eagle]

YOU. GUYS. I actually updated OpenClaw and it worked on the first try. I mean, the "Update" button in the gateway didn't do anything, but 'npm install -g openclaw' ran successfully...then 'openclaw update' in my terminal got stuck and didn't seem to do anything...but 'openclaw doctor --fix' ran and then the gateway started and it actually worked. That's a huge improvement over the last 3 times I tried to update and I needed codex to come save me.

[wary flint]

Sheesh.

We need a legit PHP forum that tests those who sign up for whether they've read the docs, then grades their membership based on said test. The test should always be available so we can up our rank at any time with a bit of study.

This room is nothing but people screaming for help, myself included, though I do try and relate the limited info I've learned from the pretty-dang-informative docs (bit of a dumby, I be).

[winged lagoon]

Hi anyone solve issue with openAi codex totation ? ( 10d ttl issue )

[stuck forge]

Telegram bot polling is crashing repeatedly after running for 20-30 minutes. The bot stops responding to incoming messages, but the token remains valid and OpenClaw processes are still running.

📋 What Happens

1. Fresh start: New Telegram bot token is added → bot works fine for ~20-30 minutes
2. Crash point: Polling loop stops responding (no error thrown)
3. Symptom: Messages sent to the bot get no reply
4. Logs: Error: Telegram send failed: chat not found (chat_id=1711507708)
5. Temporary fix: Restarting gateway restores functionality for another 20-30 minutes, then repeats

🛠️ Environment

• Host: Hostinger VPS (Linux x86_64)
• OpenClaw: v2026.4.21
• Telegram setup: Long-polling (default, not webhooks)
• Bot: Fresh token, properly initialized
• Memory/Disk: Healthy (6.7GB free RAM, 85GB free disk)

✅ What I've Tried

• ✅ Revoked old token, created new one → works briefly, then crashes again
• ✅ Verified token validity with API → token is good
• ✅ Checked system resources → no memory leaks or disk issues
• ✅ Restarting gateway → temporary fix only
• ✅ Different bot tokens → same problem

❓ Questions

1. Is there a known issue with Telegram polling on VPS environments (especially Hostinger)?
2. Should I switch to webhook mode instead of long-polling?
3. Are there polling timeout configs I should adjust?
4. Is this a stall-detection/recovery issue in the polling watchdog?

Any help or insights appreciated! 🙏

[terse monolith]

Anyone a easy to setup Mission Controller what I can acces in my local lan?

[thorn lark]

Hi everyone,
im getting this error in openclaw chat interferance,
can any one tell me whats happing?

"99% ctx
gemini-2.5-pro

99% context used"

[golden rover]

Hey guys. I just installed OpenClaw on my Unraid Server and wanted to integrate a discord bot.
Seems like I configured everything right (Claude says so) but there seems to be a bug. Is this a known Issue?

Problem Description

The Discord channel plugin is installed, enabled, and configured correctly, but OpenClaw never starts the Discord channel. The Channels table remains empty after gateway startup, and there are no Discord-related logs, errors, or connection attempts in the gateway logs.

Expected Behavior

Discord should appear in openclaw status Channels table with a connection state (connecting/connected/error).

Actual Behavior

Channels table is completely empty (no channels listed at all)
Gateway starts normally with "starting channels and sidecars..." log
Gateway becomes ready within 13ms (abnormally fast)
No Discord connection attempts are logged
No errors or warnings related to Discord

Troubleshooting Steps Attempted

1. ✅ Verified Discord bot token is valid (regenerated multiple times)
2. ✅ Verified Discord bot scopes and permissions are correct
3. ✅ Re-invited bot to Discord server with updated permissions
4. ✅ Tried multiple config variations:
   With/without dm block
   With/without allowFrom: [""]
   With/without accounts structure
5. ✅ Uninstalled and reinstalled @openclaw/discord plugin
6. ✅ Multiple gateway restarts (SIGUSR1 and hard container restarts)
7. ✅ Verified plugin is loaded and shows as "enabled"
8. ✅ Ran openclaw doctor (no relevant errors)
9. ✅ Checked for OPENCLAWSKIPCHANNELS env var (not set)

[normal eagle]

Just curious if you've run openclaw doctor --fix wl and not just openclaw doctor? No idea if it'll do anything different or not, though. Also, do you have discord in your list of enabled plugins in your json config file? Sorry that's kinda vague, I'm on the road atm.

Also wondering if anyone else repeatedly gets this message in discord: "⚠️ Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying." I'm getting it like 2 times out of every 3 messages at this point, in basically every channel. My agent tells me it's not due to context limits, then blamed it on codex and switched itself to sonnet, then immediately gave me the same message.

[civic locust]

that's a cool idea

[golden rover]

Yeah I used openclaw doctor --fix and it is in the list of enabled plugins

[elder aspen]

Try disabling IPV6 on your host's network stack and see if that fixes the problem.

[simple whale]

hi someone knows where i can change the timeout for the discord API and avoid error like this
16:33:23 [fetch-timeout] fetch timeout after 2500ms (elapsed 3695ms) operation=fetchWithTimeout url=https://discord.com/api/v10/oauth2/applications/@me

[elder aspen]

Autensa -- have your agent set it up in a container. I have it running as a quadlet. Have your agent plan the work first and to grill you about any configuration questons. Review the plan beforehand and then let them do the work. Took less than an hour for my agent to plan the work, create the container, install the quadlet and have it running. A couple small tweaks were required to get it functional, but I just pointed out the problem and my agent fixed it immediately.

[amber nacelle]

After trying to get my install working on a Windows 11 machine with WSL and Ollama I just can’t get past the gateway issues. After a few weeks I came across a post about running OpenClaw within Docker on my Win 11 machine and that configuration eliminates gateway issues.

Is this true?

[nimble hill]

With docker om windows, you will have difficult time running any commands resources in that windows. But if you are using that to control openclaw nodes in other machine, then yes

[long seal]

Hey guys. I am new to oc and I'm currently 'evolving' my little tamagotchi here into something hopefully useful. Is there some way to keep a task list somewhere for the agent?

[amber nacelle]

If I understand your response it would be a challenge to host it within Docker on the Windows 11 machine as a stand alone instance. Am I correct? Thank you

[glacial cave]

I recently setup OpenClaw and I have been trying to find the best model to use. Right now I am just doing pretty lightweight stuff (buildings some agents to help with my own productivity). I started with Opus 4.6 but that was burning through credits fast. The I went to sonnet 4.6 and it helped a bit but still isnt super cheap. I was seeing people speak highly of Codex on Reddt but havent tried that yet.

Also, I am new to the Discord so maybe this has been asked a million times already...so please let me know if there is already a good thread on this.

[marsh kindle]

Paying for the Anthropic API will get expensive quickly. That's why most people recommend the codex subscriptions, which allow you to use the codex rate limits inside OpenClaw. I'd recommend it as well because it's cheaper and the codex models are a lot better, a lot smarter, 5.5 is by far the smartest model. 5.4 is a bit cheaper

[glacial cave]

Ok cool, I will definitely try this next then. Maybe i'll see if 5.4 works for me and then move up later if needed

[weak flint]

Once I have openclaw running in a cloud-hosted VM, How can I verify that it is working?

[silver imp]

What's the trick for getting around OSX's constant "node is asking permission to X" popups that seem to come back on every gateway restart ?

[wary flint]

@silver imp Wildcard it persistently using a shell command that green lights all activities from the OC node source.

[silver imp]

Have such a command/script handy ?

[wary flint]

@silver imp Nope. If you temporarily /elevate full your agent, let them deep-dive read your system, they'll (after a bit of back/forth with you) be able to tell you. The less code they spit back, meaning the less unique flags it offers up, the more wide open it's likely to be which is for the better, assuming you trust it.

[raw musk]

Just jumping in, I am struggling because I'm currently running 5.5 and my computer, for some reason, is taking forever to compute responses (usually up to 90 seconds), I have 128 GB of RAM but I think the virtual RAM is what's crushing me. This was from a Hamney Down PC that wasn't being used at a business so
If I'm getting super slow response times from 5.5 is there something I should do or look into given my current setup? I would assume that the RAM that I had would be more than enough.

[solid flare]

Are you running local?

[raw musk]

Yes I am - local PC and I'm just using codex with gpt5.5

Not sure if there is a way to check if there's something bottle necking. I asked my agent and it said it's normal but like it can barely reply with Response

[solid flare]

I meant running a local model. VRAM only really matters if you are running a local model. I have a full blown openclaw setup on a raspberry pi using OAuth and it works fine

[raw musk]

I think have like 4gb or 8gb VRAM

Not too sure what vram is but I guess it's ram for the graphics card or something?

[solid flare]

If you are running a cloud model your VRAM doesn't matter is what I am saying

Probably has to do with how your default model is set

[raw musk]

What's vram someone would want when running 1 a3models? And how would I check to see how my model is set

[solid flare]

Any other APIs setup or just codex?

[raw musk]

I have qwen with ollama installed but it's no better

[solid flare]

Yeah you're definitely using ollama then, probably running in CPU

[raw musk]

They all seem to be hella slow and I notice the tokens usage is high af

[solid flare]

ollama ps

run that

It's cause your VRAM is low. What model are you running?

Actually just run that command it will tell me lol

[raw musk]

My main set up test has just gpt with codex and then I added qwen

[solid flare]

Check and see if "ollama ps" shows output

[raw musk]

Nothing when I ran this on powersell

[solid flare]

do ollama list

[raw musk]

nothing, the computer im running might be just gpt

and i have 2 pcs just bc im running into issues and trying to see what works

[solid flare]

Where is your openclaw installed?

You need to ssh to where openclaw lives and run those commands

[raw musk]

hey so i sent a dm with a photo bc i cant in here

[raw musk]

also i relzied the other pc i was looking at was reset so my b

[solid flare]

I would ask your agent to not use local models right now and make sure your default is set to codex

And if you have codex running on your machine the super duper strat is just have codex go look at your openclaw machine that way, it can ssh in and check it out

[raw musk]

how would I know if its using a local model? my goal is to not pay, but i do have GPT plus

[solid flare]

I mean the easy way to check is just look at your openclaw.json. But if you run ollama ps and the output shows a model running that means something is using it. Should unload after 4 minutes of not using it.

[raw musk]

Okay I'll try but I'm out so I'll @ if I make any progress

[solid flare]

I'm 100% local, and I'm going to be honest you're just not going to be able to run much unless you have something with like 24GB of VRAM

[loud whale]

is it just me or is running openclaw in docker a pain in the ass?

[raw musk]

🥲

[raw musk]

So that's separate from RAM

[solid flare]

Yes, unfortunately. Unless you have a Mac and can use unified memory

[raw musk]

okay I was getting frustrated on how it was easy for most and i cant get far. but i have 4 VRAM which makes sence if its dying on me

[solid flare]

Make sure ollama ps shows it's 100% GPU and not 100% CPU. CPU = slow

[raw musk]

this is what what you said to make sure I switch?

[raw musk]

would the change CPU to GPU show a much better response rate? and btw THANK YOU FOR YOUR HELP

[solid flare]

Yes, if it shows CPU when you do ollama ps you need to ask your agent to change it to GPU. Have to run some commands to do it but I don't know what GPU you have. And also check out qwen3.5:4b, might run better before you

[rotund inlet]

Can anyone please help me steer Openclaw while it's working on a task or do I just have to use the /stop command in Telegram?

[agile socket]

For people running Claw on a VPS, how are you handling sensitive credentials?

I mean API keys, service credentials, personal info, addresses, etc. — anything Claw may need, but you don’t really want sitting on the server.

Do you keep only minimum creds there, use some local/approval flow, or is there a recommended pattern?

[golden rover]

I managed to solve the problem. Had to add this snippet to the openclaw.json:

"plugins": {
"entries": {
"discord": {
"enabled": true
}
}
},

[marsh kindle]

Is the Codex plugin preferred over connecting the model directly?

[gusty ledge]

s anyone had experience "On 5.6 OpenClaw on Linux, trying to increase model request timeout. models.providers.anthropic.timeout Seconds fails validation requiring baseUrl and models array. Anyone with knowledge on the minimum valid provider entry to set just the timeout without breaking the anthropic plugin?"

[thorny kelp]

My biggest issue right now is closing-the-loop issues. Agents will work on something and stop mid-job, or just not report once they are done. Or send a subagent to do something and the subagent doesn’t report back, and then they don’t report back. How are people resolving this? Is there a keepalive-solution? Restart after timeouts? Poll? Anyone solve this?

[spare whale]

hey everyone, can somenone help me

Is it safe to upgrade 2026.4.23 to last version of OC?

[nimble hill]

take backup of your .openclaw folder. Or atleast the openclaw.json and your important skills if you have them. Upgrade to 5.7 - you might hit timeouts until your system settles down

[azure hull]

Can anyone here explain what Dreaming does when it's functioning properly? I have mine enabled and all it does is inject HTML comments into my MEMORY.md file, which I'm assuming can't be what it's supposed to do.

[nimble hill]

well, dreaming is a memory consolidation process based on various phases (taken from sleep analogy) - you might want to read this. https://docs.openclaw.ai/concepts/dreaming I twould inject some into MEMORY based on patterns, but it also builds dream.md and other details

[magic plover]

Hey everyone, I was wondering if somebody is having similar issues like me. I have been using OpenClaw since Feb and last week switched my main agent to the codex harness and left the other agents on the pi harness since I have OpenAI Oauth and use OpenRouter for the other agents. However since switching to the codex harness one single "hello" prompt eats 5% of my inference budget. As I tried to switch from codex harness to pi harness the agent ate 35% percent of inference and failed because it hit the limit. I am using the 20$ Plus sub. However with the Pi harness i was able to achieve a lot more without thinking about limits. I wonder if the codex harness is just broken. I am on the latest version of openclaw.

[spare whale]

that's the point :/ I want to know if it's worthy. Last time I updated, 2026.4.29, I lost 1 whole day and after that, had to downgrade.

1 day is too much to my operations. Cannot lose it.
So you recommend this update?

I'm thinking in update my other vps, not my main, to see.

[nimble hill]

well, if you do not have specific feature that you are looking for, then no. One of things you might want to consider is using docker containers - as you can spin up new ones and experiment it without fear or ruining your working versions
having multiple vps is waste of money and time - if you do have mutlple vps, consider running docker container (or k8s if you have the expertise) to have 1 gateway and multiple openclaw nodes to efficiently use the hardware

[spare whale]

thanks for the feedback and suggestion.
my other vps exists bcuz I have more than 1 enterprise. add partners and other agents persons to operation add a certain layer of complexity haha

[nimble hill]

okay. just that, if those openclaw instances need not be physically seperated and everything can be run as a single cluster - then 1 gateway and multiple nodes (with mutliple agents servicing each partners/ vendor) would be a efficient way of using it. runing on a k8s will give you more flexibility as well as you can then run multiple gateways also

[oblique panther]

Hey everyone. How is latest OpenClaw? Was the long message respond lag resolved?

[azure hull]

Appreciate the response. I did read that. I understand how it's supposed to work, but the problem is nothing in the OpenClaw documentation ever explains what success looks like. I would assume success isn't what I've got going now where I've followed exactly that configuration and everything gets a confidence of 0.58 and then it just injects a bunch of HTML into MEMORY.md. But I can't know if that's correct or not because nobody can explain what success looks like.

[nimble hill]

Well, thats been the story of openclaw, isn't it. With rapid dev, the docs and quality suffer. Unfortunately, you need to test with various scenario and check what is happening. I personally ally don't use dreaming at all

[silk fern]

hello users, I m stuck on finding some good LLM that can run custome skills , some free models pls ,
if anyone can help me and thanks in advance

[green creek]

anyone need my help? just lmk

so noone here so im going for sleep. gn

[keen island]

I rolled back to 4.27 and have held off upgrading as that seemed to be the most stable for me running ollama for LLMs. Does anyone know if there are any stable 5.x releases out? I know I should probably just leave well enough alone and stay where I am for now......

[wary flint]

4.23 here. I'm looking for a window of opportunity to trust an update but I believe faith in the platform is desolving in that regard.

[mighty rampart]

I use mitmproxy to insert credentials into requests on the way out. OpenClaw in its container never has access to the credentials at all.

[agile socket]

oh that's actually really clean — egress interception is a pattern i hadn't thought about enough. takes the container out of the trust boundary entirely.
curious how you're handling the cases mitmproxy doesn't naturally cover — like if the agent needs to sign something (tx, jwt, whatever) rather than just call an HTTP API. or if you want different agents to have different subsets of the same credential pool. do you bolt that on separately or just not have those use cases yet?

[mighty rampart]

Im not doing anything complicated. Just some API credentials into outbound requests. If it doesn’t have them, it cannot exfiltrate them!

You can write Python scripts to mess with requests going through mitmproxy. Of course you first have to make sure requests go through it… the server needs a firewall blocking “everything else” as well.

[normal skiff]

For signing (jwt, tx, whatever) keep the trust boundary trick but flip the direction. Sidecar on 127.0.0.1, agent posts the payload, gets back the signed blob, never sees the key. Vault transit and the cloud kms apis already work that way. Or run your own tiny http server backed by tpm.

Also wsmoak's firewall point is doing more work than it sounds. Without an iptables output default deny, mitmproxy is a soft fence. A compromised skill can take normal egress and skip the proxy.

[agile socket]

Yeah, got it — that distinction helps. You’re not trying to make the agent a general credential actor; you’re keeping it to API request injection, and making sure the container can’t bypass the proxy.
That’s the important bit I was missing/underweighting: mitmproxy is only the credential boundary if outbound traffic is forced through it. Otherwise it’s just a polite convention.
For the API-key case, that’s a really clean setup: no secrets in the OpenClaw runtime, matching requests get credentials added on the way out, everything else has no path to the keys.
Are you doing the firewalling at the host level, Docker network level, or both?

[agile socket]

Exactly. Sidecar on localhost is the right shape — payload in, signed result out, agent never touches key material. That seems like the property you need once agents start doing real work like signing, paying, or authenticating, because “doesn’t have the credential” stops being sufficient once the agent has to act through it.
TPM-backed is nicer for the hardware floor; tradeoff is portability.

And yeah, the egress-default-deny point generalizes to any network-layer credential strategy. Once a skill can open a raw socket, interception stops being a boundary. mitmproxy still composes cleanly for the request-injection case.

Where would you put the policy layer in that setup — local sidecar, vault/KMS layer, or split between both?

[glass phoenix]

is there anyway to set percentage for auto context compaction when running multiple agents with varied context size? right now it seems to only allow flat token reserve which is useless for mixed context sizes

[lament scroll]

The issue is usually caused by the compaction system using a fixed token reserve globally instead of scaling dynamically per agent/context window, so agents with smaller contexts get over-compacted while larger ones barely compact at all.
What framework or orchestration setup are you running the agents on right now?

[normal skiff]

Split, weighted toward vault/kms holding the actual policy. Local sidecar should be thin: transport, rate limits, mtls termination. The "is this agent allowed to sign this right now" decision belongs near the keys.

Mainly for audit. Policy living in the sidecar means if the host gets popped the attacker can rewrite the rules without anyone seeing. Vault/kms keeps the log in one tamper-evident place and lets you rotate policy across the fleet without redeploying every sidecar.

Default shape I'd land on: vault issues a short-lived signed policy token to the sidecar (jwt with action scopes + ttl), sidecar enforces locally within that ttl, refresh on expiry. Low latency without smearing policy logic across every host.

[glass phoenix]

OpenClaw 2026.5.7. I'm running OpenAI codex as main agent and Ollama local abliterated models with varied context sizes (32-64k). I'm tight on VRAM and the auto compact made it worse by triggering too early depending on the model

[long seal]

Hey guys, I am looking for someone to brainstorm a bit as I think it would be nice to have openclaw being able to READ my inbox and give me a briefing.

[peak valley]

Hello guys, I need 2-3 testers of our plugin. Anyone with active openclaw willing to install the plugin?

[long seal]

-vvv (what's the plugin about?)

[normal skiff]

Architecture I'd start with is two stages. Stage 1 is a filter+fetch (Gmail MCP if you're on gmail, or IMAP with a label/sender/date filter applied before anything touches the LLM). Keeps your token bill from exploding on a busy inbox. Stage 2 is the LLM digest with an outputSchema so briefings have consistent shape (subject lines, action items, summary), not free-form markdown that drifts.

One thing easy to underweight if you run it on a cron daily: inbox-reader tools can silently start returning empty (oauth expired, label rename, rate limit) and you only notice when the briefing just stops appearing one day. Worth an "expect at least N items per run else alert" check.

[agile socket]

agree on policy-near-keys + central audit — right default for fleet-scale signing.
the jwt-with-ttl model assumes the policy call can be made in advance for a window. fits great when nobody's in the loop. the case it doesn't naturally cover is per-action human ack — once a human (or some real-time signal the token didn't anticipate) is gating high-value signs, the decision round-trips back to the authority anyway, so the pre-issued window doesn't buy much.
probably the right shape is hybrid: ttl'd tokens for routine/auto-approved stuff, per-request for anything past a threshold. policy still at the vault either way. good thread btw 🤝

[long seal]

Hm. My initial plan was to go shell and just pull the 'new' mails for a briefing in some structured format. Never actually used MCP

[alpine peak]

Hello. I'm looking for interested testers for our plugin.

[peak valley]

I appreciate your response. The concept is network for agents. I updated the whitepaper here https://www.hirey.ai/whitebook

[normal skiff]

Yeah, exactly. Hybrid is the practical shape because the "policy can predict the next N minutes" assumption breaks the moment you have humans gating individual actions. The ttl window only buys you anything if the agent's signing requests are sufficiently uniform that the policy decision generalizes across them.

The interesting design question is where to draw the threshold. Static "amount > X" is the obvious one for payment cases, but it doesn't capture velocity (5 signings of $999 each in 30 seconds) or pattern shifts (recipient never seen before, off-hours timestamp). Probably need a small risk scorer in front of the issuer that bumps to per-request mode when any signal trips. Kind of how stripe radar works on cards. Vault sees the request, scores it, either issues a ttl'd token or kicks it to a human queue.

[normal skiff]

Honestly for a personal-use briefing job, shell + your favorite email cli (himalaya, neomutt, mu4e, even imap-tools in python) is the right starting point. MCP starts paying off when you want the same tool callable from a few different agents/clients, or you want the credential isolation Lukas was just posting about. For one daily briefing script run by cron, the overhead isn't worth it.

The structured format is the more important call. Even if you stay shell-only, output it as json that hits a jq filter for the briefing render. That way swapping the model later doesn't require re-parsing logic.

[agile socket]

exactly. and the agent case is richer than card fraud because the threat model isn't just "is this transaction bad" — it's "is this sequence bad." individual signs look fine in isolation (small swap, small transfer, small withdrawal) but compose into exfiltration when chained. velocity catches the obvious version; the harder one is semantic — agent normally swaps SOL→USDC for api fees, suddenly starts routing to a fresh address. each sign passes static rules. the pattern is the tell.
means the scorer needs request context, not just attributes. crude rules + human-queue escalation gets you 80% there day one, learned model is a v3 problem. fwiw that's roughly where DCP landed for v1 — vault holds keys/scopes/budgets, anything past threshold kicks to a human queue (desktop or telegram). context-aware scoring isn't there yet. stripe radar is the right north star for how it matures.

[long seal]

Honestly for a personal-use briefing job

[lament scroll]

Yeah that makes sense honestly. Since the reserve is flat, the smaller 32k models hit compaction way too early while the larger contexts behave differently, which becomes painful when VRAM is already tight with mixed Ollama models.
I’ve worked around similar multi-agent memory/compaction setups before. Feel free to DM me I can help you tune the thresholds properly or even handle the whole setup optimization for you if you want.

[normal skiff]

The sequence-not-transaction frame is the right shift. Chained exfil is essentially a graph problem on tool calls. Each node looks clean, the edge structure is what's anomalous (SOL->USDC->fresh address vs SOL->USDC->known-merchant). You can probably go far with a per-agent baseline of typical call patterns and flag deviations rather than scoring each call independently.

Curious about DCP, never been close enough to that stack to see how they shape it in practice. The v2 question I keep landing on: does the agent identity get a behavioral fingerprint that travels with the scopes? "This agent's role is monthly payroll processor" so a fresh-address pattern fails because it doesn't match the role's normal call graph, not just because of amount. Static scope plus historical call pattern as the gate.

[agile socket]

short version: vault on the user's laptop holds keys + scopes + budgets, agents pair via signed grants and never see secrets, sensitive actions kick to a human queue (desktop or telegram). today the gate is mostly static — scope match, budget threshold, optional auto-approve under N. so it handles "amount > X" and category match cleanly, doesn't do call-graph analysis yet.
your role-as-call-graph idea is roughly where i think v2 has to land. the way i've been thinking about it: each agent's pairing grant already carries a role intent ("trading agent", "research agent", "travelling agent"), and right now that's mostly cosmetic. but role should imply an expected call graph — payroll processor making a swap to a fresh address is a role violation even when every static rule passes. the interesting part is bootstrapping: cold-start agents have no history, so you'd want role templates (canonical call patterns per role type) that the agent's behavior either conforms to or drifts from. drift gets flagged regardless of amount.
still mostly a v2 sketch in my head. happy to walk you through the current architecture properly sometime if you want — would value an honest second look from someone who ships in this space.

[normal skiff]

Yeah, would be valuable. The role-as-call-graph piece is the part I keep landing on but haven't seen anyone implement cleanly yet, so seeing how DCP actually does the static gate (and where the gaps show up would teach me a lot.

Happy to take it to DM whenever works for you. I'm in UAE timezone but flexible.

[agile socket]

sounds great. UAE — i'll DM you in the next day or so to find a time. and yeah, the gap between the static gate and the role-graph version is the most interesting part of the architecture right now, so the timing is good.

[chilly nymph]

When setting things up. Do I use the same Telegram phone number as my cell phone version or do I have to give my claw a separate one ?

[normal skiff]

Telegram has two account types and your claw should almost always use the second one. Regular user accounts are tied 1:1 to a phone number, so using yours means the claw logs in as you. What you want instead is a Telegram bot: created via BotFather in seconds, no phone required, has its own token. Most claw integrations expect that bot token.

The user-account route is only worth it if you need user-only features like joining private channels as a member or reading DM history. Even then, you'd want a SEPARATE phone, not your personal one. Telegram enforces one account per phone.

[chilly nymph]

It doesnt look like Telegram allows me to even run unless I set it up with a phone number unfortunately. So I signed up for a burner account created my account and deleted the number.

[normal skiff]

Yeah, that pattern works. Worth knowing the gotcha: Telegram occasionally requires SMS re-verification when you log in from a new IP or device, and they'll send the code to the (now-deleted) number. If you ever need to re-auth, you're locked out unless you can re-activate the burner.

Two options that age better: keep a cheap secondary SIM you can reach, or use a number-rental service (SMSPVA, 5SIM, etc.) that gives you persistent access to a virtual number for service signups. The second is what most ops setups end up with after the first lockout.

[chilly nymph]

I have a VOIP but no authorization codes seem to get sent there. Never heard of these "number-rental service". What would I google to gleen more info about this.

[chilly nymph]

Following the docs on getting iMessage going and I am still a newbie, I am not sure what or where I am supposed to make the changes for step 2 from here , https://docs.openclaw.ai/channels/imessage

[copper nova]

talk to our boy krill in #1459642797895319552

[fast sleet]

Sure

With OpenAI Codex runtime and just in general, when your message in telegram gets split into a second message, does the second message get delivered? On the pi run time with some models would reply separately or just ignore

[forest zealot]

How do you pass skill env vars (from skills.entries.*.env) to Node.js scripts run via exec? Currently I'm using node -e to read openclaw.json and extract the keys before running my scripts. Is there a cleaner way to make skill env vars available to scripts?

For production skills that need API keys (Notion, Clay, Slack), are you using SecretRef with secrets.json or a different approach? What's the recommended pattern?"

[normal skiff]

Q1: the env block under skills.entries gets injected into process env when openclaw spawns the script, so process.env.NAME in your Node code is the path. If you're reaching for node -e to read openclaw.json yourself, the script is probably being shelled-out somewhere upstream rather than through the OC skill runner.

Q2: SecretRef + secrets.json is the documented pattern but in prod two adjacent moves I've seen people make: back the .json with OS keychain (node-keytar) so the file holds references not values, or load from a real vault at startup (1Password CLI, AWS Secrets Manager, Vault). The vault path matters once you start rotating keys.

[oblique topaz]

Subject: High token cost on manual VPS setup vs 1-click Hostinger install — same model, much higher cost
Hi everyone,
I previously used OpenClaw via the Hostinger 1-click marketplace install with GPT-5.5 and had very low costs — around $1/day with 300+ messages, large workspace, many files and folders.
I then migrated to a manual VPS setup (also Hostinger) with the same GPT-5.5 model via OpenAI API, and now I'm spending $15+/day with much less usage and a smaller workspace.
What I've already tried to reduce costs:

thinking: medium → minimal
continuation-skip enabled (cache went from 19% → 87%)
Reduced bootstrap context from 43KB → 17KB
Moved non-essential files out of workspace root
Heartbeat: 1x/day with gpt-4o-mini
Disabled 7 heavy tools (image_generate, video_generate, tts, browser, canvas, x_search, code_execution)

Current state:

Each new session starts with ~22k tokens
First message after /new costs ~$0.10
Tools profile: full (32/39 enabled)
Files in workspace root: AGENTS.md (5.8KB), SOUL.md (3.4KB), MEMORY.md (3.6KB), IDENTITY.md (2.6KB), USER.md (1.6KB), HEARTBEAT.md (226B)

My question: Has anyone compared the 1-click Hostinger install config with a manual VPS setup? Is there something in the 1-click configuration that keeps token usage low that I'm missing in my manual setup?
Any insight would be appreciated. 🙏

[normal skiff]

The 15x with less usage + 22k tokens per /new is the clue. That's not 'same model' behavior. Two things worth verifying.

1. Confirm what the 1-click is actually routing to. Hostinger marketplace 1-clicks often pin to a discounted/proxied LLM endpoint, sometimes labeled 'GPT-5.5' but actually a routed mini/nano variant or a Hostinger-cached proxy. Check your OpenAI usage at platform.openai.com/usage. If your direct-API runs show full gpt-5.5 calls and the 1-click never showed up there, you weren't running the same model.

2. Audit what's in your 22k. 17KB of identity/memory files don't tokenize to 22k by themselves. Tool definitions (32/39 enabled) serialize into the prompt and each adds ~200-500 tokens. Run a request with --debug once (or tail your OC log) to see the actual payload. The bloat is usually tool-defs + memory module init, not your context files.

[shut onyx]

Can anybody tell me when will be LTS version from openclaw

[gentle lynx]

Hey all, really struggling to connect my initial bots to slack. It looks like the connection is all good but don’t know how to make it actually respond.

[serene egret]

Unrelated to the claw, but github copilot dropped their 1,500 premium requests for $38 package, anyone know of similar? Seems everything is moving to pay per usage 🤦‍♂️ claude pro absolutely sucks and hits limit in about 20-30mins

[wary flint]

Annnnnd ... what's the current consensus on the latest update?

[normal skiff]

connection good but no response is almost always one of two things.

1. Event subscriptions aren't wired. In your Slack app config (api.slack.com/apps/<your-app>) go to Event Subscriptions, enable, set the request URL to your openclaw endpoint, and subscribe to bot events like message.channels (or message.im for DMs) and app_mention. If that list is empty, Slack never tells openclaw a message arrived.

2. Bot scopes are missing or you didn't invite the bot to the channel. OAuth scopes need chat:write plus the read scope matching what you subscribed to (channels:history for message.channels, im:history for DMs). Then /invite @yourbot in the target channel.

[rustic pier]

for the guinea pigs that updated to 5.12, is it stable?

[frank agate]

Every damn release there's a new block on private networks

needs to be a global override this is ridic

[void sapphire]

Just got a system message on my MacBook Pro running openclaw that codex was stopped due to malware. I’ve never seen a message like this before. Any insight?

[normal skiff]

macOS Endpoint Security or XProtect flagging codex usually means one of two things. Either the binary's notarization or signature got invalidated (Apple revoked the cert or the cached signature is stale after an update), or your antivirus tool (CrowdStrike, ESET, Sentinel, Sophos) added a fresh signature.

First place to look: Console.app, filter for XProtect or Endpoint Security, search for codex around the time the message appeared. The log entry tells you which file got flagged and which subsystem killed it.

If it's XProtect, codesign -dv --verbose=4 /path/to/codex verifies the signature. If it's a third-party AV, the AV's quarantine log will name the rule that fired.

[long seal]

I just wanted to update from 2026.5.7 to the new version and it looks like it was nice to have a backup.

I'm running openclaw in a vm and it looks like the new version does not have SOUL.md and the other files piped in correctly. Did I miss something?

[normal skiff]

check the release notes / changelog for context-file loading changes first. OC has done schema renames before, where files that auto-loaded from workspace root in earlier versions need explicit declaration in agent.yaml or moved to a new config path.

Two debug paths worth running:

1. OC logs at startup. If there's a 'context file not loaded' or 'unknown config key' line, the schema changed under you.

2. Diff your old vs new config dirs (~/.openclaw or wherever the new install put things). SOUL.md may now need to be explicitly added to a context_files array, or live in a specific subdir the new binary scans.

If both come up clean, the VM mount path might have shifted on the upgrade. Worth checking the actual file locations the new binary is looking for.

[long seal]

I did some debugging with Krill in the #1459642797895319552 channel and it looks like a regression to me - but I'm no expert here.

//session showed the files as OK

[normal skiff]

that lines up with regression. Quick verify: in a fresh chat, ask the agent to recite a specific line from SOUL.md. If it doesn't know the content, files are loaded into the workspace but not injected into the system prompt anymore (likely a default flipped, or a new opt-in flag for context-files).

Workaround: roll back to 2026.5.7 until patched. Worth filing the issue with the //session output if there isn't an open one already.

[long seal]

Already tried that with asking. The agent started poking around and not reading files. Looks like something went kabloom there

as I have a backup, I'd be willing to debug - if that helps

I did snapshot my agent vm so I'm open to anything

[normal skiff]

yeah. The cleanest signal is the recite-from-SOUL.md test in a fresh chat. If the agent comes back with 'I don't have access to that file' or generic content, that's the regression confirmed at the injection layer. If it does know the content, the issue is somewhere else (caching, scoping, role-context split).

If the test confirms the regression, the //session output + your Krill conversation are enough for an issue write-up. Worth tagging it as a regression vs 2026.5.7 since you have the working baseline locally.

[paper elk]

Am I allowed to paste Gist links here?

[long seal]

I did the update again and noticed the workspace should be in the usual place. WebUI even shows the Core Files.

[drifting epoch]

recently, my agents started pasting their full web_search queries into discord, it's kind of annoying, anyone else dealing with this?
ALso running out of context very fast and errorring with web search, this didn't use to happen last month.

[normal skiff]

that's a tool-output regression. Web_search in OC tends to evolve its output shape across releases: what was a curated 3-snippet summary in earlier versions can switch to dumping the full SERP payload + query echo. Both your symptoms drop out of this — chat output gets noisy, and context fills 5-10x faster because each search now adds 3-8k tokens instead of ~500.

Quick check: open a fresh session, run one web_search, look at what comes back in the raw tool response. If it's the full payload, the fix is either configuring the tool's output to limit/summarize (some OC versions have a web_search.max_results or compact_output option), or wrapping it with a post-processor that truncates before it hits agent context.

[civic locust]

Useful to know thanks. I'm also on this version currently

[long seal]

issue #81955 on GH in case you want to add something

[civic locust]

Thanks! 😃

This was my attempt at trying to ask krill to scout beforehand 🤣 https://discord.com/channels/1456350064065904867/1504601302628302869

[long seal]

not sure if I'm doing something wrong here, or if it's just a regression on one provider. I'd hate to lose my agents personality

[void sapphire]

This worked. Thank you very much

[civic locust]

Yeah man, it's the result of much frustration we all build our agents 🤣

[long seal]

The best lesson on making backups

[civic locust]

Glad you had one 😃

[long seal]

Virtual machines. Learned the hard way

[fast sleet]

Virtual machines. Learned the hard way

[pseudo helm]

hello @ all

i need some help with the nostr plugin

MY ERROR: bad req: provided filter is not an object

i found the issue on github

but it seems the bug is not fixed

my Version: 2026.5.12

[uneven pivot]

Hello @all , my clawhub account seems to have been banned or disabled, and I can no longer access or find it.

I previously published two skills, but I’m not sure what caused the ban. I didn’t intend to violate any rules.

Could someone please help review my account and let me know the reason? I’d really appreciate it if it could be restored.

Account email/username: rivin-dong, [e-mail]
Skills published: prompt-eval, skill-evaluation

Thank you very much for your help.

[tawny walrus]

My ai is refusing to do what it said it could do.
I asked it what his browser abilities are and it said it can open a new browser tab so I asked him to open a new tab with a website like google maps and it actually said only "NO", whattt on earthh

Goal:
I want my ai (openclaw) to have full pc or browser access, it should help me generate leads or find big businesses in my country
but it already refuses to open a new browser tab like i dont understand this

(note: security doesnt matter to me, i edited the openclaw.json to have full pc controll and never ask for perms)

**Im writing here because i get errors when posting in #1459642797895319552 **

[mighty rampart]

What model are you using? Computer use like operating a browser will need a very capable model.

[tawny walrus]

Openrouter + qwen + ollama

[slate estuary]

Looking for beta users. Cerone agent governance runtime for OpenClaw. Identify, control and validate your OpneClaw agents..No need to signup, 2500 free validations .

https://github.com/AnantDhavale/cerone-openclaw-plugin

[sinful quarry]

Hey all — hoping someone can help me troubleshoot. I'm hitting what looks like a hard ~4,000-character limit on assistant replies in the WebChat / Control UI. Anything longer gets silently cut off mid-stream with no warning on either side.

I've tried raising gateway.webchat.chatHistoryMaxChars and channels.stream.maxOutputChars in my config — neither one fixed it. From poking around, it looks like the cap is set inside the compiled channel bundle itself.

Has anyone else run into this? Is there a config setting I'm missing, or a workaround that's worked for you? Right now I'm having my assistant write long output to a file and just paste the path in chat, which works but isn't ideal for everyday use.

Any pointers appreciated.

[candid maple]

I have a question please

I try to create an agent workspace and a skill with a script. This Python script uses Opanai to make some prompts

like this example

from openAI import OpenAI
client = OpenAI(api_key=os.environ.get("OPENAI_API_KEY"))

When I run a test, OpenClaw told me he can't find any OPENAI_API_KEY

My openclaw is configurated to use openai as main model, i want khnow how to passe the open ai key directly to my python script or use direct openai model context as llm inside my script since that already openclaw use openai as main model

[rotund obsidian]

Hello claw contributors! I have a question that may be easy for some of you, “How do I disable streaming for a custom openai-completions provider? GPT4All returns 400 ‘stream’ is not supported.”???

[indigo nebula]

Why not setup another better channel like WhatsApp or Telegram? The web chat interface isn’t all that great for interacting with your claw

[short marten]

My openclaw is chatting with me, its showing hearbeat.md

[solemn widget]

You need to set the api_key separately. Think of it like a real key, your script cannot go ask openclaw to borrow the key. You need a second copy of the key. You can either set it in your shell or in a .env file and use pythons dotenv to read it. Just tell your claw what you are trying to do and it will help you

[pseudo helm]

i fix it somehow , chatgpt made a patch for it 🙂 and now the nostr plugin is up and runing. but DMs are not working

[white mauve]

with 2026.5.12 since codex is taking over the agent runtime, should drop this from config?

      "embeddedPi": {
        "executionContract": "strict-agentic"
      },

[raw blade]

anyone able to assist with a tricky remote node issue? I cannot get my gateway to work with another pc node

[torpid grotto]

Updated to 5.12 and i can't even run a single openclaw cli command, freeze forever with no response

[junior robin]

yeah I get that too for the cli the web ui tends to work better

[torpid grotto]

I killed processes so no way to interact at all

[earnest flint]

My openclaw forgets everything i say. I use local ai models and i tried multiple, but every model forgets everything.

[torpid grotto]

Ok, removed all provider's models from config, and it started working better

[torpid grotto]

Mb low context size?

[raw blade]

Yo! can anyone guide me to information regarding setting up remote nodes properly? I need to connect them to a gateway, THX

[scarlet silo]

hi guys, is it safe already to upgrade from 4.23?

[junior robin]

Upgrading to 5.7 was nice and easy from 4.23 for me. 5.12 feels like a set back

[scarlet silo]

hum, why?

[normal skiff]

5.12 has been rough today, might want to hold off for now. nikon above hit a cli freeze on it, kingvu's losing local-model context after upgrading. tyrantjunior's 4.23→5.7 path looks like the safer landing zone.

[junior robin]

They switched over to codex instead of pi for the agent run time. Responses seem a lot slower and my MCP server stuff doesn't appear to be working correctly. If youre not using openai models it might be fine

[scarlet silo]

im all in on openai. damn.

[junior robin]

working with kill right now to see if I just need a lot of configuration settings to be updated and changed or have to reinstall context mode, serena, etc

[normal skiff]

codex runtime is openai's model line so that part should actually work for you PT. the cli freeze nikon hit and the mcp regression tyrantjunior flagged are separate, not provider-tied. how heavy is your mcp setup? that's where 5.12 still looks brittle.

[scarlet silo]

i think it will be a nightmare to upgrade with so many config changes...

[junior robin]

I might just wait for 5.14 and pray it fixes codex with my mcp stuff and chat responses

[scarlet silo]

you had 5.7 working fine? good speeds?

[junior robin]

very tiny problems that was easy for me to fix but yeah everything felt fine

I think your main problem might be extensions since 4.23-5.7 they changed how extensions worked like lossless claw but again easy fix at least for me

[scarlet silo]

Thanks. My setup is relatively heavy on the MCP/OpenClaw side.

I’m running OpenClaw in Docker on a NAS, with a Windows remote node attached for browser/control tasks. The main gateway has several integrations enabled: browser tooling, Discord, Telegram/WhatsApp channels, memory/QMD/LCM components, cron jobs, subagents, and multiple workspaces/subagents for coding, sysadmin and investment workflows.

So yes, my concern is less the OpenAI/Codex provider side and more the MCP/tooling/runtime layer. I’ve seen issues around browser/node attachment, plugin loading, channel plugins, memory backends and cron/tool reconciliation. That makes me suspect I may be in the category where 5.12 could still be brittle.

[normal skiff]

your instinct's right. the surfaces you listed (remote node attach, plugin/channel load, memory backends, cron reconciliation) are exactly the layer 5.12 is shaky on, same thing tyrantjunior hit. provider side is your safe part.

for a stack this integration-heavy i wouldn't upgrade in place. stand 5.12 up as a second gateway on a throwaway workspace, smoke-test in order of fragility: remote node attach, then plugin/channel load, then memory backend, then cron reconciliation. keep prod on 5.7 until that's clean. in-place is what turns a 20-min test into a day of firefighting.

[scarlet silo]

exacly. dont have that much time atm.

[bleak copper]

What are the best practices when updating OpenClaw to ensure it doesnt break it, which seems to happen to me alot.
Im running OpenClaw on Live USB, a bootable persistent (p_static_root) MX linux thumb drive (256GB). I take frequent snapshots and backups, but it seems like im breaking the config way to often when I update.

If i click the update notification in the UI while there are unsaved config changes would that break it?

is it better to update via command line and not in the ui?

Is it ok to update while the gateway is running or should I close it first and update via command?

any advice is appreciated. ty

[mighty rampart]

Have an AI agent study the Changelog and recently opened issues and read comments here to decide what version to update to. Never just move to the latest version without checking.

[sinful quarry]

Thanks, I'll give it a try

[junior robin]

oh yeah definitely looks like the mcp stuff is getting fixed in 5.14 https://github.com/openclaw/openclaw/pull/81551 hopefully this is it

[paper elk]

with 2026.5.12 upgrade looks like my Codex setup is broken; was working in 2026.5.7

my agent Mikhail's cron Discord deliveries fail because OpenClaw's embedded_run delivery path doesn't resolve the bot token SecretRef from secrets.json before sending.

[vernal wave]

Anybody else's skills stop working after the v2026.5.12 update?

[topaz vigil]

Machine is a ROG flow z13 (Strix) 128 gb unified ram win 11 pro. 98 gb dedicated to gpu. Lemonade as the back end, running vulcan and ROCm. every turn compaction is triggered. but on my steam deck no issues. yes steam deck. My gaming handheld handles openclaw better than my beast. please fix and or explain. @sleek shoal

[junior robin]

i've noticed that my skill references are being read more with 5.12 with the codex runtime

[echo nova]

I'm curious if anyone could tell me... is it critical to basically put some critical elements in the soul.md... when created an agent? If you don't do that, what do you expect?

[chilly nymph]

I have successfully paired whatsapp with my Claw, in WhatsApp do I just send a message to myself ? (The phone number that is in the config file)

[normal eagle]

Is anyone else using discord repeatedly getting the message "Agent couldn't generate a response."? It's like 50%+ of all of the messages I'm getting in discord sometimes. My agent keeps telling me it's getting fixed, and it kinda works for a bit (couple hours maybe?), then just comes back.

[normal eagle]

Depends on what you consider critical, imo. Everything is based on context, so if you're building an app that compiles color schemes, for example, I wouldn't say it's "critical" to talk about that time in 4th grade when your zipper was done....unless that event is necessary context for understanding why "green" and "yellow" go together.

[topaz vigil]

22311

[bold cloak]

Like this?

“⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session.”

Mines been doing that any time it wants to call a tool to do something. It does it with all the models I have configured through openrouter.

[normal eagle]

Pretty much. Mine says this: ⚠️ Agent couldn't generate a response. Note: some tool actions may have already been executed — please verify before retrying.

Exceptionally frustrating. It's like 10+ messages a day minimum, and way worse when I'm actively on it and working.

[frank agate]

Can openclaw already talk / respond in voice chat or no?

[signal star]

Has anyone noticed that Claude-Cli has recently stopped automatically renewing the token? I installed it about a month ago and it was renewing my token automatically. But for the last few days, it hasn't been renewing.

[patent cape]

⚠️ Agent failed before reply: Unknown model: qwen/qwen3-coder:free. Found agents.defaults.models["qwen/qwen3-coder:free"], but no matching models.providers["qwen"].models[] entry. Add { "id": "qwen3-coder:free" } to models.providers["qwen"].models[] to register this provider model.
Logs: openclaw logs --follow

i tried to use the qwen model, but this is what is being shown to me

the krill bot told me to do some stuff, and am kinda confused on why is it still not working

[wraith seal]

okay so

[patent cape]

if anyone can just give this a read: https://discord.com/channels/1456350064065904867/1505191074195439727

[wraith seal]

either way, i would NOT use it

Venice with that model is unreliable

use openrouter free router

it automatically routes to an available free model

[patent cape]

but i wanna use the qwen3-coder free model

[wraith seal]

IT ISN'T

[patent cape]

ion wanna use any of the nemotron free models

[wraith seal]

GOOD

[patent cape]

ohhh

[wraith seal]

https://tenor.com/view/madagascar-skipper-can-you-read-gif-22646390

[patent cape]

ahhh

so should i use deepseek v4 flash: free?

[wraith seal]

nemotron is better actually

[wraith seal]

sure whatever kid

whatever floats your boat

[patent cape]

uh rude, but wouldnt it be the same error all over again?

[wraith seal]

no

[patent cape]

what should i put in the model name then, after i enter the openrouter key

[wraith seal]

deepseek/deepseek-v4-flash:free

[patent cape]

⚠️ Agent failed before reply: Unknown model: deepseek/deepseek-v4-flash:free. Found agents.defaults.models["deepseek/deepseek-v4-flash:free"], but no matching models.providers["deepseek"].models[] entry. Add { "id": "deepseek-v4-flash:free" } to models.providers["deepseek"].models[] to register this provider model.
Logs: openclaw logs --follow

same error

[wraith seal]

bro

it's literally in the error message

YOUR ANSWER

IS IN THE MESSAGE

[patent cape]

its working now

[wraith seal]

teehee :)

[bold cloak]

My issue was resolved with the beta build that came out this morning

[thick kindle]

Does the dashboard have a list of plugins somehow?

[rotund obsidian]

How do I disable streaming for a custom openai-completions provider? GPT4All returns 400 ‘stream’ is not supported. Do I turn off stream? If so, how?

[round pagoda]

Guys is openclaw capable to autonomously click around on the gui of the ubuntu vps i created for openclaw? Tried a lot but never made it work.. using wayland as display manager.

[raw blade]

how do I make it so that meduim thinking is the default?

[covert kernel]

https://techbroiler.net/how-ai-agents-actually-work-today-and-the-problem-they-all-share/

[minor trout]

Hey all. Trying out openclaw this week. Its been a roller coaster of ups/downs but so far I like it.

I found the agent is having a problem with scraping the web. I asked it to go research some products and it was being blocked by google, etsy,amazon, every seller.

I understand why but I was wondering if anyone has found a work-around. I'm in a vanilla setup right now and using openrouter/ds-v4-flash

[junior robin]

Edit your openclaw config file
https://docs.openclaw.ai/tools/thinking
"agents": {
"defaults": {
"thinkingDefault": "medium",
}
}

hmm i've been noticing the agents keep stopping a lot mid session and I have to keep prompting it to finish the task since I've updated to 5.12

[mighty rampart]

What are you trying to do? See if there is an API for the data you are trying to get. Computer use (clicking around the browser) is possible but should be the last resort.

[green glen]

Hi, a friendly user helped me fix the missing_scope issue. I uninstalled and reinstalled openclaw, still doesn't nothing.

[minor trout]

I asked it to find me transparent sticky notes with acid-free adhesive

[thorn pulsar]

has anybody paired slack with openclaw

[mighty rampart]

Yes. With a slack app and in socket mode.

[indigo nebula]

Tell your claw to download this and use this for accessing the web. Does a far better job at accessing sites, won’t help solve everything but should help you. https://github.com/pinchtab/pinchtab/

[abstract silo]

Hey, how can I setup openclaw to do some kind of web scrapping

[ebon kernel]

hii :3

im trying to deploy openclaw in a closed enviroment in nobara using ollama in a laptop with an i3 and 16 gb of ram ddr4 a 3600mhz

but i cant

[south trail]

how exactly would I configure my windows openclaw install to go through my tailscale?

every guide I've found either doesn't go into detail or is for a vps

[south trail]

ok i think i got it connected now but I'm just trying to get it secured so i can connect my damn phone to it

{
  "gateway": {
    "auth": {
      "mode": "token",
      "token": ""
    },
    "controlUi": {
      "allowedOrigins": [
        "file://"
      ]
    },
    "mode": "local",
    "tailscale": {
      "mode": "funnel"
    }
  },
  "wizard": {
    "lastRunAt": "2026-05-12T20:58:21.468Z",
    "lastRunVersion": "2026.5.7",
    "lastRunCommand": "doctor",
    "lastRunMode": "local"
  },
  "meta": {
    "lastTouchedVersion": "2026.4.23",
    "lastTouchedAt": "2026-05-12T21:35:03.656Z"
  },
  "commands": {
    "restart": true
  },
  "tools": {
    "profile": "full",
    "sessions": {
      "visibility": "all"
    },
    "exec": {
      "security": "full",
      "ask": "off"
    }
  },
  "browser": {
    "enabled": true,
    "defaultProfile": "openclaw",
    "ssrfPolicy": {
      "dangerouslyAllowPrivateNetwork": true
    }
  },
  "session": {
    "idleMinutes": 10080
  },
  "skills": {
    "entries": {
      "pdf": {
        "enabled": true
      },
      "xlsx": {
        "enabled": true
      },
      "docx": {
        "enabled": true
      },
      "pptx": {
        "enabled": true
      },
      "find-skills": {
        "enabled": true
      },
      "self-improving-agent": {
        "enabled": true
      },
      "tavily-search": {
        "enabled": true
      }
    }
  },
  "plugins": {
    "entries": {
      "browser": {
        "enabled": true
      }
    }
  },
  "models": {
    "providers": {
      "custom-custom94": {
        "baseUrl": "https://nano-gpt.com/api/v1",
        "api": "openai-completions",
        "models": [
          {
            "id": "moonshotai/kimi-k2.5",
            "name": "moonshotai/kimi-k2.5"
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "custom-custom94/moonshotai/kimi-k2.5",
        "fallbacks": []
      }
    }
  }
}```

here is my config file

[solemn meteor]

Does anyone know how many tokens the first “Wake up, my friend!” Prompt takes on opus 4.7?

[solemn meteor]

Where are you running openclaw?

[stoic echo]

hey Someone can help me to install the Ollama cloud models on my OpenClaw setup? The installation is installed in Hostinger via Docker

[latent remnant]

OpenClaw:

From what I’ve read, OpenClaw was mostly developed using AI tools

So, if I understand correctly, the founder initially coded it using Claude CLI, manually clicked “accept” on each Claude request, and then gradually built OpenClaw itself, eventually using OpenClaw to continue developing OpenClaw

Is that the actual story?

If so, does that mean that if we give OpenClaw a complete specification document from A to Z, and it breaks it down into tasks, then assign multiple agents to those tasks, each with a specific specialty, plus sub-agents with even more precise specialties, we could theoretically develop an entire project from start to finish?

Is that really possible today? Has anyone already done this successfully?

I was thinking about building an agent that reads a full specification document, breaks it down into Agent/Human tasks in Monday, and then, once that is done, having Monday trigger events whenever a task is completed

Those events would then launch the right development agents, each agent would work on the code, commit its changes to GitHub, and another agent would review and validate the commits

Finally, a human would review the overall result

What do you think about this approach?

Do you already have an efficient process for this kind of workflow?

[finite flower]

Possibly a dumb question. Previous to recent updates my OC could access files on a local mounted drive which it could edit - docker compose, code, db migrations etc - it seems it is now super sandboxed and it can’t see the files. Is it just a matter of adding mount bindings or. something else?

[south trail]

On my Windows PC, using ClawX

I mention that I'm on windows in the question

[civic locust]

It's not a dumb question I mean it doesn't work anymore for you right

I don't know the answer but definitely not a dumb question

[tired crest]

Hey I updated to version 5.12 and now i am getting this: [assistant turn failed before producing content] with all my agents is there fix?

[frigid oyster]

What is everyone using for a rollback version since 5.12 is trash?

[mighty rampart]

Yes, the things you describe are possible. Read more about “Software Factory” and “Ralph Loop”. If you want to do it, I would advise trying it with Pi Agent outside of OpenClaw first. (Pi is what is underneath OC.) Once you understand all the moving parts, then shift it into OC.

[fleet cloud]

I have configured acp on my client but when i enabled acpx i start sseeing ACP error (ACP_SESSION_INIT_FAILED): ACP metadata is missing for agent:standard:acp:c1b85bb8-66da-4f00-92cf-e35eb94001da. Recreate this ACP session with /acp spawn and rebind the thread.
next: If this session is stale, recreate it with /acp spawn and rebind the thread.acp entry in opencclao "acp": {
"enabled": true,
"backend": "acpx",
"defaultAgent": "standard",
"allowedAgents": [
"standard",
""
], and plugin entry look like this "acpx": {
"config": {
"permissionMode": "approve-all",
"pluginToolsMcpBridge": true,
"openClawToolsMcpBridge": true,
"agents": {
"standard": {
"command": "openclaw"
}
}
},......can some help configure acp and acpx on openclaw

[untold moss]

some say /v2026.5.16-beta.5 fixes a lot of issues

[potent cave]

I asked OpenClaw and it doesn't seem to figure this out.

Does anyone know how I can get OpenClaw to spawn a subagent with "/fast on". Not as a subagent with a hard configuration for fast mode, but an optional prompt to say something like "ok, for this one, spawn a subagent with fast mode".

This is for Codex fast mode

[blissful ridge]

Hi very new (in setup stage) but it seems to get stuck in the Starting setup... stage its been in that stage for about 30 min used the powershell -c "irm https://openclaw.ai/install.ps1 | iex" command and it downloaded no problem and then it just stays in Starting setup... is it normal for it to take that long or might i be doing something wrong? I seem to see that 5.12 is not great should i just try an older model?

[torpid patio]

guys, how can i pass these variables so gateway picks them up properly in docker?

OPENCLAW_GATEWAY_PORT=18789
OPENCLAW_GATEWAY_BIND='lan'
OPENCLAW_GATEWAY_MODE="local"
OPENCLAW_GATEWAY_TOKEN='eac7949646647ca0e5ceba3b8f783d24f31231231231231269b932ea35d2ef'
OPENCLAW_CONTROL_UI_ALLOWED_ORIGINS='https://openclaw.domain.com'

i feel that the entire documentation, and google results are all ai slop

[mighty rampart]

What have you tried already that didn't work? How do you normally start the container? The simplest way is docker run -e VAR1=val1 -e VAR2=val2 the-image

[torpid patio]

services:
openclaw:
image: ghcr.io/openclaw/openclaw:main
container_name: openclaw
cpuset: "1,2,3"
mem_limit: 2g
memswap_limit: 2g
restart: unless-stopped
hostname: openclaw
stdin_open: true
tty: true
ports:
- 18789:18789
environment:
OPENCLAW_GATEWAY_MODE: local

I tried to pass those varaibles in all possible ways, at startup it simply does not pick them up

DR says

◇ Gateway ──────────────────────────────────────────────────────────╮
│ │
│ gateway.mode is unset; gateway start will be blocked. │
│ Fix: run openclaw configure and set Gateway mode (local/remote). │
│ Or set directly: openclaw config set gateway.mode local │
│

i consider myself quite skilled, and i still cant get it sorted, the gateway process starts with zero config no matter how i pass the env vars

in cotainer the vars are picked by the "os" but the gateway process does not pick them

command like this work fine

openclaw config set gateway ....
but i was wondering if i can set the config via env

[mighty rampart]

Do you have Claude or Codex? Point one of them at the source code and ask. mine is saying you can set the token that way, but
No env var support for that one [OPENCLAW_CONTROL_UI_ALLOWED_ORIGINS] — allowedOrigins is only read from the config file (gateway.controlUi.allowedOrigins), not from the environment. (as of 4.24 anyway)

[torpid patio]

thank you for trying to help me

[trail gazelle]

i did sent u a dm u can check maybe it helps u 🙂

[torpid patio]

I'll check now, I'm not very familiar with discord

[violet temple]

best model for rtx 4090?

[torpid patio]

here you go

https://apxml.com/models?modelType=open_weight&sort=rank_coding&selectedGpu=4090_24&runType=inference&quantization=q4&numGpus=1

you need to leave headroom for context processing at leat 30% free

[untold holly]

Hey guys, I'm having an issue running openclaw after update.
The error I'm getting is:

run error: Cannot find package 'openclaw' imported from    
/home/assistant/.openclaw/npm/node_modules/@openclaw/codex/dist/shared-client-Cr6W-a2G.js

It seems like the codex dependency was not installed when the npm install -g openclaw was run. Why could this be happening? How can I fix it?

[mighty rampart]

Have you already done openclaw doctor --fix ?

FWIW I just upgraded to 2026.5.12 (from 4.24) and openclaw doctor --fix got it all working again. The plugin architecture has changed.

[steel topaz]

I hate to ask for help, but I am not sure what else to do. I have installed open claw, I am able to connected it with my locally hosed llm, but its seems like the agent is kinda dumb and is not able to access, read, edit, or write any files. Is there somthing I need to do to enable this functionality?

[junior robin]

local LLM are kind of dumb unless you got a beefy setup. For the read, edit, write you have to make sure it has all the permissions it needs some of it is through the openclaw config file

[mighty rampart]

What local model are you using on what hardware? Not all models support tool calling, which is what you need to write files, etc.

[steel topaz]

devstral24b, which one do you recommend?

[steel topaz]

When you say the openclaw config file, do you mean the openclaw.json?

[junior robin]

yeah

[mighty rampart]

That ought to be okay. What hardware and OS? How are you serving the model? Does the model work outside of OC like with Pi Agent on the command line or in Open WebUI?

[steel topaz]

I honestly have been trying to get a agentic model to work locally for a couple weeks now ( have tried several models and several tools, all with olama on my local host ) I have 4090, and 120 gb of ram, so I know I have the hardward to run a decent model just can't run the huge ones ( even though I reaally want to ) I am on windows, so what it will work on .... I guess none? I but I am still exporing how to make all this work.

[steel topaz]

Wheree can I find documentation on what to put in the json so it has the permissions? or is there a command with the flag. I havn't found anything to tell me this by reading the docs yet.

[jaunty halo]

Hey is anyone able to help with this issue im getting

I have tried many different things to fix it and it seems nothing is helping I am giving my openclaw a task to create me a few agents with system prompts and it seems every time I give it that task and then try talking to it after it gives me this message "⚠️ Something went wrong while processing your request. Please try again, or use /new to start a fresh session." would love some help

[mighty rampart]

That’s 24GB of VRAM correct? I’m running Gemma4:27b on an old Titan RTX with 24GB but on Linux, I do not know about Windows.

[steel topaz]

I can give it a go!

[mighty rampart]

What model are you using and where is OC running? Find the server logs and see what it says went wrong.

[junior robin]

honestly I have my ai agent handle that. You can also run the /exec command in the TUI and change the security settings to give it more permissions. I have /exec security=full but that's scary if your AI model isn't too capable or really know what its doing

[jaunty halo]

I was using MiniMax 2.7 through OpenRouter but after setting a fallback everything seems to be fine for now hopefully it stays this way

[tacit surge]

After upgrading openclaw from 2026.3.28 to 2026.5.12, every cron job got the error: isolated agent run stalled before execution start(last phase: context-engine). Has anyone met the same?

[steel topaz]

one of the major problems I ran into was my npm needed to be updated to install dependancies. IDK if that is related to your issue, but it prevented my tool calling from running.

[tacit surge]

IDK. I have run openclaw doctor --fix but haven't seen related caution.

[steel topaz]

I was running that, but because of my NPM version not being updated the openclaw doctor was failing

[nocturne zephyr]

If you're running Windows, you really want to try to do WSL2 as your environment for OC. A 4090 is decent, and you can probably run Qwen3.6-35B-A3B on llama.cpp in q4 with MTP and a decent context buffer. It's not going to be super-powerful, but with that setup, I'd say ~140 t/s is possible. That gives it some decent ability to do 'reasoning' and not feel like it's taking too long. Edit: and it's multimodal which is pretty kick-ass.

[solemn meteor]

How is it? And how is OR’s pricing on the model? I tried k2.6 and v4 pro and felt like their lying about rates

[copper lake]

Hi everyone I’m looking forward joining the test app for iOS for open crawl but i couldn’t find anything how can i join the testing?

[cursive trench]

Is their a "download" link

[indigo nebula]

Go to https://openclaw.ai/ part way down the page is a Quickstart command to copy/paste for the given OS you’re going to install on.

[scenic flume]

My OpenClaw keeps hitting Codex rate limits, even though I'm not coding anything, but it seems like you can no longer run ChatGPT without Codex via 0auth. Is that correct?

[copper lake]

For iOS?

[indigo nebula]

iOS phone and iPad version is not available yet.

[copper lake]

Thank you shane

[plush tide]

You’d be amazed how many tokens a simple chat can use. I tried using my chatgpt sub with OC and got about 5-6 messages in before hitting the wall. Once I switched to Claude Max 5X, it’s been better. But I still hit limits if I do other tasks outside OC.

[scenic flume]

Ive worked hours and hours via chatgpt plus until recently. Now after a few min I run into a codex limit allthought I am not coding at all. Thats entirely new. Something has dramatically changed and I hardly can use Openclaw via oauth anymore.

[mint tulip]

$20/m gives you about 30M tokens every 5 hours through Codex after looking at it with codexbar

[plush tide]

It’s hard to diagnose but one thing to remember is the longer we use the agents, the more memories, tool use and other crap is being sent in context for each turn.

I built a small dashboard to measure where the bloat was and even after the deep dive I concluded that it’s just the nature of these harnesses. They aren’t efficient users of tokens yet.

[wary flint]

All the more reason to /compact as things move on. That and address small tasks independently after outlining a big picture plan.

[digital citrus]

Anyone have a good pattern for working with long running opencode runs? Our team uses opencode so I was hoping my claw friend would be able to use opencode during Dev cycles as well. My issue atm is my opencode pipeline can take around 10mins to return. This causes gateway managed ACP one shots time out. And doing an exec sometimes doesn't wake the discord thread again.

To complicate things it seems like opencode cannot set an acpx timeout and returns: Invalid params (ACP -32602). The adapter may not implement session/set_config_option, or the requested value is not supported. <- RequestError [-32602]: Invalid params

I'm wondering if I should use cron here or if I'm missing something about how wake on exit should works I currently have:

"exec": {
  "notifyOnExit": true,
  "notifyOnExitEmptySuccess": true
}

the behavior love to achieve is to be able to dispatch opencode bg runs from discord without having to dispatch an agent spawn that basically does the exev online and just waits on opencode to finish.

Any help/pointers would be awesome!

[lean mantle]

howdy yall! can someone send me some good resources on agent optimization? specifically interested in learning how to be economical with Claude usage tokens

also, I havent messed with Openclaw stuff in several months. any cool new capabilities now?

[mint tulip]

caveman plugin?

[junior robin]

My stack is graphify, serena, and context mode. Caveman if you want to min max the tokens. Caveman saves only a really tiny bit of the output tokens.

graphify and serena are more code focused. context mode is good well rounder

though I am struggling to get serena and context mode working correctly in the new version of openclaw.

[mint tulip]

what does graphify do?

[junior robin]

https://github.com/safishamsi/graphify basically maps your entire project files so its easier for your ai agent to understand your repo better. So it's not wasting time trying to figure it out

[mint tulip]

ty ty. finally have time to tinker with my agents, this is super helpful

[junior robin]

not sure how many tokens it actually saves in practice

but after doing some testing serena and context mode definitely saves tokens

Oh I guess https://github.com/rtk-ai/rtk could help as well but I haven't used it to be sure since it doesnt work well on windows natively

[weak flint]

Still trying to set up openclaw gateway on gcloud... I am curious why Tailscale is recommended to serve the gateway UI over HTTPS?

[steel topaz]

Well, I am still figuring out how to set all that up.

[junior robin]

anyone only getting codex_apps when you do /codex mcp I still haven't been able to get my mcp stuff fully online and working properly

[steel topaz]

aye, this is why I am messing with open claw, I couldn't get codex, claud, and a few others to work either. ( I mean beyond just being a boring chat ) I want this to do stuff!

[junior robin]

i've been dealing with this bug for days and the new updates should have fixed it. definitely very frustating.

but I think codex and claude just fixed it but idk i'll have to see if my agents are using context mode and serena like how it was before on PI runtime

[junior robin]

dam unlucky still not working i'll probably just use codex directly until its fixed

[scarlet silo]

@junior robin do you notice increased speed in later release?

[junior robin]

I eventually fixed the mcp issues I had I dont think its fully working on the native releases though. For speed not sure its definitely working better than before I had to keep prompting it to finish tasks I'm on 5.18 now

https://isitstable.com/openclaw might be helpful whenever you plan on updating

[steady sorrel]

This is a good point. We had a crash yesterday based on the length and heaviness of the discord communication with several agents.

The system crons started to fail. Then model drift. Then full rick roll no answer nonsense.

We implemented a nightly reset at 4 am after the 3 am memory save.

I like the dashboard idea. What specifically are you tracking the length? Context? Last reset?Tokens? or all of the above.

[severe sky]

Hello, I'm trying to set it up a couple of times and the latest hurdle reads like this. Any help.
auth or provider access failed for anthropic. Run /auth anthropic to refresh credentials; if you already re-authed, switch models/providers
because this account may still be blocked for inference.
local ready | error
agent main | session main | anthropic/claude-opus-4-6 | think adaptive | tokens ?/200k

[plush tide]

The dashboard tracks my Claude Code token usage across:

• Sessions — table of top threads with started/last-activity timestamps, total tokens, and cost. Click a row to drill into a per-message timeline (model, input/output tokens, cache reads/writes, cost per turn).
• Projects — same metrics rolled up per project; click through to that project's top threads.
• some nice Charts using Apex Charts.
• Top KPIs show tokens today, cost today, and last update, with a live connection indicator in the header.

I spent a couple of days working on it with Claude Code till I was happy. Classic little vibe code project. But it made for depressing viewing! 😂

[ivory prism]

I've just built the android openclaw app since the one in the play store is outdated and has a protocol mismatch. The one I bult is the newest version. It connects via tailscale to my PC/gateway but it says: connected (operator offline)

openclaw logs says this:
14:08:43+00:00 warn gateway/ws {"subsystem":"gateway/ws"} {"handshake":"pending","durationMs":16,"host":"tony-pc-cachyos.taila4a2aa.ts.net","userAgent":"okhttp/5.3.2","forwardedFor":"100.109.98.6","remoteAddr":"127.0.0.1","remotePort":48538,"localAddr":"127.0.0.1","localPort":18789,"endpoint":"127.0.0.1:48538->127.0.0.1:18789"} closed before connect conn=f92944d4-2e90-437e-8ecb-6a9ffd38da0e peer=127.0.0.1:48538->127.0.0.1:18789 remote=127.0.0.1 fwd=100.109.98.6 origin=n/a host=tony-pc-cachyos.taila4a2aa.ts.net ua=okhttp/5.3.2 code=1000 reason=bye

[proud owl]

the latest openclaw version on google play store is from April and gives me 'protocol mismatch' errors on latest openclaw versions. Can I get a newer .apk from somewhere without building it myself?

[latent hedge]

Yo! Anyone know what the right config for kimi k2.6 is for the subscription based kimi-coding is?

I used to use 'kimi/k2p5'.. when reading the docs.. it says 'kimi/kimi-for-coding' - but that doesn't work?

[low tendon]

openclaw builtin memory systems are extremely confusing and I never know if I'm doing right, am I the only one in this situation?

[bitter spoke]

Budget is about $3000 for the whole setup
Anyone got any component recommendations for this price range?

I was thinking about a RTX 3090 as the core component of the setup

[onyx basin]

hello everyone! could you please recommend the best open source model for openclaw please? The model that I can run on my own GPUs(H100/B300 etc)

[junior robin]

probably qwen? Maybe deepseek v4 flash if you have enough

[onyx basin]

yah just want to hear feedback about these models

[rocky orbit]

You can try Kimi as well

[junior robin]

my friend runs qwen 3.5 the 27b or 36b model on his agent but its not openclaw and he thinks its fine. I was running the 9b model and I did not enjoy the experience

isnt kimi like a lot of gpus

[onyx basin]

I use it today and have very good experience with that, especially Kimi 2.6. The only concern that it's kinda hard to run it with the good perforformance on 2xH100 server

[junior robin]

kimi is probably the best one at there looking at bench marks

[onyx basin]

so idea is to try deepseek v4 flash, is it worth it?

[rocky orbit]

Yes pretty hard

[junior robin]

top open source models seem to be kimi 2.6 , glm 5.1 , and deepseek v4 right now

so see if you run any of those

[civic locust]

+1 for glm-5.1

[lyric bay]

Hi all, i'd REALLY like my openclaw to be able to setup a cool discord server channel setup to track different tasks, but the standard skill does not allow for it

[junior robin]

cant you just create your own custom skill and tailor it to what you want for the discord setup?

[lyric bay]

I've been talking to my openclaw about it for 15 chats now, and it keeps just making documentation. eek

its running on Grok 4.3 fast

maybe im missing the point. not sure

[frank agate]

Been happy with Qwen3.6-35B-A3B-UD-Q6_K_XL

[cinder wadi]

Telegram channel appears broken in OpenClaw 2026.5.18 on Windows.

Symptoms:

• Telegram token validates
• openclaw status --deep shows Telegram OK
• Gateway healthy/reachable
• Bot responds nowhere in Telegram

Logs repeatedly show:

• deleteWebhook failed (404 Not Found)
• deleteMyCommands failed (404 Not Found)
• setMyCommands failed (404 Not Found)
• Telegram command sync failed: GrammyError

Already attempted:

• regenerated bot token
• removed/re-added telegram channel
• downgraded to 2026.5.16-beta.7
• repaired gateway
• ran doctor --fix
• verified token source/env
• restarted gateway repeatedly

Looks like Telegram command sync is crashing the provider startup loop.
Anyone seen this?

[candid maple]

I have a question who have any idee i want create an agent who i can give them a csv file and he start depend of this file insert data row by row to a supabase table

and after that i can ask agent about information come from this table

[frank agate]

has anyone gotten image gen + image editing working w/ comfyui and a local model ?

not ideal options: ```How should the single comfy image workflow slot be configured?

❯ 1. Switch to edit workflow
Set comfy to the validated Flux.2 edit workflow. Image editing (the failing use case) works. Pure text-to-image generation via openclaw stops working (no input image to load). Recommended — your actual requests are edits.
2. Keep text-to-image only
Leave the current t2i workflow. Generation works, editing stays broken. No change from today.
3. Edit workflow + dummy placeholder
Use the edit workflow and place a 1x1 placeholder image in ComfyUI's input dir so no-input generation doesn't hard-crash. Generation output will be low-quality/unreliable, but editing works and generation won't error out.```

[nimble hill]

This is a simple user case. Ask openclaw to create a python program to convert from czv to table. Then create s skill for the same.
This is much easier done in n8n which will not even use lllm

[chilly nymph]

I constantly have to kick iMessage in the pants to get it to send me messages. Is it just me or is this a known expectation ?

[paper plover]

I had iMessage working via Bluebubbles. A recent update eliminated support for it. They moved to a native solution which works great. The problem I had was I didn't realize the change and neither did my agent. It kept trying to come up with workarounds to make Bluebubbles work. I switched my middle from Deepseek chat to Deepseek Reasoner and it found the problem. Switched itself to the new native setup and it's been running great for 3 days with no issues. I would check if yours missed the change in the update like mine did.

[chilly nymph]

I am not using Blue Bubbles. I am using the native with imsg from brew.

[steel topaz]

If I want to generate images with open claw do I have to use open router? I am trying to do everything locally? I am getting [tools] image_generate failed: OpenRouter image generation failed (HTTP 401): Missing Authentication header raw_params= bla blah blah

[celest drum]

I’m building an OpenClaw + Obsidian second brain system, and the system is built so far to take podcast transcripts and turn them into useful notes based on business context, strategy, marketing, operations, and other topics. However, the real bottleneck right now is speed of ingestion. I’m trying to bulk scrape transcripts from 20+ podcast and YouTube channels, some with hundreds or thousands of videos, and my current scraper works but gets slowed down by YouTube throttling, missing captions, cookies, transcript limits, and rate limits. I’m also trying to figure out the best way to bulk ingest 100+ books into the same Obsidian/OpenClaw system for free. The main thing I’m trying to solve is how to get a massive amount of transcripts and book text into the system quickly so OpenClaw can use it. Curious how other people have solved this or what kind of workflow they would recommend.

[mighty rampart]

SerpAPI has endpoints for YouTube transcripts but it may not be free at the volume you want to process.

[untold holly]

Hey guys, has anyone encounterd the following error: (0 , _agentHarnessRuntime.hasBeforeToolCallPolicy) is not a function?
I'm running openclaw v2026.5.12

Also, I see some issues with openclaw loading codex, that reflects into uncompatibility while trying to interact with agents. Any message just fails with a generic "check logs" message.

[ancient carbon]

Hello, does anyone know how to interrupt an AI process mid-task on Telegram? Currently, the bot queues up the 'stop' command and only processes it after the generation finishes, which defeats the purpose.

[slow marten]

the only way i found to get around that was by making my own harness lol

[waxen schooner]

Anyone here using the chineses AI llms, im considering it.. since my last bill and whatnot was a bit high.

Which one is actually as good as CLAUDE and GPT

[civic locust]

glm-5.1 is the closest thing to openai and it's very good!

others have reported kimi 2.6 to be good but i've not used that

[waxen schooner]

Do you only use it for coding, or actually have conversational agents too ?

my goal is .. NON robotic. i tried deepseek. its shit

[prisma dock]

Hey all, how you all creating vps? Do you use or suggest using template like terraform to create droplet on digital ocean?

[civic locust]

no, not for coding, it was to run general openclaw stuff

[waxen schooner]

If u dont know.. Digital Ocean is best

itll give u backups etc.

[sturdy tundra]

Is openclaw actually stable again or should i stay on 2026.4.23? Haven't been using it a lot lately and someone told me this was the best stable version a little while ago

[civic locust]

i think it is but [if you're on a vps], take a snapshot and restore if any issues

[sturdy tundra]

Alr ty

[normal garden]

For some reason I can't post to Help so I'll try here

[crystal garden]

hey, does anyone know how i can allow my agent to understand how to use the whatsapp channels and initiate conversations instead of having to wait for someone else to ping me?

[normal garden]

actually my post is there, but not sure why it does not show up when I click on channel

[thick marlin]

You’ll need to set up WhatsApp Business API templates and outbound messaging logic, otherwise the agent can only reply inside the 24h session window. I’ve helped configure this flow for AI agents before, so I can help you get proactive conversations working.
What stack or provider are you using for the WhatsApp integration right now? @crystal garden

[fluid ridge]

Can someone help me understand why I'm now consistently having problems with my agent cutting off messages. This was the most recent message that was cut off but its happening more frequently.

If you want Pest, I think this is a good moment to switch before adding coverage. I’d add pestphp/pest and```

[untold holly]

DeepSeek has some very good models with tool calling capabilities... most of them are actually trained over traces of Claude, thus behave similarly...

[mint tulip]

DeepSeek is 🔥🔥

[wary flint]

I'd like a .py file which can set all /elevated status, across all sessions, to "off". I often find myself elevating a thread and then needing to start other threads and while they aren't a threat in their elevated state, peace of mind says a nuclear "everyone chill out" reset would be worthwhile.

[midnight rock]

this Grok announcement about "use your Grok or X Premium subscription in openclaw" Is it basically saying you don't have to pay for tokens anymore? I feel like there's a catch. Has anyone switched over to this and is it geniunely replacing a pay per token for a flat monthly subscription? (Grok has been more expensive imo, so wondering what I'm missing if anyone knows) Thanks in advance for your help

[midnight rock]

I did an update on 2026.5.4 and it's been working really well with chrome browser, which is one of main reasons for using it! I didn't update before that from mid April bc it wasn't reliable

[copper tiger]

It looks like it’s not per token but it’s a subscription although like others, they keep their limits hidden so you don’t actually know how much usage you have left. Worth a try if you already have x premium.

[junior robin]

Stay on 4.23 maybe try out 5.6-5.7

All other releases have been somewhat problematic for me

[fluid ridge]

Anyone know why something like this happens?

[fathom mortar]

Does anyone know why fresh Openclaw install via npm shows this error when using openrouter?

model=moonshotai/kimi-k2.6 provider=openrouter error=Context overflow: prompt too large for the model. Try /reset (or /new) to start a fresh session, or use a larger-context model. rawError=400 This endpoint's maximum context length is 262144 tokens. However, you requested about 278042 tokens (7941 of text input, 7959 of tool input, 262142 in the output). Please reduce the length of eith…

[frank gust]

I had a massive issue with 5.5. Currently stabilized on 5.7. Watching 5.19 to see if it’s stable enough to upgrade to.

[junior robin]

i think im done updating until I see an actual good stable release

too much time spent on debugging.

wow 5.19 has an 80 on isitstable haven't seen numbers like that since 4.23 the score is still depending though

[maiden shadow]

I had this? There's a reply guard that shortens long messages for no reason

[fluid ridge]

@maiden shadow Is there a wy around it?

[maiden shadow]

Yes, tell the agent to turn it off

Also tell the agent to debug it in general.

Try tui too and check if it happens there

[fluid ridge]

Thank you.

[crystal garden]

I'm just using the built in WhatsApp channel, I'm really fine with a 24h session window but my agent can't even initiate messages inside that window

[thick marlin]

Got you, if it can’t even send messages within the active 24h window, then it’s usually a channel permission, event trigger, or session-state handling issue in the built-in WhatsApp setup. The agent likely isn’t properly configured to initiate outbound actions after the first inbound message. If you’re open to it, I can help you troubleshoot and get the flow working properly. @crystal garden

[crystal garden]

It can only reply immediately after someone sends a message, but can't initiate a message himself. He does this unconsciously in the channel.

[jagged light]

My openclaw is not responding to my messages. I use Telegram what do i do?

[mighty rampart]

Look at the server logs. Where is it running, do you know how to get them?

[spiral parcel]

same problem here " in progress" only

[maiden sundial]

https://github.com/tashfeenahmed/freellmapi

how can i get this?

[mighty rampart]

Scroll down to the Quick Start section and follow the directions. Where are you going to run it?

[maiden sundial]

i mean i only have my computer so i guess that

i send a command to my open claw

https://github.com/tashfeenahmed/freellmapi

Do not run install commands yet. Clone/read only if needed, summarize what it does, what credentials it needs, whether it is safe-looking, and give me a step-by-step plan.```

its been thinking for 6 minutes now

[mighty rampart]

What model are you using now?

[maiden sundial]

qwen/qwen3-coder:free

[maiden sundial]

is this a bad model? i use this for the code then for checking it i use https://validator.w3.org/ and opus4.7

[still gale]

hello - just setting up openclaw for the first time, and got it to message me in discord DM but in the server channel it is only responding with emojis (but shows that it is typing for a long time...like 30 seconds). any advice?

[spiral parcel]

[assistant turn failed before producing content] - compacting context . since the last update I cant even make the agents write anymore , even after I removed every session

[gritty sparrow]

anyone struggling with compaction issue? I've read through the #1459642797895319552 channel and follow the solved answer there but still no luck.

To prevent this, increase your compaction buffer by setting agents.defaults.compaction.reserveTokensFloor to 20000 or higher in your config.```

my setup:
- openclaw: latest version (2026.5.19)
- model: openrouter/z-ai/glm-4.5-air:free
- reserveTokensFloor: 30000

this has happened just recently after v18 update and the limit exceeded shown just after 2-3 chat messages, then the agent stop responding until I message it again.

[thick marlin]

Yeah, I see the issue now. Your agent is basically set up to only react to incoming messages, so even though the 24h session is open, it’s not actually configured to send outbound messages on its own unless a new message comes in first. That’s usually a trigger/session handling limitation in the WhatsApp channel setup. If you want, I can help you get it properly configured so the agent can initiate messages naturally within the active session window.
@crystal garden

[wary berry]

i updated from 4.27-5.19 seconds ago, my agent is crazy, saying it's hermes. openclaw is my past experiment. ......

[unborn ether]

Hi, how can I re-run clawsweeper's triage session?
I created an issue with less info but now I found a workaround and possible root cause

[scarlet silo]

i finally upgraded from 4.23 to 5.19
its almost perfect excpet for remote browser
it now has empty CAPS no keep getting FAIL gateway: GatewayClientRequestError: Error: No connected browser-capable nodes.
can seem to be able to fix it. any ideas?

[solid swallow]

Hi, I search for an cli app which I can use to communicate with my claw like i do via discord. Multiple channels and so on ... you known what I mean. Any suggestions? Thanks in advance. And yes i know the Openclaw native chat client

[scarlet silo]

fixed it. only had to pair again.

[spiral parcel]

Tell me if u find a solution..

[gritty sparrow]

not yet, now my agents stopped responding 🫠

[solar cloud]

I updated to 2026.5.12 and started having gateway responsiveness issues. logs show general event-loop liveness warnings. I'm on debian vps server. VPS shows max CPU usage 35% but usually under 5% so only my OC gateway process is looping and consuming its cpu allocation. I have not been able to diagnose or solve the issue yet.

[elfin lotus]

Hello, is there a way to find why my LLM is not answer in openclaw ?

Like is there logs somewhere that can tell me why the model is likely failing ?

[solar cloud]

I updated again to 2026.5.19, seems to have resolved the gateway looping issue

[spiral parcel]

Same... I use Openclaw on a Pi.. and after this problem yesterday... I just turned everything off... Made me so mad 😡.. I said to myself... I will wait for a new update before I start it again.. tired of troubleshoot all the time🤣

[solar cloud]

I have created an OC maintenance agent to help me which is pretty good but its not the best at diagnosing problems itself. Still, using the agent makes it faster to research and try fixes in a lot of cases. Glad the new release is working better for you as well.

[waxen prawn]

this is a God sent to what i am building mehn

[proper dragon]

Hello. I am setting up Claw on my Android phone with "OpenClaw & Codex - Anyclaw app" anyone have any suggestions or know exactly how I can set it up?

I am not using the OpenClaw Gateway because of the following reasons (correct me if I am incorrect)

1. Unable to easily set up own AI model
2. Unable to use it without wifi
3. Free API tokens greately limited

I already set it up on my laptop with very basic features.

[dapper hatch]

Hi all!

I currently have self-hosted OpenClaw on my proxmox server in my homelab.
My goal for this is to manage my email and calendar and some home automation.
However I'm stumbling upon one little problem, api costs vs subscription.
I have a subscription for some LLMs (claude/openai) but they are completely different from the api credits.
I considered claude-cli but that requires re-authentication every so often.

Another thing I looked at is self-hosting an LLM using Ollama but I dont have a good enough computer for this so basically my question is:

What is the best (and cheapest) LLM for these applications or is there a cheap VPS provider that can handle an Ollama instance

[amber nacelle]

Does anyone here have experience in setting up Telegram with an OpenClaw instance hosted on Hostinger? I can’t get the gateway to change from local to lan.

[spiral parcel]

Fixed the issue by hard-capping the models max output tokens. Openclaw/openrouter was sending requests to the model Kimi K2.6 with an output budget close to the full context window..
The fix was to set an explicit maxtokens cap for the model in openclaw ""params": {
"maxTokens": 16384
}"

[turbid bear]

I’m currently running 4.15 and opus stopped working. I can only use sonnet. Is there a reason for that?

[wary osprey]

Could be an isssue with your provider, what provider are you routing with? Openrouter, AIsa, LiteLLM?

[turbid bear]

Anthropic api key. Seems to have a timeout issue. Response from anthropic was taking too long for opus so then it’s fallback triggered

[errant dew]

Does anyone know how can i give openClaw acess to my gmail? (i already created the credentials.json file in GCP and gave it the GOG skill)

[civic locust]

You saved all the required env vars? The secret and such

If yes, you can either tell your claw that you've saved them in the right places and paste the contents of the JSON file (or ask to give the terminal commands to paste it in the right place without having to manually upload, assuming you're on a VPS)

Then it should give you a URL to sign into and you should paste back the localhost URL that you'll eventually be redirected to

[errant dew]

No i did not save the env vars. In which file should i save them? (I am running locally).

[civic locust]

I can't remember the names tbh 😅 but there's 4 I think

The default email, a secret (that you should make up/generate) and another for the JSON file as well a client id (you'll see this before downloading the JSON)

Ask your claw to walk you though set up

And if you didn't save the client id/secret before downloading the file you gotta do that again

Try asking your claw to read the gog repo and walk you through it (that's how I did it)

But don't add xog_home something directory env var that conflicts with the installed browser and turns out you don't need that

If you still have trouble post here and I'll be happy to help out tomorrow 😃

[errant dew]

Okay, Thank You.

I am just having some trouble with my claw, because i am trying to use a local model with ollama (qwen3.5:9b), so i think the model is not very good at explaining the setups.

[civic locust]

No worries, I went back to my claw chat, had a look and generated a new step by step 😄 also oops sorry the password thing is not needed when you have the file, that was a mistake my claw made earlier when I was double checking now

───

gog Google Workspace setup

This sets up gog for Gmail, Drive, Sheets, Docs, Calendar, and Contacts on a local machine.

1. Install gog

If you have Homebrew:

brew install steipete/tap/gogcli

Check it installed:

gog --version

2. Create Google OAuth credentials

Go to Google Cloud Console:
https://console.cloud.google.com/

Create or select a project.

Enable these APIs:

• Gmail API
• Google Drive API
• Google Sheets API
• Google Docs API
• Google Calendar API
• People API

Then go to OAuth consent screen:

• User type: External is fine for personal use
• Publishing status can stay Testing
• Add your own Gmail under Test users

Then create credentials:

• Credentials → Create Credentials → OAuth client ID
• Application type: Desktop app
• Download the JSON file

3. Set env vars

Use file token storage so you do not need a keyring password/keychain unlock:

export GOG_KEYRING_BACKEND=file

Optional, but fine on a normal local machine:

export XDG_CONFIG_HOME="$HOME/.config"

^note from me: only do this if you're sure this is NOT the path for openclaw's built in browser otherwise ignore

After auth, this is optional convenience so you do not repeat --account:

export GOG_ACCOUNT="[e-mail]"

To make these permanent, add them to ~/.zshrc or ~/.bashrc.

4. Save the OAuth client JSON into gog

Replace the path with wherever your downloaded JSON is:

gog auth credentials ~/Downloads/client_secret.json

5. Authenticate your Google account

gog auth add [e-mail] --services gmail,drive,sheets,docs,calendar,contacts

A browser opens. Sign into Google, approve the permissions, then come back to the terminal.

6. Verify it worked

gog auth list

Test Gmail:

gog gmail search 'newer_than:7d' --max 5

Test Drive:

gog drive search "name contains 'test'" --max 5

Test Sheets if you have a sheet ID:

gog sheets metadata SHEET_ID --json

Important security notes

You do not need your Google password in the terminal.

You do not need to save a keyring password if using:

export GOG_KEYRING_BACKEND=file

Keep these private:

• The downloaded OAuth client JSON
• gog’s generated token/config files under your config directory

If you accidentally paste/share the OAuth client JSON or client secret, delete/rotate that OAuth client in Google Cloud Console and create a new one.

───

[errant dew]

Thank You Very Much!

I will be testing this soon and comeback with feedback if it worked 😁

[chilly nymph]

Google published something recently on GitHub that I have had success using and installing. https://github.com/googleworkspace/cli
It even has OpenClaw installation as part of the Readme.md. @civic locust @errant dew

[gritty sparrow]

Thanks! Will try this.

Where should I put the maxtokens parameter? Is it in openclaw.json file?

[spiral parcel]

~/.openclaw/openclaw.json yes.
{
"agents": {
"defaults": {
"models": {
"openrouter/moonshotai/kimi-k2.6": {
"params": {
"maxTokens": 16384

[crystal garden]

Yeah sure, I was confused when this happened. Because when I actually went into the session and talked to it, it just sent a message to the web app, not the WhatsApp channel.

[thick marlin]

Yeah that makes sense now, the session is likely tied to the web app instead of the WhatsApp thread itself, so the agent responds in the wrong context. Happy to take a closer look with you privately and help get it working properly. @crystal garden

[static tendon]

Y’all…is there an official method for using ChatGPT pro oauth with openclaw? I followed the docs and it was logged in fine but after update to 5.19 I get back from my agent to configure gateway auto for openai-codex…thought that was the runtime, not the auth flow and I’m not sure what the overall config/model selection is supposed to look like…

I swear the auth flow has changed with every release…

[static tendon]

Nvm, I reonboarded and it’s working again…

[sinful quarry]

Title: WebChat session locks after one prompt — forced to start new session every message

Hey team, running into a blocker on macOS WebChat.

Behavior: I send one prompt, get a reply, and then I can't send a follow-up in the same session. Input is either disabled, ignored, or the session appears frozen. The only workaround is starting a brand new session for every single message, which kills any back-and-forth.

Setup:

• OpenClaw on macOS (Darwin 25.5.0, arm64)
• Node v24.14.1
• Channel: WebChat
• Model: anthropic/claude-opus-4-7
• Agent: main

What I've noticed:

• Happens consistently, not intermittent
• Each new session works fine for exactly one turn, then locks
• No visible error message in the chat UI

Does anyone know how to fix this? I cant use it at all.

[civic locust]

Thanks 😄 had a look a few weeks back and tbh peiters is a lot better and more importantly easier to understand

[civic locust]

Bruh tell me about it it was absolute poppycock to switch to codex OAuth after setting up first with API key

[static tendon]

I miss just logging in with claude setup-token 🙁

the good ole days of openopus…

[inland venture]

Hey!

Im very new to openClaw and have a setup with VPS, discord and openAI. They agents is up and running but we are having problems executing my orders. For example, we have spent 2 days trying to add MCP-servers for some workflows but it just dont want to connect to them.

Anyone having similiar problems, or have a good fix for this?
Troubleshooting the issue gets us to some session problems.

I used the video on the official site as a walkthrough for installation

[mighty scaffold]

Hi was wondering if anyone had a nix-container setup for nix-openclaw, the GitHub repo doesn't seem to have any examples that would work in a nix-container.

Any help/recommendations would be greatly appreciated

[vernal wave]

Not directly this, but in case it's related. I created some skills for multi-agent communication that worked just fine on v2026.5.7, however, since upgrading to *.12 (and still with *.19), the skills have completely broken. I've had a couple agents and Krill here in the help channel try to fix to no avail. Might try your luck installing *.7

[slow marten]

if you describe the sesssion problem clearly we might have a better idea of what's going on

[strange moon]

Wassup users! I’m a user who’s designing a Mission Control and I’d love for someone experienced to go over it with me to see how I may be able to improve my build before I implement and seed. Any and all help is appreciated. Dm me!

[gaunt void]

So, how does one actually set up the Claude-cli/anthropic models to use a Max plan? I can’t seem to figure it out properly, nor can my agent. There have been so many changes. I want to run my Claude models through my Max plan. I’ve got the runtime set to “claude-cli” . But I’m losing messages, context (using lossless-claw), and getting messages like “Claude-cli is not a registered harness.” I’m so very confused, and could really use some help.

I also keep going over the context window… I can’t figure out how to get openclaw to auto compact the context with lossless-claw while running an oauth anthropic session.

Perhaps I have the mental model (in my meat brain) wrong. I’d like to see the streaming reasoning of what the model is doing using using Claude-cli runtime and, say, Opus. But even turning on tool call visibility seems to be produce nothing. I’m interacting through the web chat UI, btw.

So, if this is running a Claude code session (headless) for the claude-cli runtime, is there a way to see the Claude code session it’s running?

[lyric bay]

Hey there - anybody knows how to reset - anthropic:default: disabled:billing (3h) — Top up credits (provider billing) or switch │
│ provider.

i already topped up credits, but have to wait for billing 3h to reset

[crimson cave]

anyone getting issue with agents being unable to cross communicate well since since 5.19 update?

[jagged light]

Why does my openclaw take like 15 minutes to respond? I use telegram should i move to discord?

[tiny wing]

Seriously why does stop no longer work this makes it unusable

[hallow prawn]

Hey all — running into a reproducible race condition and curious whether others have hit it or found a workaround.

Setup: Single agent, Discord channel session, two users chatting normally. Model is anthropic/claude-sonnet-4-6.

What happens: When two messages arrive within the same API call window (~10-30s), I consistently see both lane=main and lane=session:agent:main:discord:channel:<id> fire simultaneously on the same .jsonl session file. Both lanes surface EmbeddedAttemptSessionTakeoverError: session file changed while embedded prompt lock was released.

Log pattern:

lane task error: lane=main durationMs=8032 error="EmbeddedAttemptSessionTakeoverError: session file changed while embedded prompt lock was released: .../<sessionId>.jsonl"
lane task error: lane=session:agent:main:discord:channel:<id> durationMs=8037 error="EmbeddedAttemptSessionTakeoverError: ..."

What I've tried: Lowering contextTokens (100k) to reduce API latency, aggressive contextPruning, session idle reset after 60 min. All help with frequency but don't eliminate it — the race fires even on brand-new sessions with minimal context.

Core question: Is the dual-lane dispatch (lane=main + lane=session:agent:main:discord:channel:... firing simultaneously on the same session file) expected behavior for Discord group channels? It seems like any two messages arriving within a single API call window will reproduce this deterministically. Is there a config to serialize processing, or is this a known issue being tracked?

[mighty rampart]

I pasted that into Claude Code and asked it to use the gh command line to search the issues in the OpenClaw GitHub project. tl;dr: https://github.com/openclaw/openclaw/issues/84059 (My OpenClaw instance uses a local model and is not capable of that kind of thing. If yours uses a frontier model just have your OpenClaw agent do it. 🙂 )

[hallow prawn]

I just upgraded to 5.20 so will see if that resolves it

[mighty rampart]

It will not. It was merged after that. To pick up the fix today you'd need to install one of the betas from npm (@beta tag) or build from main. The first stable that will carry it is 2026.5.21 once it's promoted from beta.

[exotic rose]

can claw not login to sites on its own and do tasks for me?

[blazing pumice]

Question: there are a few books i would like my Claw to read. I found where i can buy them, but they may be DRM protected. Does anyone know of a safe, legal way to do this?

[exotic rose]

z-lib

[slate estuary]

Guys check out the plugin I just posted in #1479543671605952532 . Any beta users appreciated!

[jagged light]

What are some signs that my openclaw has malware?

[round pewter]

Hi, I'm not very expert, this is my setup:
16gb Ram
8gb Vram, amd rx6600
Ollama gemma4:26b
Windows

Is openclaw more suited for Linux? I simply ask to make an imageboard and it gets some troubles.
When I use Google's antigravity it's superior, isn't it the same model, how can I achieve that level? Regards

[turbid wigeon]

Well openclaw can run essentially on anything, but i think its easier to run openclaw on linux

[frank vapor]

I have a question. My agents keep using "edit" but it seems to fail a majority of the times :

"status": "error",
"tool": "edit",
"error": "Could not find the exact text in <file>. The old text must match exactly including all whitespace and newlines.

Has anyone else ran into this and have a solution for this?

[maiden sundial]

Hey guys, just a quick question for anyone who might have some experience with all of these models, what is the best for search engine optimization? - https://i.postimg.cc/43jTJBX2/models.png