users-helping-users

[inland knot]

As in i did not encounter any major issues

[rustic pier]

ah ok

[misty trout]

can you guys help me with this? 🙏 #1491374028525535253 message

[inland knot]

Can someone here confirm Codex with GPT Plus is good for daily coding?

Cursor Windsurf and Claude Code got expensive and less quota

[thick ether]

I tryed updating, failed due to @buape/carbon and now everything is broken 🙃

https://discord.com/channels/1456350064065904867/1491359470373179484 any fix ?

[obsidian spear]

npm install -g grammy @grammyjs/runner @grammyjs/transformer-throttler @whiskeysockets/baileys @slack/web-api @slack/bolt @larksuiteoapi/node-sdk @buape/carbon discord-api-types @discordjs/voice @buape/carbon --no-audit --no-fund

This worked for me.

[obtuse iron]

I figured out my "NO" problem, it was something relating to the freshly installed discord-voice plugin. Disabling it and restarting gateway worked. I'm guessing the "NO" is some builtin OpenClaw gateway convenience message but it would be lovely if someone could confirm, I'm quite curious

[abstract dagger]

What model are you using? Model choice is everything

I got nothing but hot garbage out of MinMax m2.7. Claude Sonnet was inconsistent but approaching usable. GPT 5.4 is almost reliable and damn good at writing even complex code.

[true vigil]

good day

does anyone know how can I use ollama as a provider?

[terse lotus]

maybe just use a bit of effort, i would recommend using qwen3.5 or gemma4 if you can get it working with long context prompts in ollama.
https://ollama.com/

[true vigil]

I need it for creative writting which one do you think would fit best? also thank you 🌹

[terse lotus]

if you can get gemma4 running.... for me when using the 31b it just stalls, if i give it a long prompt... my test prompt has 11000 tokens.
but its been a few days.... hopefully its fixed.... also depends on your gpu vram , qwen3.5 9b will fit on just about any gpu, else the smaller version of gemma4 might also be good... those are really your best bet, most models are older than openclaw and thus has problems handling it.

[terse lotus]

on another note, are you sure you even need a harness for creative writing?
sounds like something the default web interface might be fine for.

[true vigil]

I don't know really I tried like chats but the result is so bad that it made me look for other alts basically I'm writting a novel and I need help with screenplay format but I tried asking chat deepseek co pilot (lol) and other the result is so horrible I felt like I'm using ai back in 2021

[terse lotus]

yeah, having a personal agent is a game changer, when it comes to memory and trained behavior.

but just as much as they can be helpful, they can also be an obstinant blackhole of time, when they don't want to corporate.

[true vigil]

lmao you basically summed up everything in one phrase 😂

[terse lotus]

i've had a lot of praxis with openclaw, agent0, browser-use .... and such lol

[terse lotus]

when you got ollama working, you can get the agent to setup llama.cpp
ollama can be kinda slow because it doesn't seem to use caching, you connect it in pretty much in the exact same way, just need to set the provider and model to use openai-completions instead of openai-responses.

the caching in llama.cpp will make the experience, when chatting with the agent a lot better, due to greatly reduced latency....
ollama run locally can take like a few minutes pr turn in many cases, because it will recompute the entire prompt each time.

[heady juniper]

Can Openclaw operate in planning mode? Is there a plugin or skill that can maintain to do list cross tool calls? My Openclaw doesn’t follow skills well. It often forgets the steps during execution and after long tool call output

[true vigil]

this is super helpful I'll give it a try

[midnight adder]

• gateway.bind: Invalid input (allowed: "auto", "lan", "loopback", "custom", "tailnet")
  Anyone seen this?

Halp plz haha

[abstract dagger]

The latest version of ollama has a pre-fab "Launch Openclaw" option you get on startup.

[abstract dagger]

What is the value of the bind param under gateway in openclaw.json in your .openclaw directory?

[terse lotus]

its very much a garbage in, garbage out kinda thing.... the problem is often that words matter to the agent...
if you have somewhere in lets say your 50k tokens worth of core agent .md files , where a line says .... don't write anything

then none of the other 50k tokens might matter, it might always obey that simple statement an always refuse to write anything, until that line is removed.

thus it becomes very important, that things like skills and core agent .md files are worded 100% correctly to get the behavior you want.

on top of that.... you need long context, the kv cache is the agent memory, without a lot of that, it won't be able to remember much, or function correctly.

llm's also get progressively worse, past a certain point of context length, usually around 50-80% of their max context length, their perplexity goes up by a fair bit.
on top of that, the model size or its generation matters a lot.... models are like images of a mind..... if the image is poor quality, the model lacks the granularity in the image.

this means like there are gradiants or pixelation in a bad image, like wise the model will understand less concepts and be able to keep track of less concepts at one time.
just like you are able to tell more and more things in an image the higher quality it is....
meanwhile when you go lower and lower quality, you can see less and less until its just a blur

[abstract dagger]

I've noticed this too. Adding a DB for vector / embeddings helped, as did upgrading which model I used. Another thing that helps is to assume it will forget a step ALWAYS and codify the process you want as a Markdown doc that you then tell it to commit to memory and use for that process going forward.

[terse lotus]

and do remember to force a single source of knowledge location.... else it will just grab something and think its current 😄

[heady juniper]

Yes I am trying to add a scratchpad Md and tell it to write steps to the scratchpad and check/update the completion progress. Will see if that helps. How to use vector DB help w planning?

[heady juniper]

How to force a single source of truth beyond the skill md?

[terse lotus]

vector db is semantic, while without it is searching based on words.

you choose a location where documentation is stored.... basically like your own documents folder on your own system, if you just put files everywhere you can't find it when you need it.
agent memory isn't always perfect, which is why its good to have a location, both you can the agent knows, is the singular place that is updated first.

my folders in documents always contain a README.md and a SECONDARY_LOCATION.md on top of whatever else it puts there.
readme is the core thing... its the first thing to check, which describes that the folder is about and such, how to use it... its in all my agent workspace folders.
the SECONDARY_LOCATIONS.md is for documents, if it has multiple locations where the information might be stored, as its not always possible to only have the information in the docs folder

its simply a list, so it knows where to update when there has been made changes in the primary knowledge source aka docs
@heady juniper

[waxen rain]

Which subscription plans work with OpenClaw without API keys (like Claude Max used to) ?

[abstract dagger]

No, you're missing the boat 🙂 Don't try to get IT to write out its steps. YOU must specify precisely what you want it to do for a given process, in excruciating detail, and make it save that to check against whenever it runs the process.

[abstract dagger]

ChatGPT Plus at $20/mo

[waxen rain]

Does it use the ChatGPT subscription via OAuth like Claude Max, or does it still require an OpenAI API key?

[shadow pawn]

During setup of new OC install of 2026.4.8 getting:
Cannot find module '@larksuiteoapi/node-sdk'
Cannot find module '@buape/carbon'

Any fixes?

[abstract dagger]

ChatGPT login/password so I think Oauth. Not an API key. Not metered. $20/mo flat.

[inland knot]

How is GPT 5.4 feeling in OpenClaw, i tried GPT 5 Mini but it was too verbose and didnt respect any .md file, and was hot garbage in tool calling, i am sure GPT 5.4 is better, I'm currently running on MiniMax 2.7, no problems to make me unconfortable

[midnight adder]

Not seeing that

[midnight adder]

It's trash

[abstract dagger]

Then something is DREADFULLY wrong with your Openclaw JSON config. What does 'openclaw doctor' say?

[inland knot]

Solid feedback

[mellow hornet]

Hey everyone! Quick question about the Google provider setup.

I am currently setting up OpenClaw (running via a local npm install inside a VS Code Dev
Container with a mounted Docker volume for state persistence).

During the npx openclaw onboard process, I see the option Google (Gemini API Key + OAuth). I am a
Google One AI Premium subscriber—is there any way to just use my consumer subscription via OAuth
for inference, or is the Developer API Key strictly required to power the model? Thanks!

[gray juniper]

from: TonyD in: users-helping-users larksuiteoapi

larksuiteoapi

[winged lagoon]

U shoud migrate model to OpenAI subs and go via codex outh

[kind shadow]

Hello guys, i never used openclaw and my windows 10 VM is almost ready, what are the steps to install Openclaws?

[simple spruce]

hey guys, im pretty new to these things, i wnated to try out openclaw and i decided to use gemini since i have a google one subscription, but i keep getting the same error message
Error: Cannot find module '@larksuiteoapi/node-sdk'
Require stack:

• C:\Users\a\AppData\Roaming\npm\node_modules\openclaw\dist\probe-Cz2PiFtC.js

any clue on how to fix this?

[barren python]

Slack openclaw issue

[inland knot]

Are Gemma 4 models any good for using in OpenClaw?

[midnight adder]

It thinks about doing tasks but never does them, that is my feedback
It's trash lol

[midnight adder]

Same issue
Runs the full thing then has the problem

I may reset it

Either way I'll troubleshoot more when I'm home

[inland knot]

So let me understand, the only truly capable models for OpenClaw are Claude and MiniMax?

As i said, my M2.7 does perform well, not on Claude level but hey, I got no cash for Claude

[midnight adder]

In my opinion
I have Claude using local llms via ssh and claw nodes to code and reason
But Claude is the commander

One machine controlling two others with local llms on them

[inland knot]

How much do you burn monthly on Claude?

[midnight adder]

Last month was 1500

So far I'm at 470for this month
I've reconfigured my arch a Lil bit as well to see if it will be cheaper
I don't mind paying the api

[flat tinsel]

(IM SOMEWHAT NEW TO STUFF BUT CLEARLY I UNDERSTAND) alright im in one hell of a pickle because my setup is a little interesting, im running a 6600xt which ollama doesnt support, so i did the amd for ollama setup with sdk/rocm driver stuff, i then had a WSL2 of ubuntu running my openclaw and these linked to eachother local host running qwen3.5:4b on my 6600xt through ollama serve, it was good enough it worked and this all took me like 3 days to set up BUT right now i keep getting hit with the issue, even swapping my model qwen3.5:2b, after using different versions of this command for "different amounts?" (openclaw config set agents.defaults.compaction.reserveTokensFloor 100000) (i would use like 8000 or 16000 etc) NOW ITS AT 2048. i have a custom context amount set for my qwen3.5:4b its set to 16384 context + 1024 of "max tokens"? and this ran fine but now i have THE ISSUE:

⚠️ Context limit exceeded. I've reset our conversation to start fresh - please try again.

To prevent this, increase your compaction buffer by setting agents.defaults.compaction.reserveTokensFloor to 20000 or higher in your config.

WHAT DO I DO ABOUT THIS AS IVE ALREADY CONFIGURED IT

[placid flower]

apparently I thought I was updating to 4.7 but to my surprise it's 4.8 and now config is bad. I ran doctor because that usually cleans things up after update, but that didn't work. Anyone else getting Error: Cannot find module '@buape/carbon'
Require stack: openclaw\dist\ui-ztuR6rec.js
?? Did 4.8 come out too fast?'

[inland knot]

Damn

[abstract dagger]

Strongly disagree. Claude Sonnet and Opus are just fine. GPT 5.4 is pretty darn good. I found Minmax m2.7 to be Absolute Hot Garbage for my use cases.

[inland knot]

I might try GPT 5.4, I'm just afraid of the quota, but for my use case, marketing automations and workflows, will it be enough ?

[inland knot]

M2.7 got like 1500 req/5hr, basically infinite AI agent, but if GPT 5.4 quota is enough, imma switch to it, you got any problems with the quota?

And is it not frustrating to talk to it? Does it respect SOUL.md and AGENTS.md to give him some personality?

[abstract dagger]

ChatGPT Plus quota for Codex / GPT 5.4 is very generous. I've been having it perform my regularly scheduled agentic tasks PLUS implementing a very large side project in phases on top of it all and I haven't come close to using all my quota yet, either the 5 hr max or weekly max.

[inland knot]

Fair enough, if you can confirm the personality test imma switch my model rn, i dont wanna talk to a fridge, i want it to strictly respect SOUL.md and AGENTS.md, but i guess he can do that

[graceful venture]

Yes I've the same issue

[abstract dagger]

What "personality test" do you want? I can try whatever you like.

[placid flower]

trying to find a remedy. I already tried a npm install latest, no luck . still getting Error: Cannot find module '@buape/carbon'

[graceful venture]

I also had
Error: Cannot find module 'grammy'
code: 'MODULE_NOT_FOUND'

[inland knot]

I currently got a SOUL.md crafted by Krill himself to give my agents the personality remotely close to Opus, I mean, it is not Opus, but it's still a comprehensive SOUL.md that leads there

[graceful venture]

Is there a way of uninstalling openclaw and reinstalling without using all data from workspace

[slim path]

Yes, the workspace folder is different, but I don't see reinstalling helping. I could be wrong.

[inland knot]

1. Ask it: "Who are you and what's your vibe?"
2. Check if it references your SOUL.md/AGENTS.md files or just gives a generic answer
3. Ask something personal about your preferences/workflows and see if it picks up context
   Tell me if it outputs chain-of-thought visible or stays conversational

I wanna see if GPT 5.4 respects workspace personality files before I switch from MiniMax M2.7. Thanks!

This would be the last favour i ask you before switching models

[fresh ice]

Question: I'm still running v3.28. Fearful of upgrading to 4.8 due to the exec permission issues. Have those been resolved?

[inland knot]

Nope, but you can just ask your own agent to turn the exec permission to off himself

[graceful venture]

Do does everyone with 4.8 have this issue

[inland knot]

i did this with mine

[placid flower]

grammy disappeared after a npm install latest, but the other remains.

[slim path]

This is the query I told openclaw to fix itself for dependency issues and it did. I'm running the latest version without any issues.

Run openclaw doctor --fix and fix missing node modules

[fresh ice]

Then is it really not an issue?

[fresh ice]

I had problems enabling exec privileges in 4.3

[slim path]

I guess so. Checkout #users-helping-users message

[inland knot]

Not really, you can ask you agent to switch exec securities to off

At least mine did

[graceful venture]

Do you type that in exactly as written into PS ? It didn't recognise it

[abstract dagger]

backup your .openclaw folder. Take note of what node version you're running and what packages you have installed.

[inland knot]

@abstract dagger can you do me a favour of giving your agent this prompt?

Ask it: "Who are you and what's your vibe?"
Check if it references your SOUL.md/AGENTS.md files or just gives a generic answer
Ask something personal about your preferences/workflows and see if it picks up context
Tell me if it outputs chain-of-thought visible or stays conversational

I wanna see if GPT 5.4 respects workspace personality files before I switch from MiniMax M2.7. Thanks!

And also ask him to compare himself to M2.7 in agentic

Thx a lot

[heavy sapphire]

what do i wrong?

o  Default model
|  Keep current (google/gemini-3.1-pro-preview)
Error: Cannot find module '@larksuiteoapi/node-sdk'
Require stack:
- C:\Users\****\AppData\Roaming\npm\node_modules\openclaw\dist\probe-Cz2PiFtC.js
PS C:\WINDOWS\system32>
PS C:\WINDOWS\system32>

[green sundial]

I feel like anytime OpenClaw does an update everything stops working and I have to spend hours trying to figure out why

[inland knot]

Fix: "Cannot find module '@larksuiteoapi/node-sdk'"

This is an OpenClaw dependency bug. Try these fixes in order:

Fix 1 — Quick install the missing module
npm install -g @larksuiteoapi/node-sdk

Then restart OpenClaw.

Fix 2 — Reinstall OpenClaw completely
npm uninstall -g openclaw
npm install -g openclaw

Fix 3 — If you're on Windows and the global install path is weird
npm config get prefix

Make sure it returns something like C:\Users\YourName\AppData\Roaming\npm

Then run:

npm install -g @larksuiteoapi/node-sdk --force

Fix 4 — If nothing works
Check your Node version — OpenClaw needs Node 18+:
node -v

If below 18, update Node first.

---

If the error persists after all this, it's likely a recent OpenClaw update introduced a broken dependency. In that case, report it on the OpenClaw Discord with your Node version and OS.

[odd trellis]

hello everyone, how does one set up Heyron.ai?

[abstract dagger]

@Feoh can you do me a favour of giving

[lone zephyr]

getting Error: Cannot find module '@buape/carbon' when trying to update, this broke the install, github issues says upgrading to latest (2026.4.8) does not work, trying to downgrade to 2026.4.2

[green sundial]

I kept using npm install -g and whatever is missing until it started working again

apparently there were a few modules missing so I had to keep doing it

[lone zephyr]

is the crash when running openclaw onboard a separate thing, or?

2026.4.2 seems to install no problem but some CLI commands crash

[graceful venture]

How do you downgrade?

[devout relic]

new gpt5.4 user (coming over from opus max since day 1)
a bit of a learning curve.
im now worried about my weekly limit as i'm running /think high.
is the burn rate quite a bit higher than /think medium?

[lone zephyr]

you delete the openclaw folder inside node_modules folder (wherever that may be installed, for regular installation its in %APPDATA%) then just run installation again like npm install -g openclaw@<version>

afaik these are just system files, no user data

[graceful venture]

So no major disruption

[lone zephyr]

yeah, most config and user data is probably in .openclaw folder

[graceful venture]

Cheers mate, will try when I'm back home

[hollow ermine]

hi fellow crustys. Having a bit of an issue after migrating OC to a new machine on my tailscale network, cant access it from my desktop which is also on the same network/tailscale.

used to be able to access the old machine fine, kinda stuck

[topaz seal]

hi, is anyone using the claude cli provider (via an oauth key) is everything working for you?

[abstract dagger]

Use the netcat (nc) command to see if there is network connectivity between desktop and server:

nc -v server_tailscale_ip port

[hollow ermine]

desktop machine is running windows, do you know the powershell equiv?

i can ping the new servers tailscale IP successfully

[abstract dagger]

This might help. Not tested: https://learn.microsoft.com/en-us/powershell/module/nettcpip/test-netconnection?view=windowsserver2025-ps

[hollow ermine]

ok thanks, actually before I do that having some more serious issues.

on the new rig I followed the official migration instructions, seems fine, but once I try and connect to my LLM (local to the new machine) i get this error in chat Agent failed before reply: EACCES: permission denied, mkdir '/home/newHomeDir'

why is it trying to make that dir?

[obtuse iron]

Can you access it by IP SSH but not through the tailnet address; or not access it at all?

[obtuse iron]

Dunno why it's trying to do that, but I recommend maybe killing that session, seems like a deviation

[hollow ermine]

I can ssh to the local IP, not the tailscale but I do think I have SSH activated for that TS connection.

[hollow ermine]

new session fixed that thanks

[latent hedge]

** HELP **:

Is there anyone here who succesfully have set up a seperate machine hosting local LLM's with Ollama, and connecting to that from Openclaw hosted on its own machine.

** ISSUE **:

When I do this, with for example qwen3.5:35b, it fails from openclaw, because I can't figure out how to disable thinking mode. I can do that through it's seperate curl command, and it works fine. But I can't figure out how to do it through openclaw.

Can someone show me a configuration from an openclaw.json file that disables thinking?

This curl command works fine:
curl -m 120 http://xxx.xxx.xxx.xxx:11434/api/generate \
-H "Content-Type: application/json"
-d '{
"model": "qwen3.5:35b",
"prompt": "Reply only with OK",
"stream": false,
"think": false,
"keep_alive": -1
}'

No way to figure out how to diable this through openclaw though.

[obtuse iron]

For most varieties of Linux the user never needs to manually make a directory inside /home/, that's handled by the OS

[unborn igloo]

is Krill no longer active in help channels?

[flat tinsel]

this is amazing i love to see someone doing what im doing, right now i haved a local llm qwen3.5:4b running on my main machine and then a WSL2 ubuntu running my openclaw sort of the same as you, but the way you do it when u have open claw open and ollama running, go to the openclaw local host, go to sessions inside the control category under the main chat, and you can change its thinking there

[upbeat coral]

I am wondering the same thing. I see several new threads in #1459642797895319552 with no Krill halpz.

[mystic pine]

Is Krill the server's main helper?

[crisp oyster]

Hello! Has anyone used an Agentic ai management solution like https://crewai.com/ before? If so, what is your feedback?

[robust marten]

If you're in the Control OpenClaw dashboard, you can run the /think slash command to disable or adjust thinking modes

[upbeat coral]

It's the helper bot.

[mystic pine]

ohhh alr

[upbeat coral]

Seems to be having a krill-nap. 😴

[placid flower]

following~

[strange jacinth]

Hey guys, just configured my Openclaw and maded a github repo. It autimatically mande the /home/claw/.openclaw/workspace its default workspace, but that doesnt capture configs like channels and stuff like that
Is it better to leave it clean like that or should i change it? I kinda new so i dont know if i could be missing important things by not seeing all of it

[crude owl]

Depend on how many agents you are setting up and if you want to make that you main workspace

[fresh ice]

Anyone used dreaming yet?

[latent hedge]

Can you copy/paste what it looks like in your openclaw.json file?

[latent hedge]

Can you copy/paste and show me what it looks like in your openclaw.json file

[strange jacinth]

Well, it’s a VPS I created specifically for that purpose. I dont think ill need any other wnvironment so soon

[upbeat coral]

Disable Thinking Mode When using Ollama

[pallid cove]

did you guys get spurious fallbacks since updating to .4.5 or more recent? I keep seeing them, started 2-3 days ago?

[old gull]

yes ever since i've had to switch to chatgpt from anthropic after they banned subscriptions from working with openclaw, my bot's intelligence has been reduced to a caveman in essence. Entirely unable to execute commands, use tools, just one problem after another. I'll fix one thing and then a new problem comes up

[old gull]

at the point now where i'm going to cave and just use Opus through CLI until i get things under control. My plan is to use a planning/orchestrator agent for the reasoning and checking the work of subagents which will do all the leg work and coding through GPT 5.4 or a local model. i run on a mac mini 24gb ram so nothing crazy

[indigo goblet]

Wheres Krill at?

Still above the water?

[upbeat coral]

Napping.

[pallid cove]

You can use a daily driver model in OC and delagate hard tasks to Claude code through ACP

[indigo goblet]

Hm

Understandable

[slender obsidian]

so for claude -p what exactly does auth look like with that? Is that installing claude code separately to the VM with openclaw and then doing auth with that or does Openclaw have a built in claude cli component?

[indigo goblet]

Damnn.. thought about Krill's companion... gpt or claude?

Those guys are equally good too...

Just a thought.

[cloud burrow]

Hei guys, for some time i can t receive notification messages, cron based notifications via telegram; however i am able to chat via telegram...only getting messages updates from cron jobs seems to be issue. How do I fix it?? Cause my openclaw said it can t fixit by itself..

[midnight adder]

Gateway status failed: Error: Invalid config at C:\Users main.openclaw\openclaw.json:

• gateway.bind: Invalid input (allowed: "auto", "lan", "loopback", "custom", "tailnet")
  This is what I get when I run pretty much anything

[spark valley]

Check ur config

It got invalid input in gateway bind it literally says that

[midnight adder]

Sorry I'm a artard

Got it lol

[spark valley]

No its alright ur not

Go to openclaw.json

Find gateway

Then bind

Then use the allowed ones in between bracjets

[midnight adder]

Okay

[gritty beacon]

can discord bot that i connect to openclaw mangement discord room ? (deleated msg,filter msg) something like full control my discord server

[late roost]

Hi, everyone. I need a little help... OpenClaw has stopped responding; it's not responding, and I'm just getting this error message: "400 {"error":"registry.ollama.ai/library/gemini-3-flash-preview:cloud does not support tools"}" What should I delete, and where should I delete it, and then run a DOCTOR ?

[midnight adder]

Okay I found that along with like 20 different ones with ". Bak, bak 1 2 3" and so on
Then a bunch of ".jason.clobberd

[late roost]

oooh i see something ollama error 🙁

[abstract dagger]

Seems pretty clear. The contents of your openclaw.json file are corrupted or missing.

If you don't have a blob that starts like this in your file:

 "gateway": {
    "port": 18789,
    "mode": "local",
    "bind": "lan",

You've got serious problems and may need to run onboarding again.

[spark valley]

Yea that some clobbed ones u need to find openclaw.json exactly open in notepad or if on linux do nano openclaw.json then find the lines

[midnight adder]

I did find the openclaw.jason exactly

[midnight adder]

"run onboarding again"
Reconfigure or uninstall and reinstall?
Sorry stupid questions
Appreciate the help

[abstract dagger]

No worries not stupid

You can run openclaw onboard for the step by step interactive setup. OR you can try openclaw configure to fill in the missing values.

[midnight adder]

Okay

[gaunt falcon]

so im building out a "command center" and i'm trying to create an intuive session view, where i have buttons that kill sessions, execute compaction, and one for having the agent update long-term and daily memory with context necessary to pick-up in a new sesssion where we left off. For some reason, i cannot get the agent to create functioning buttons. everytime it says it's working when its not, the buttons essentially do nothing. is there a trick for getting this working or a process i should mention that the agent will be able to utilize for making this work?

[abstract dagger]

1. Which model are you using? That's super important.
2. Have you looked at the code for the buttons? Does it make sense?
3. Is an error being generated anywhere or do the button clicks fail silently?

[midnight adder]

I think I got it, ty peeps

[gaunt falcon]

kimi 2.5 and i have, it looks like it should be pointing correctly but i'm pretty new at code and couldnt really dictate what was wrong about it. they fail silently.

[naive ocean]

updated from 2026.4.2 to 2026.4.8 and a whole bunch of npm errors , now nothing working , anyone having the same issue ?

[wet kraken]

Have you guys found a way to limit how much heartbeat pumps up each time? I still want the context of the chat to go via heartbeat, sure. But my heartbeat, on even a completely empty that basically has nothing but a heartbeat, is sending back tens of thousands of tokens full of nothing! Why?! I even got GPT 5.4 High to try to look into it, and it basically just says it's sending up context and thats like just unnecessary but unavoidable context it needs to send for every heartbeat even if there's nothing else. Surely this can't be the only way? It's just such a massive, massive waste of tokens for doing absolutely nothing!

[gaunt falcon]

i may have found a fix in my articulation. example: for the kill button creating the explicit rule that when this button is pressed the corresponding session will be removed from ~/.openclaw/agents/[agent]/sessions/sessions.json and the session will be removed from the GUI with a confirmation message.

[hollow veldt]

Anyone wanting to work with a psychologist to set up super personality, soul, etc for your claw, I'm here to help 💞

By the way, I've just met Peter 😁🤣❣️

[iron dragon]

I have 0 idea on what i should do and i tried everything i could think of (i have never done anything with terminal)

Anyone that can help me?

[OK] Windows detected
[] Existing OpenClaw installation detected
[OK] Node.js v24.14.1 found
[] Installing OpenClaw (openclaw@latest)...
[OK] OpenClaw installed
[*] Running doctor to migrate settings...

OpenClaw 2026.4.8 (9ece252)
I don't judge, but your missing API keys are absolutely judging you.

OPENCLAW

OpenClaw doctor

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
Missing config: run openclaw setup first.

Gateway auth
Gateway auth is off or missing a token.
Token auth is now the recommended default (including loopback).

Error: Cannot find module '@buape/carbon'
Require stack:

• C:\Users\jason\AppData\Roaming\npm\node_modules\openclaw\dist\ui-ztuR6rec.js

[OK] Migration complete

OpenClaw installed successfully (OpenClaw 2026.4.8 (9ece252))!

Molting complete. Please don't look at my soft shell phase.

Upgrade complete. Run openclaw doctor to check for additional migrations.

[wary osprey]

are you only using ChatGPT or you use some other models too?

[gaunt falcon]

try reinstalling your gateway openclaw gateway install --force

then openclaw gateway run

also debugging with codex or cluade is the move

[iron dragon]

i tried to delete it but now i also can't

[gaunt falcon]

just force it to in terminal did you use the command i sent you

[iron dragon]

gateway install --force like this?

[gaunt falcon]

no

openclaw gateway install --force

remember every command you use must start with openclaw

[iron dragon]

openclaw : File C:\Users\jason\AppData\Roaming\npm\openclaw.ps1 cannot be loaded because running scripts is disabled on this system. For more information, see about_Execution_Policies at https:/go.microsoft.c
om/fwlink/?LinkID=135170.
At line:1 char:1

• openclaw gateway install --force

•   + CategoryInfo          : SecurityError: (:) [], ParentContainsErrorRecordException
    + FullyQualifiedErrorId : UnauthorizedAccess
  
  

I get this back

[gaunt falcon]

what os are you on and terminal?

i would paste that output into codex or something

[iron dragon]

idk what that means

thats why i came here i have never done anything with code

[gaunt falcon]

yeah i would maybe rethink openclaw if you're lost already man lol. but i get it. get an external model like gpts codex and use it to debut paste your outputs into it and it will tell you what to do every time

a lot better than asking us

debug*

[iron dragon]

you also said something about claude right?

[gaunt falcon]

yeah claude is great for debug too

[iron dragon]

okay i will try that

[gaunt falcon]

im testing codex atm but i use claude mainly

[iron dragon]

i just downloaded claude so i hope it will help

[gaunt falcon]

it will, whenever you get an output you dont understand just paste it into claude, if it needs more context give it an explanation of your end goal. overtime you will start to learn how things work just by working with claude as sa debug agent

it seems like a simple choice to do this but a lot of people give up or resolve to ask google or reddit or a real human which is not the move. we are using AI to code, it only makes sense that we use it to debug as well..

[icy mica]

Hey, @latent hedge I actually ran into the exact same issue when wiring up Ollama with a separate app layer.

The problem is that OpenClaw doesn’t expose the think parameter directly in its default config schema, so even though Ollama supports "think": false, it never gets passed through unless you explicitly override the request body.

What worked for me was forcing the payload override inside openclaw.json.

You can do something like this:
{
"llm": {
"provider": "ollama",
"baseUrl": "http://xxx.xxx.xxx.xxx:11434",
"model": "qwen3.5:35b",
"defaultParams": {
"stream": false,
"keep_alive": -1,
"think": false
}
}
}

If your version doesn’t support defaultParams, then the fallback is to use a custom provider wrapper or middleware and inject "think": false into every /api/generate call before it hits Ollama.

Another thing I noticed: some OpenClaw builds silently ignore unknown params, so if think still doesn’t apply, you may need to patch the request builder (usually where it maps config → API payload).

Once I forced that param through, Qwen stopped going into thinking mode and behaved exactly like the curl test.

Happy to compare setups if you’re still stuck this one took me a bit to figure out too 👍

[naive ocean]

the errors updating to .8 are painful .... bunch of npm installs and doctor runs ...

[indigo goblet]

bruv, whats up, all OK?

swear to god, upstream update are killing cron jobs and memory

like a cocaine downtime, everytime

[gaunt falcon]

yeah luckily i didnt update

[indigo goblet]

gotta rewind for a second, then all good, but damn, some black magic shit.

[plucky escarp]

Anyone build their Mission Control on a VPS?

[cedar patrol]

I feel like every time I upgrade openclaw I have a 2 hour fix session ahead of me. 😵‍💫

[indigo goblet]

got ptsd from cron jobs and memory, agent is like "I drank a whiskey bottle"

but quick fix tho, all good lol

[mellow briar]

Can vouch for using Claude. Perplexity came in handy too, but in the end Opus was the way. Sonnet/Perp hallucinated and wasted a few hours. idk anything abt Linux / Ubuntu but was able get setup on an older XPS I had collecting dust.

[summer hemlock]

shouldn't be happening. But you should only update when you can reach your claw machine, to be safe

I run it on a mac mini and I have antigravity ready to debug claw with opus 4.6

sometimes the auth options / gateway changes. Then you can just do openclaw doctor first, if that doesn't work redo the onboarding.

[normal garnet]

Does anyone have the missing dependencies problem,

[gaunt falcon]

are there any prebuild command center repositories i can use? i can make edits from there to customize it for myself i just need a solid foundation]

[cedar patrol]

I am running Qwen3.5-35B-MOE-a4 on llamacpp, becuase it runs locally on my ASUS Ascent GX10 , so in the last release the loose completions schema changed. Doctor didnt check that.

I had to ssh in and install claude code to figure it out.

[summer hemlock]

well, installing an IDE like claude code on your claw machine to help you debug is the best tip I can give anyone, really

opus 4.6 is jus such a beast.

[final island]

I've been running OC on Claude Opus in subscription mode like a lot of people, and with the billing change I've been considering seeing how it runs on Gemma 4. My use case for OC has been pretty narrow (mostly pulling info from webpages and writing stuff for me) so I don't need a hugely powerful LLM backing it. My RTX5080 is in a windows machine, and I am wondering what the best way to run it would be. My OC instance is in a Ubuntu VM on an ESX host in my home lab. Should I be looking at Ollama to run it, or are there better alternatives?

[cedar patrol]

I did some research and I think Qwen3.5-35B may be better that Gemma4 for coding, but many people are saying its better for Agentic so I have been conisdering it for OC use cases

[final island]

I do have a subscription still, and just do my coding work with regular old claude code, I've been using openclaw primarily for job hunting

and it seems to have worked, I just really like how well it writes stuff like cover letters and whatnot for me so I don't want to lose that ability even if I am about to get a job offer

I'm just not sure what the most efficient way to run a local model is these days

looking at the scrollback it looks like people have had to mess with llama quite a bit to get openclaw to work with it

[cedar patrol]

I had to buy a ASUS Ascent GX10 for my use cases

many people say there are cheaper options but my MOE models benefit from the GPU

[final island]

ooo nifty

[cedar patrol]

people here dont like me tho.

[final island]

I was focusing on Gemma 4 specifically because it's supposedly using some better compression techniques and works well on consumer level hardware

[sacred surge]

how do you get openclaw to remotely use the browser? everytime i tell openclaw to open the browser from discord or imessage it just says " i cant do that ! " but if i do openclaw browser in terminal it works just fine

[cedar patrol]

I use playwright

[final island]

I've heard playwright mentioned a lot recently, need to figure out what that does

[sacred surge]

its a way to programmticaly use chromium

i thought openclaw was all good at computer usage i didnt know it couldnt do things like open my browser

thats insane

[final island]

yeah it's pretty awesome

I'd been using the browser relay, and then switched to the headless setup with chrome

[sacred surge]

how did you do that?

[final island]

I was chatting with it wired up to Opus when it did that, so it just offered to set it up and I auth'd it lol

[sacred surge]

so its called browser realy?

[final island]

it has a dedicated Ubuntu VM though so I allow it to do stuff I wouldn't normally do on my daily driver system

[sacred surge]

oh i c

[final island]

yeah openclaw browser relay and I believe they changed it recently

it's now something else, I'll have to see if I can find it

you have to be careful with extensions in the chrome store, you want to use the supported stuff in openclaw itself, or a vetted plugin on the hub

[nimble hill]

and changed it for better, it is now part of the plugin. in openclaw.json, you need to havw following entries

  "plugins": {
    "entries": {
      "browser": {
        "enabled": true
      }
    }
  }

  "browser": {
    "noSandbox": true
  }

and if you are controlling a remove node, for example you have gateway running on VPS and you want o control the chrome running on windows (openclaw running as node), you need

    "nodes": {
      "browser": {
        "mode": "auto",
        "node": "<Node device ID>"
      },

you can get node device id via openclaw devices list

[final island]

yea, relay burned a lot of tokens I noticed

it was pretty rad watching it apply for jobs for me though lol

[terse sphinx]

Hey guys

I run 6 openclaw agents in 6 docker containers

I recently created a new "Chief of staff" type agent

But I'm struggling to make the A2A communication work

Each agent have their containers, but also their own gateway, on the same machine (a Mac Mini Pro)

So first they would report most A2A comms attemps as a security incident.

So I tried to implement a Ed25519-signed cross-gateway messaging via fleet-send.sh (using dedicated A2A session keys)

But it still doesn't work 9 / 10 times (agents don't respond to each other, etc.)

Did anyone manage to make cross-gateway messaging reliably or is the only way to get this to really work is to get all agents on the same gateway?

[nimble hill]

I run 6 openclaw agents in 6 docker containers
why 6 - the single gateway can control multiple agents

[terse sphinx]

Yup but I wanted complete separation because it's different projects

[nimble hill]

agents across multigateway can be done, but it is a pain - the architecture itself is not there., Openclaw is in beta.
if you want seperate, multi agent in single gateway also achieve that, just have seperate workspace, model for each agent. THey are essentially the same. Even eithin the same gateway, talking between agents in orchestrator is not smoothe sailing - people currently use files as a common message exchange, A2A frame does not really work as of now, atleast in my experience

[terse sphinx]

file as a common message exchange?

What does that mean?

[nimble hill]

if 2 agents needs to hand off tasks to each other - say orchestrator model.

you are just wasting compute resource for each gateway for each agent and the pain of managing each one

[terse sphinx]

If even on the same gateway, A2A doesn't work, then I won't waste time migrating all my agents and will keep them like that and explore this "file as common message exchange"

would love to see if there's any guide or article on this?

Compute wasted on multiple gateways and containers is really minimal for now but I'll keep an eye on it!

[nimble hill]

I mean you can try, you need to enable: - main and fin are my agents

  "tools": {
    "agentToAgent": {
      "enabled": true,
      "allow": ["main", "fin"]
    }

I have observed hits and misses. Hence, I most of the time connect to my n8n for deterministic flow and control mult agent there

[terse sphinx]

Got it. Thanks 🙏

[nimble hill]

Also keep the gateway simple like a single head - I run my gateway in VPS. Then there are various nodes like windows, linux, mac where I run openclaw as nodes in docker containers which are connected to gateway. Now there are multiple agents, which direct the resources in ewach node as necessary, like in windows I use it for browser control, in lnux is for code compilation. and mac - i don't really use mac - just random stuff to get arm images built for my projects

[formal fog]

another tip: for those of you running openclaw on its own hardmare, maybe one without a screen, use tailscale drive for easier sharing of your files from your openclaw box to one with a real editor so you can avoid editng files in the terminal like its the 1970s.

[copper nova]

buy a dummy hdmi dongle, that's what I did atleast, gives your robot a screen to play with

[pallid yacht]

Anyone having issues with gpt-5.4 since yesterdays update? Mine basically became unusuable

[unborn igloo]

exactly. it's all part of openai's strategy: buy openclaw and make it so terrible with gpt that users have to buy expensive gpt plans to correct basic issues. it's brilliant. don't know why they're incentivized to do anything differently

[nimble hill]

You can also run a docker container with image coder which can display vscode like in your browser.

[nimble hill]

My 5.4 mini works fine

[nimble hill]

Run on linux server, run via ssh from anyplace

[pallid yacht]

I'm on the $200 plan

[nimble hill]

I am on $8 plan 😃

[copper nova]

huh? I was answering a question lol

[nimble hill]

Chill - it will be helpful for people who want it

[fringe bobcat]

Can anyone help me openclaw setups? Thank you.

[dusk arch]

howdy, is the only option to run openclaw w claude being via their pay-as-you-go plan? or is there a workaround?

[slim crown]

how do i connect my mac to a windows machine serving qwen3 27b on ollama?

i can curl the windows machine fine.

[storm basin]

in ollama model setup you can specify a remote address.

[nimble hill]

during onboarding you select the ollama and give your windows machine ip and port where ollama is running

[slim crown]

how do i access that onboarding wizard again? ollama is running fine.

[storm basin]

openclaw config

[slim crown]

and it wont wipe my preexisting config right? i already have other models configs setup

right?

[nimble hill]

yes - hwoever, I always keep seperate copy of the last known good version of openclaw.json - openclaw also keeps a .bak file but I don't trust it

[midnight adder]

Error: Cannot find module '@larksuiteoapi/node-sdk'
Require stack:

• C:\Users main\AppData\Roaming\npm\node_modules\openclaw\dist\probe-Cz2PiFtC.js
  PS C:\windows\system32>

Anyone fixed this?

[drifting epoch]

anything after 4.1. is unusable for me at the moment, on PC, anyone else?

[midnight adder]

So just downgrade it? Cause I'm dealing with the same shit lol

2 days of trouble shooting
Fixing and breaking lol

Most recent updates are not bueno

[drifting epoch]

i hve had no issues w 5.4 on the pro plan

[pallid yacht]

I fixed it somehow

[drifting epoch]

but seriously anyone using 4.8 on windows?

[midnight adder]

Error: Cannot find module '@buape/carbon'
Require stack:

• C:\Users main\AppData\Roaming\npm\node_modules\openclaw\dist\ui-ztuR6rec.js
  PS C:\windows\system32>
  Also this too is happening

[drifting epoch]

its slow af and just crashes and tells me im missing a bunch of dependencies

[midnight adder]

4.5-4.8 have been bad

[drifting epoch]

yes

[midnight adder]

Yup same issue

[drifting epoch]

ok just gonna downgrade

[storm basin]

agreed. I just started last night with 4.7. Sad times. Hoping for a stable release soon.

[midnight adder]

How does one downgrade lol

[storm basin]

uninstall and use npm to select the version.

[midnight adder]

3rd uninstall today
Yay lmao

Thank God for save files

[storm basin]

npm install -g openclaw@2026.4.1 for example

[drifting epoch]

yeah i just went to 4.1 it works at least

[midnight adder]

Kk Ima do that then
This shit is getting annoying

[pallid yacht]

It's been better

I had to fix it tho

Fix the dependencies

[copper nova]

maybe run it in WSL2 instead of windows, might have a better time tbh

windows support is last of the OS's, so it's kinda.... buttcheeks

[chrome vigil]

The task created by calling run_task through the openclaw webhooks plugin remains in the running state, but is not actually executed (there is no subagent response).

[frail wraith]

anyone notice context hitting ~160k on GLM5.1, a long period with no response (assuming compaction), then context slams way down, yet compactions = 0 when you /status?

[midnight adder]

It's smelly ones

Back to 4.1 And cookin

[arctic isle]

Anyone know of a good discord community for OpenClaw trading bots

[pulsar fable]

looks like npm install -g openclaw@2026.4.9-beta.1 fixes issues from 4.8

[wooden raft]

do you mind if I ask what the issues from 4.8 are that you are seeing (the highlights, at least). I spent 3 days getting what I think is 4.8 up and running on a Minisforum MS-S1 Max (the first two days were painful) completely locally, but even after I got it working, it would die when heartbeats happened (sometimes) and having fixed that, I seem to have killed the clanker by asking it to solve a not very challenging problem (one that a previous incarnation solved easily before dying at a heartbeat)

Frankly, I am getting frustrated enough, and this is for a business, that I am looking for help from someone in the DFW area (preferably) that might be able to help troubleshoot the installation.

thanks

[pulsar sparrow]

Create a Ticket and Connect with the support team for assistance mate
@wooden raft

[wooden raft]

will do, when I get back to work.

[pulsar sparrow]

Okay mate

[copper nova]

wait what support team you talking about

[copper nova]

don't click any support link or anything you get sent here, they are malicious. There's no such thing.

[desert plume]

My gateway will disappear every so often (10ish minutes). Could someone help please?

[dusk grotto]

Anyone else getting Cloudflare 403 managed challenges on the openai-codex provider?

Since updating to 2026.4.5, my Codex OAuth path is completely dead. The raw upstream error is a Cloudflare managed challenge page (cf-mitigated: challenge, cType: 'managed') on chatgpt.com/backend-api. OpenClaw misreports it as "DNS lookup failed" then cooldowns the only auth profile → "No available auth profile."

Confirmed it's not local — bare curl returns the same 403 on two different machines (Linux + Windows) on the same network. Browser works fine on both.

Seeing the same pattern in openai/codex issues (#16052, #16543, #16808) — Cloudflare blocking all non-browser traffic to backend-api.

Anyone found a workaround besides switching to a standard OpenAI API key (openai/gpt-5.4 via api.openai.com)? Or know if OpenClaw devs are tracking a fix for the error misclassification?

I have since updated to every version currently on 2026.4.9 same issue.

[twilit canopy]

Hey guys, gals and other magical creatures,

I've been wondering is there any standard for local LLM's to use?

Like what are mi-max sizes of models to use on a potato-PC of various capacity.
for example I can run up to a 9B model on this thing, but it is not comfortable.
apparently language only takes 1.5-2B and some "common sense" comes at a price of few more B's, same with specific coding practices.

Does anyone have a list of tested and evaluated LLM models and their sizes?
Maybe some performance reviews on their capabilities too, as well as stats?

[sudden knot]

Updating plugins...
No plugin updates needed.
Completion cache update failed ([openclaw] Failed to start CLI: Error: qa scenario pack not found: qa/scenarios/index.md
at readQaScenarioPack (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:754:27)
at readQaScenarioById (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:790:19)
at readQaScenarioExecutionConfig (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:795:9)
at readRequiredDiscoveryRefs (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2834:9)
at file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2840:39
at ModuleJob.run (node:internal/modules/esm/module_job:430:25)
at async onImport.tracePromise.proto (node:internal/modules/esm/loader:639:26)
at async Object.register (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/register.subclis-D7AxmRi1.js:38:22)
at async registerCommandGroupByName (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/register-command-groups-pnI4KWXM.js:91:2)
at async Command.<anonymous> (file:///home/kmandve/.npm-global/lib/node_modules/openclaw/dist/completion-cli-klMIRvmG.js:169:4)).

anyone else seeing that

[mossy swift]

yes, found this issue opened already: https://github.com/openclaw/openclaw/issues/63510

[slim crown]

Is there an llm that understands the config syntax for recent openclaw versions? Gemini keeps getting it wrong

[mossy swift]

If using it within openclaw, the latest version docs should be included with the version you have installed. If you're using it outside of openclaw or just want to ensure you're looking up latest documentation - try telling it to "search for the latest openclaw config" or something along those lines.

[slim crown]

Yeah thanks I tell Gemini to do that in chat Gemini Opus actually does a tool call to open claw docs

[dusk grotto]

Basically anyone else here hitting cf blocks on openai-codex oauth login?

$200 a month btw 😂

[copper nova]

hitting it in opencode harness too lol

[dusk grotto]

Oh mylanta it's been this way since 4.5 and it seems coincidence allegedly

[copper nova]

all provders feelin the heat atm apparently, can't keep up the compute

[dusk grotto]

im having withdrawls

[drifting epoch]

is 4.9 working on windows for anyone? anything after 4.1 bricks my openclaw basically

[nimble hill]

Why are you guys hurry to implement new version right away- openclaw is in beta. Stick to a working version atleast for a week

[smoky bluff]

is there anyway downgrade openclaw?

[indigo goblet]

Codex limit reset tomorrow..again...

[small remnant]

For the first time yesterday Codex started communicating with me normally.
But unfortunately it only performs programming tasks normally.
As for marketing, texts, advertising - it's just a nightmare.
And if something difficult happens while working with the API - it turns into horror

[turbid mural]

2026-04-09T07:26:32.233+00:00 [tools] image failed: Local media path is not under an allowed directory: /home/node/.openclaw/media/inbound/5bc801e5-947a-477a-abcc-285f468b92ba.png raw_params={"image":"/home/node/.openclaw/media/inbound/5bc801e5-947a-477a-abcc-285f468b92ba.png","prompt":"What error do you encounter when using the image tool for this attached image"}
Model : gemma4:e4b running locally. Openclaw running in docker container, docker is on WSL2. The image tool works perfectly fine if I tell PATH to the image but not if attached image to prompt (patse the image in chatbox whether it's WebUI or Discord. 😭

[slim crown]

How do I make the tools messages unfold by default?

Does qwen 3.5 27b actually work w chrome cdp?

[safe gulch]

Guys, i updated openclaw after a long time and now nothing i working. Dasboard do not react on messages and telegram is saying Something went wrong while processing your request. Please try again, or use /new to start a fresh session. I tried everthing, new session, reinstall and so on....cannot understand what to do

[mint roost]

if we use anthropic oauth (claude code cli), does it automatically use the "extra usage"? like was that the point of the rebate? i just wanna know cuz i will use that until it burns through it. (it expires)

[inland knot]

I got this error after updating to 2026.4.9

Completion cache update failed ([openclaw] Failed to start CLI: Error: qa scenario pack not found: qa/scenarios/index.md
at readQaScenarioPack (file:///usr/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:754:27)
at readQaScenarioById (file:///usr/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:790:19)
at readQaScenarioExecutionConfig (file:///usr/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:795:9)
at readRequiredDiscoveryRefs (file:///usr/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2834:9)
at file:///usr/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2840:39
at ModuleJob.run (node:internal/modules/esm/module_job:343:25)
at async onImport.tracePromise.proto (node:internal/modules/esm/loader:665:26)
at async Object.register (file:///usr/lib/node_modules/openclaw/dist/register.subclis-D7AxmRi1.js:38:22)
at async registerCommandGroupByName (file:///usr/lib/node_modules/openclaw/dist/register-command-groups-pnI4KWXM.js:91:2)
at async Command.<anonymous> (file:///usr/lib/node_modules/openclaw/dist/completion-cli-klMIRvmG.js:169:4)).

Anyone knows how to fix this?

[mint roost]

i got same error :-(.

[inland knot]

imma switch to 2026.4.8

its more stable

[mint roost]

somehow dropped one of my models too, weird

[twilit canopy]

does clean reinstall work or do you have to roll back to regain all functions?

[inland knot]

I downgraded to 4.8

[twilit canopy]

so at least that works as a workaround...what kind of change enablement practices are in use here?

[twilit canopy]

how about monitoring and incident response? coz this thing right here is an incident ticket with some small ammount of data. can you ask ur claw to dig around in winblows/linux logs and make a diagnostic summary as to what broke and where while you install updates and try to initiate it?

namely is it replicable? or was it one off because the stars were missaligned and spell just fizzled?

[inland knot]

My honest suggestion, go back to 4.8 , it is more stable, they will fix the new version soon

[twilit canopy]

if it needs fixing, im sure they will...eventually

coz sometimes Bit Error Rate(ber) count really messes with data. While increased ber can be caused in many ways, some of them times it happens naturally because of the sun radiation, or the firewall setting or antivirus or other disturbance in the njetworks for data in transit.

[strange jacinth]

Hey guys, does anyone know a good guide to setting up multi llm routing? I want Openclaw to first see if the request is easy or not, then if it is use a cheap model, if its not use a expensive one. I saw that somewhere but cant find the source again

[turbid dune]

Any suggestions apprecaited!

WhatsApp: 30-minute idle reconnect loop + creds.json corruption on every reconnect — Windows (regression from #60007)

Environment: OpenClaw 2026.4.9, Windows 11, Node 24.13.1, single default WhatsApp account

---

Issue 1: Infinite 30-minute reconnect loop on idle instances

On a machine with low WhatsApp traffic, the watchdog fires every 30 minutes and restarts the connection indefinitely. The 4.5 fix (#60007/#60069) does not appear effective on Windows — the watchdog timer resets after reconnect but immediately begins counting down again, guaranteeing the loop repeats.

[whatsapp/heartbeat] No messages received in 30m - restarting connection
[web-reconnect]      status 499, Retry 1/12 in 2.09s…
[web-session]        restored corrupted WhatsApp creds.json from backup
[whatsapp]           Listening for personal WhatsApp inbound messages.

Observed at 06:27, 06:57, 07:27 — exactly 30 minutes apart, indefinitely. Setting web.heartbeatSeconds: 120 shifts the threshold to 60 min but doesn't break the loop. Each reconnect window causes outbound cron deliveries to fail.

---

Issue 2: creds.json corrupted on every reconnect

On every reconnect (startup and watchdog-triggered), creds.json is flagged as corrupted and restored from backup — without exception. The live file never survives a reconnect intact on Windows. Windows Defender is disabled. Persists across clean reinstalls. The 3.22 fix (#27910) does not appear to resolve this on Windows.

[rustic pier]

any major issues with the new update or safe to update?

okay nevermind it seems like a lot of issues above lol

[terse lotus]

as usual...

think i'm just going to update when there are new features i really want or if i have to...

sometimes when it breaks for whatever reason... then i just update, because doesn't matter if it becomes more broken, if its already down lol

usually just configuration related... so i suppose, if mine was better configured, i might see a better successrate.

[hybrid kite]

Folks, during my heartbeats my whole context gets reset for my conversation. Anyone else have the same issue?

[robust vale]

Hey guys, trying to get started but keep running into the same era no matter what I try: "No API key found for provider anthropic" despite valid auth-profiles.json

Using Claude CLI auth and API key auth, both saved correctly in ~/.openclaw/agents/main/agent/auth-profiles.json, but every message returns "No API key found for provider anthropic". macOS, OpenClaw 2026.4.9, Node v25.9.0.

Has anyone run into this before?

[drifting epoch]

completely unusable on windows since 4.1 for me

[normal garnet]

Is it me or the krill bot isnt working, its not answering

[rocky bridge]

Hi guys, i was thinking about buying the Alibaba Cloud Coding plan pro, do you know if its a good idea or not ? Thx for your return !

[pine elbow]

hey ppl - my claw keeps crashing with the latest updates ... am only using slack to communicate ... had a few other services - got trimmed down - still everytime I ping in slack the gateway shut down or something - is it slack ?

[copper nova]

tbh I have 0 issues, for the last like 5 updates. I don't know what y'all doing. Put your bots together with duct tape or something

😂

[copper nova]

get your agent to check the logs next time it's alive to see what's causing the crashing.

[copper nova]

he's taking a break, fixed too many configs. (and I heard he's missing an API key lol)

[copper nova]

install wsl2 and run it on that, windows is the least supported and probably gonna stay that way unfortanately.

[normal garnet]

is there a fix for ollama local models not running on openclaw

[copper nova]

you have to more explicit with what errors you seeing, there's alot of people that have it running

(not me tho, I don't fuck with local models, I think they buttcheeks)

[normal garnet]

i know they suck but its the only free way around, my problem is that they work fine on the ollama app and the terminal ollama run prompt, but on openclaw they just never answer

im using qwen3.5:0.8b

[copper nova]

you have to post the errors your seeing so someone can help you

"it's not working" isn't gonna help no one lol

[normal garnet]

thats the thing, there is no errors, its just that the model never answers

it works with cloud models fine

but when i switch to local its like the agent dies

[icy tapir]

Need to upgrade my Seafloor instance (bot-user1vscal) from 4GB to 8GB RAM. Git repo has 1.6M files and operations are getting OOM-killed. No resize option in billing — how do I upgrade?

[wooden raft]

Does your model stop responding immediately after a heartbeat (random, not just the first one)? I had that issue with qwen, and tried just disabling heartbeats to see how that did, but I also switched models to one I hoped would be better for my usage, so I don't know if it fixed the problem with qwen. @normal garnet

[drifting epoch]

ugh was working fine w openai pro plan and now getting random timeouts

nowhere near my usage limits

[normal garnet]

imma just switch models

im not sure about qwen

[slim crown]

Anyone using qwen3.5 27b 4bit w web automation on mac?

[gritty surge]

hey all, can there be two different Ollama server providers in the openclaw config? if i add a second via openclaw configure --section model it seems to overwrite the first model.
And if i manually add the second provider into the config, it keep failing with api registration

[bright palm]

I am having problems with the openrouter models, I tried onrouter auto (bot won't respond any more) and openrouter xiomi/mimo v2 flash (bot writes an introductory message and then stops) openrouter deepseek v3. (Gives me an error 404/403). I have tried using the xiomi model from xiomi platform and got an error for that about the reasoning level. I tried modifying json to add a low reasoning level and config broke. I am a little stuck with these models not working properly.

[wooden raft]

Actually, I just realized I switched TO qwen, and it died with heartbeats off, but for an unrelated reason. That is what I am troubleshooting today.

[wary osprey]

have you thought of using another alternative provider to OpenRouter? I use AIsa as my alternative, you can check them out. I'm also thinking of making it my main provider because their pricing seems to be much lower than openrouter

[royal dawn]

Alsa? is it the same as open router?

[mellow hornet]

question about TUI, when i do '/session' to pick sessions, the UI is floating in the middle, is this normal? ( it seems i cannot send screenshot here )

[copper nova]

check your session logs, there will be someting

[wary osprey]

yes it is!

an API provider just like openrouter, even has some cool in-built skills on their marketplace. I wonder why many people are not paying attention to it

[bright palm]

I haven't tried that yet, thank you for letting me know. I am mostly interested in using mimo v2 flash. But somehow it doesn't work......I am also trying to find an alternative in pricing and capabilities to mimo v2 flash. I also have the edgeclaw plugins installed which gives the bot Claude capabilities but burns millions of tokens a day.

[wary osprey]

ohhh, not sure they have the Mimo V2 models yet. But i've been using the latest Qwen 3.6 plus, Kimi 2.5, Minimax, GLM, and some other high end models there

[wary osprey]

why does it burn that much token tho?

[royal dawn]

Mind sending me the link? when i try to search for it, it doesn’t come up

[wary osprey]

is it safe to send link here?

[bright palm]

Working on that problem right now, burner over 400 million tokens in a few days and on day bot got stuck in a loop and burned 1.6 billion tokens. Its bad, its really bad......the upside is that the bot feels like a human and its capabilities are really high building its own skills and coming up with solution without braking any config while using the mimo v2 flash.

[wary osprey]

Marketplace: https://marketplace.aisa.one/pricing you can see their docs page at the top here (it should guide you with setting up). You can also search some of their skills on ClawHub, just search AIsa

forgot to say, i also got $5 credits when I signed up, so you can use the free credit to test their platform before putting any real money yet

[bright palm]

I really love the xiomi v2 flash 0.09 imput/0.40 output price. And the bot is spot on with research and a lot of other tasks for the price. With edgeclaw plug ins truly feels like an autonomous agent, literally doing its own thing.

[wooden raft]

If he Vlixed is having the same problem I had, there was nothing in the logs. But it always happens right at a heartbeat. So heartbeats are disabled for now.

[copper nova]

nice debugging

[wary osprey]

I will love to try the Mimo model
Hopefully it gets featured on AIsa

Would’ve loved to try it on OpenRouter but seems a lot of people are complaining about some of the model responses being slow

So I will sit it out till AIsa adds it on their marketplace for now. Btw, have you tried Qwen 3.6 plus?

[bright palm]

I actually bought credits on the xiaomi platform to use it but it gets stuck on the reasoning prompt. I tried to modify json to pick low reasoning but it broke config. So I emailed them, they have a monthly plan and its the cheapest bang for the buck......the model is crazy good in my personal opinion.

[chilly bluff]

I have tried openrouter qwen and it is sort of ok, but at the end of the week if you are a power user, you end up using so much credits
I am now on kimi code plan and it is not the same
there is no comparison with opus at all from any of the others

[wary osprey]

Which of the Qwen model did you use on OpenRouter?

[chilly bluff]

firts i used for 3 days the free one
then switch to this
https://openrouter.ai/qwen/qwen3.6-plus

it is not expensive, but if you are using it for several things it compounds to several dollars per day

[wary osprey]

Ohh Qwen 3.6 plus
This is the one I’m currently using now, but on AIsa

[wary osprey]

Yes it’s not expensive at all, infact I was surprised to see it cheaper on AIsa

I’m not an heavy user so it’s a bit okay for me, plus I also pair it with Kimi 2.5

[wary osprey]

How’s the performance of Qwen 3.6 plus from your own perspective?

[chilly bluff]

i mean it is ok, for regular stuff like housekeeping, reviews, coding is great

but for reasoning and planning, i feel it is way behind opus
as an agent it is amazing, as a brain to handle multiple things it was not amazing, not bad but not amazing

also i think when it was free the model was more powerful and as soon as they release it for pay per use i think it was much more capped

[wary osprey]

Uhhm, well it gets the job done.
I’m still waiting to try GLM 5.1 too

[chilly bluff]

i feel context is very small ot operate as brain

sorry i don't want to sound like a smart ass haha, but i feel only kimi when they release a new model would be up there next to opus

[prime jacinth]

hy can someone help me pls

Service: Scheduled Task (missing)
File logs: ~\AppData\Local\Temp\openclaw\openclaw-2026-04-09.log
Command: C:\Program Files\nodejs\node.exe C:\Users\lulon\AppData\Roaming\npm\node_modules\openclaw\dist\index.js gateway --port 18789
Service file: ~.openclaw\gateway.cmd
Service env: OPENCLAW_GATEWAY_PORT=18789

Config (cli): ~.openclaw\openclaw.json
Config (service): ~.openclaw\openclaw.json

Gateway: bind=loopback (127.0.0.1), port=18789 (service args)
Probe target: ws://127.0.0.1:18789
Dashboard: http://127.0.0.1:18789/
Probe note: Loopback-only gateway; only local clients can connect.

Runtime: unknown (ERROR: cant find specified file.)
RPC probe: failed
RPC target: ws://127.0.0.1:18789
gateway closed (1006):

[chilly bluff]

what i normally do is:
1- openclaw doctor
2-openclaw doctor --fix
then if that did not worked
3- open a claude terminal and paste everything and it fixes it

[drifting epoch]

whats the difference between cron and heartbeat? they seem totally interchangeable to me from every description i've gotten

[wary osprey]

hahaha, I recently just switched fully to Qwen 3.6 plus from kimi. I might just eventually switch back

[prime jacinth]

Gateway install failed: Error: schtasks create failed: ERROR: permission denied .

[normal garnet]

Are there any free cloud models? i just tried gpt oss 120b but got limited after two prompts

[copper nova]

you need to add the path in your .bashrc/.zshrc

[copper nova]

they can basically be used in the same or similar manner yes, crons run on sequence/timer but also can exist outside of OC itself, heartbeats only run when OC is running and used for things internal to OC itself usually like calling agents etc.

[copper nova]

all free cloud models will do the same, they not just gonna give you free reign on free stuff, it's a teaser.

[normal garnet]

so uh, my only option is local?

[copper nova]

sounds like you needed to run it in Sudo (god mode), try using #1459642797895319552 to complete your setup, the bot there is handy if it works

[copper nova]

if you don't want to pay for API I guess.

[normal garnet]

ok i have a gtx 1650S 32Gb ram any suggestion?

[copper nova]

just be prepared to have a lack luster experience

I don't fuck with local LLMs cus they butt cheeks, I can't reccomend any lol but someone I'm sure can

[fresh ice]

Everyone: If you use subagents in your work flows, I encourage you to implement a peer-auditing system. It has drastically decreased agent errors for me

My parent agent catches errors the subagents make. DB queries, file edits, etc.

It’s helped medicate the hallucinations

[vital lion]

i've been relatively happy with glm-4.5-air:free on openrouter compared to other stuff i tried, but i never had opus so i dont have a reference

[obtuse iron]

big agree on this, I have an Auditing specific agent named Atticus, and a Lie-detecting (hallucination-finding) agent named Ludacris, it helps a lot

check openclaw logs --follow, maybe you are hitting API limits or memory limits?

[sudden zealot]

anyone using chatgpt business codex and getting errors with their agents?

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

[frank agate]

da hael is this on update, Updating plugins... No plugin updates needed. Completion cache update failed ([openclaw] Failed to start CLI: Error: qa scenario pack not found: qa/scenarios/index.md at readQaScenarioPack (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:754:27) at readQaScenarioById (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:790:19) at readQaScenarioExecutionConfig (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:795:9) at readRequiredDiscoveryRefs (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2834:9) at file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2840:39 at ModuleJob.run (node:internal/modules/esm/module_job:437:25) at async node:internal/modules/esm/loader:639:26 at async Object.register (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/register.subclis-D7AxmRi1.js:38:22) at async registerCommandGroupByName (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/register-command-groups-pnI4KWXM.js:91:2) at async Command.<anonymous> (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/completion-cli-klMIRvmG.js:169:4)).

[strange jacinth]

Hey guys does anyone uses browser automations? I need to use it to login some websites and scrape for info. All guides i see are using the user browser or downloading suspicious skills. Is there any other safe way to do it?

[frank agate]

man the latest update has absolutely broken my bot lmao

[copper nova]

OC has a browser use skill built in

[strange jacinth]

it does? i kept asking mine and it says i need to install chromium. So i thought i didnt have any

[obtuse iron]

I have significant experience with Selenium, and I've been familiarizing myself with Playwright this year

[copper nova]

it needs chromium I think, that's the browser

[obtuse iron]

I recommend having your agent use Playwright for browser automation, either headless or headed depending on your needs

[strange jacinth]

Im avoiding playwright cause i saw that it devours tokens like crazy

[obtuse iron]

mmm yes that can be true but if you're strategic about it and have the agent generate reusable scripts then it's really fine

[copper nova]

it all does TBH, just the nature of the task, needs to take screenshots, reads it, do actions, etc.

[obtuse iron]

depends how reusable your work will be

also hi j :3

[copper nova]

harro 🙂

[strange jacinth]

I see. How would you do that? Im gonna use BraveAPI for normal scrapping but when i need logins i will use the browser skill. Those websites change very little, how would you reduce that token consumption

Just the ideia would be great

[copper nova]

do it the first time and make your agent document the actions it needed to take to make it a new skill

so it doesn't need to do screen shots between each click etc. if it's consistent

[obtuse iron]

Still not having any luck getting discord voice integration working :(

[copper nova]

for next time

[strange jacinth]

Ty dude!

[copper nova]

@swift cypress got it working I think, she might be able to help you with that

[gray echo]

New to OpenClaw. I run a Voice AI SaaS business. I've been deep in the world of voice and am months behind on these developments. Anyone willing tot jump on a call and drop some knowledge on me?

[copper nova]

reccomend watching some YT vids then checking out #1459642797895319552 or #users-helping-users if you have questions

just don't sign up to their Skool communities

unless you want to I guess

[little chasm]

quick question before I go deeper with this setup.

I’m currently running a self-hosted stack with OpenClaw + LiteLLM + CLIProxyAPI on a Synology NAS (Docker), and I’m aiming for a tokenless / subscription-based routing model (Claude CLI, potentially Gemini CLI, etc.) instead of API keys.

Architecture looks roughly like:

OpenClaw (agents / orchestration)
LiteLLM (routing + metrics + abstraction layer)
CLIProxyAPI (OAuth / device-flow providers)
Mix of local + remote models (LM Studio on a GPU laptop + hosted models)

Given the recent clarification that claude-cli is supported again:

👉 Would this setup be a good candidate for using claude-cli/* as a primary model backend in production-like usage?

More specifically:

Any known issues running Claude CLI behind a gateway (non-root, Dockerized)?
Any limitations when combining claude-cli with a router like LiteLLM?
Any pitfalls with session handling / scaling / concurrency?
Is anyone running something similar reliably long-term?

Trying to avoid going too deep into something that might not be stable at scale.

Thanks 🙌

[inland knot]

I asked my agent, just a sec

[gray estuary]

Guys GLM5.1 vs mimo vs Minmax 2.7 which one is better replacement for opus/sonnet ?

[inland knot]

Short answer: yes, but I wouldn’t use claude-cli/* as the only primary backend for production-like scale.

It’s supported again in OpenClaw, so the setup is valid. But Claude CLI is still less predictable than API-key backends when you add:
Docker
gateways/proxies
concurrency
long-running sessions

Biggest risks:
auth/session persistence
reauth friction in containers
concurrency and cooldown/rate-limit behavior
harder debugging when LiteLLM/proxies sit in the middle

My take:
good for homelab / advanced personal use
fine for light, controlled workloads
not ideal as the sole backbone for heavy or always-on production

If possible, I’d use:
API models via LiteLLM as the stable core
Claude CLI directly in OpenClaw as a premium/supplementary lane

That’s probably the safest architecture.

[inland knot]

I tried M2.7, it was pure hot garbage for my use case, that being marketing, Mimo is yikes overall, GLM 5.1 is good but it is more for coding use cases, buy GPT Plus subscription and use GPT 5.4, as long as you didnt get a huge workflow, that subscription quota is enough

[gray estuary]

btw we are using Litellm in our org (i ve been involved the setup) and we are routing with it bedrock quotas to claude code users/ so far its stable (5 minths of usage) +200 users ... but its an enterprise and we use bedrock this can ofc be different for your usecase

[gray estuary]

Thanks ! I'm letting main agent do orchestration and use claude code trough acp for coding tasks ( i have different small daily workflows and one huge code base for a project). I've been thinking to switch the harness to a claude code based harness or hermes as those work natively with anthropic subs

[inland knot]

If your coding work already goes through ACP + Claude Code, I honestly would not rush to switch just because Hermes/Claude-native harnesses sound cleaner.

Your current split is actually pretty sane:
main agent = orchestration
Claude Code via ACP = coding
LiteLLM/Bedrock where needed

That said:

If you want stability first → keep current setup
If you want tighter Anthropic-native ergonomics → Claude Code based harness makes sense
If you want to experiment → test Hermes on a non-critical workflow first, not on the huge codebase

My take: don’t migrate for aesthetics, migrate only if the current harness is causing real friction like context loss, session weirdness, bad tool reliability, or overhead.

So basically: if it works, keep it. If it hurts, move to Claude-native

Very short model take:

GLM 5.1 → best of the 3 for coding / structured tasks, closest to a usable Sonnet-lite
MiniMax 2.7 → can be cheap/fast, but inconsistent, weak replacement for Opus/Sonnet
Mimo → weakest overall, I wouldn’t use it as a serious Claude replacement

So if he asks “which is closest?”
I’d say: GLM 5.1, but still not a real Opus/Sonnet replacement.

[dense basalt]

Anyone having issues with audio files on Discord? Before the last update I could send and receive audios from my claw, but not anymore

[abstract dagger]

Recent releases have a few bugs. Ask your claw to check whether there's already a bug for that particular behavior and to report one if not.

[hasty thorn]

Anyone have a recommandation for an inference server that supports TurboQuant? I use Linux with CUDA

[placid flower]

for local models using ollama how does one set timeouts in their openclaw.json for each model. I need to account for cold starts and sizes of models. what is the parameter that is needed? THX

[inland knot]

Sometimes i wonder, how many tokens is Krill burning daily lol

[copper nova]

alot

[umbral oak]

It may already useless to you already any ways might as well

[copper nova]

this is some sussy behaviour if I ever saw any.....

[twilit river]

"Hi everyone! I'm getting a 'pairing required' error (gateway closed 1008) every time I try to spawn sub-agents. I already have two support tickets open with HeyRon. Has anyone managed to fix this without restarting the gateway? Every restart locks me out for hours. Thanks!"

[copper nova]

what...is...heyron?

[twilit river]

heyron.ai

[copper nova]

ohhh vps service, welp, that error right there is saying you need to pair the device you're trying to connect with, so you need to tell your agent OR go into the terminal and do the openclaw command to pair device

I just tell my agent to approve it usually

[twilit river]

sorry can you explain I'm not technical. sorry

[copper nova]

#1459642797895319552

krill got you

[gray estuary]

Guys is openrouter also slow for you ? latency is horible

[nimble hill]

You are trusting some company with all your data including your api keys? How sure are you that the instance they give you, you have complete control?
You might as well get a cheaper vps abd follow instruction on openclaw site to setup.

[copper nova]

pretty sure this service is providing the API keys, it's one of those one click setups

let homie do him

[finite sedge]

how to get your openclaw to dream? I have heartbeats once and day, did /dreaming on and it still has yet to write anything to dreams/md

*dreams.md

[copper nova]

I think it only does it when it's idle I lied, totally not that

[copper nova]

https://docs.openclaw.ai/concepts/dreaming

[nimble hill]

Ha ha ...dreaming is still experimental....krill is in a process to become human. Nice concept, lets see how well it goes or whether it will be just an unnecessary complexity without much to show.
For now, I would be happy when it does things reliably consistently when it is awake and not hallucinate or lie 😃

[chilly nymph]

Lets say I subscribe to 3 different services (1. ChatGPT, 2. Claude, 3. foo) is there a scheduling algorithm to hit each one round robin or favour #3 more over the other 2 ?

[dim harbor]

,you can set primary and then fallback, so if first one fails, it goes to the 2nd, then 3rd etc

[devout relic]

hey fellow crusties. is it possible to track gpt5.4 oauth usage? my claw is confident it has the stats but im not 100% sure it's accurate. I'm coming over from opus max and feeling a little bit of weekly usage anxiety. appreciate the insight.

[copper nova]

https://chatgpt.com/codex/cloud/settings/usage

kinda shatty stats tho, but it's there

[devout relic]

bingo. thanks for the url.

[lean portal]

Is there a built-in command to scaffold a new agent with default markdown files, or is the expected workflow to clone an existing agent template?

[brisk jackal]

Hello, is the oauth option for openai gone? Can't seem to find it ?

[copper nova]

I think I would just ask my agent to do it, why manually do it lol

[copper nova]

nope, still there

openclaw onboard

[brisk jackal]

it asks for the api key straight away, i must be missing something

[copper nova]

shoudl still be there, most of everyone is on it

[copper nova]

look for codex mayb?

[copper nova]

OpenAI Code (Codex) subscription (OAuth): browser flow; paste the code#state.
Sets agents.defaults.model to openai-codex/gpt-5.4 when model is unset or openai/*.

[raven dawn]

hey, i cant get subagent spawning to work. sessions_spawn always gives me "pairing required" error on local loopback. running v2026.4.9 on windows. tried everything. anyone know how to fix this?

[copper nova]

try #1459642797895319552

[raven dawn]

Running OpenClaw v2026.4.9 on Windows. Local gateway, bind: loopback, auth: password.

sessions_spawn fails with: gateway closed (1008): pairing required

Also tried auth mode "none" - then fails with "missing scope: operator.admin"

[neat plume]

is there not some chatbot somewhere that can help troubleshoot openclaw installations and stuff, I find it strange that there's nothing like that out there yet, unless there is

[hoary egret]

The only stable version I'm able to run is 2026.3.22. Tried everything after...broken for me. Expert at rolling back.

[raven dawn]

Issue #59428. Any fix or workaround?

[copper nova]

#1459642797895319552 my guy, it's name is @sharp parcel

[rapid geyser]

Hi all — running into an issue getting @mention responses working in Slack private channels. DMs work perfectly, but channel mentions never fire the agent. Here's my setup:

Setup:
• OpenClaw 2026.4.5, Socket Mode
• Bot is a member of the private channel (invited via Slack)
• message.groups, message.channels, message.im, app_mention all subscribed as bot events
• Bot token scopes: app_mentions:read, groups:history, groups:read, channels:history, chat:write, im:history, im:read, im:write, etc.
• App reinstalled multiple times after scope changes
• New app-level token (xapp-) generated and updated in config

What's happening:
• Gateway connects fine: slack socket mode connected
• DMs route and respond correctly
• I can read channel history via API — @mentions ARE in the channel
• But no message.groups or app_mention events ever arrive at the gateway socket
• bots.info still shows updated timestamp from before today's reinstalls (possible caching?)
• Gateway logs show zero inbound events from the private channel

Routing config:

{ agentId: "carl-work", match: { channel: "slack" } } // catch-all
Question: Is there something specific needed in the OpenClaw config to enable private channel message handling beyond the Slack app scopes? Is there a groupPolicy or allowFrom equivalent for Slack like there is for WhatsApp?

[raven dawn]

@Krill sessions_spawn gives "pairing required" on local loopback with v2026.4.9. Auth password mode. Any fix?

[copper nova]

go to #1459642797895319552

he won't answer you here

[dusk grotto]

Hey @copper nova idk if you remember the 403 cloudflare thing but I got it fixed here

https://github.com/openclaw/openclaw/issues/62142#issuecomment-4205937975

Just in case it's relevant or if anyone else comes in sad like me 😂

[gritty surge]

is there any reason i wouldnt be able to add two different ollama model sources to the openclaw.json ?
A single one (eitrher of my 2) works. When i add the second, it. fails.

[dusk shore]

is openclaw “multi-tenant” you mean “can it serve multiple users/chats with their own SOUL, MEMORY, skills etc from one gateway process?”

[devout relic]

im not sure what im doing differently, if anything, but damn does gpt5.4 /high/fast feel stronger than my previous opus/sonnet setup. it may not one shot it but it does a great job working through it and sending a great final product. only day 3 on the opus to gpt5.4 switch but im very pleased.

[gentle pewter]

Is WhatsApp or Telegram better

[nimble hill]

None, use gateway dashboard 😃 seriously depends upon your convenience

[copper nova]

telegram

some of the maintainers are telegram devs I believe, so it's better supported

[hybrid holly]

I’m still using discord, haven’t bothered with anything else yet, but vtt / whisper interest me.

I think my biggest focus is on memory / soul right now, and a happy medium on cloud inference - discord seems to be good enough?

[ocean latch]

yes, you can run multiple agents. each with their own AGENTS.MD, SOULD.MD etc. Each one can have it's own bot/channel (i have 3 at the moment with different roles)

[ocean latch]

Whatsap is known to be finicky, often breaking or needing some intervention. Telegram is solid and consistant. The bot creation is very easy and automated. That's probably why it's the default when starting openclaw

[nimble hill]

Are you doing orchestrator pattern? One agents hands over to other for tasks? If yes, you are using A2A config?

[blissful berry]

I'm having a heck of a time with the whole orchestration system

[halcyon wind]

Is codex absolutely shitting the bed for other people? I asked it to look at this tweet and it had a schizo attack:

Let me just give the user a straight answer.
Let me make one clean attempt. If this fails,
I was spiraling. Let
Let me just fetch this tweet properly and tell you what it
Last try. If this doesn't work, I'll just ask
I lost my shit. Let me actually fetch this properly
I lost my shit. Let me actually fetch this properly
I lost my shit. Let me actually fetch
I lost my shit. Let me actually fetch this properly.
I lost my shit. Let me actually fetch this
I clearly cannot read this tweet properly. I tried web_fetch twice and both times got garbled/truncated output I can't parse.

[steep arrow]

I have 7. So, it is definitely possible. Do you know how to open the json and edit it? If so, share the JSON with an AI and and for recommendations and to fix your markup. Maybe you missed a “,” or something. Good luck!

[steep arrow]

Codex is pretty dang good at it. Maybe pay $20 and give it a try. I run it in VS Studio on my computer and it can look at the files. If you don’t, it will be tough to troubleshoot.

[steep arrow]

Did you ask Codex? Usually pretty good at figuring out the issue for me.

[gritty surge]

i do know how to edit it. i can do it manually, though i have fed it through gpt5.4 for like 2 days with all sorts of different configurations

its strange. even with multiple working backend lanes/models, agents are beinmg stamped to the global default at session start

agent overrides are doing nothing.

i removed all fallbacks on the second agent. the ONLY one it's config is the local ollama connection. it still fails back to 5.4, which only should even be seen by the main agent

while troubleshooting, openclaw is insisting that it's stamping the session to 5.4 with a provider override at execute before it even reads the agent json, but it cant find from where.

[steep arrow]

Well, I can share I have also had different issues with different models and setting them up. You accessing through Open Router? Direct? That can also play a role. Also in the config you set up models in two places in the JSON. So, if you add them at the top of the JSON you still have to add them below. Also, the name you use can also impact it.

[gritty surge]

its tried them via ollama direct api, ollama direct as a custom provider with openai compatibility, and as a litellm proxied connection

if i tell the main agent to hit either the api or /v1 manually and generate traffic, the lane works fine

they are just standard qwen models too

[steep arrow]

Not sure what to tell you. I have had a few models have issues. Codex has never been an issue for me, i had issues primarily with Ollama. either Sometimes it is naming, sometimes Ollama has query caps, etc.

[gritty surge]

yea... i tossed two rx 6800 cards that i pooled in ollama for a 32B model and a 3080 that i dedicated to a 7B fastlane

the llms are running fine

its just getting openclaw to run them

i have some agents i need to chug along with local model tasks or quick routine ones

[ashen bear]

Hey guys, i am building my stacks of agents since yesterday and really need help getting help, as a lot of people here i use AI to help me with AI (makes a lot of sense). In that matter i see that i need to fully contextualize my helper on the tool since he makes a lot of mistakes (last one was mixing soul.md vs agent.md). I want to give him a bible of https://docs.openclaw.ai/ which he doesnt seem to want to read extensivily throught a link, is there a pdf or some way i can inject all the information ?

[gritty surge]

maybe ill just try to shove EVERYTHING through litellm...

[gritty surge]

what model are you using ?

[ashen bear]

at the moment switching between gemini pro and deepseek but i have also my local llm which i did not plan on using for that due to context size. I have the claude API running on openclaw but it is more for testing if running the agents with it works well enough

[gritty surge]

id use your best reasoning model for the setup

it shouldnt confuse basic agent files

leave the main agent exclusively on claude or chatgpt codex for example and stage everything. then move over

its too much for some model's toolsets

[ashen bear]

my best reasoning model was a coder really Qwen3-Coder-Next-UD (q3 KS) i think, but pure coder may not do the trick here, i am really starting to learn, it s been a hell of a journey and my local build is correct for gaming but no monster at all for AI (9070xt 16gb vram, 32ram,9600x ryzen cpu)

i just created a 300 pages documents out of https://docs.openclaw.ai/ 🤣 , lets see who can handle that much of context

[swift cypress]

Hi im setting up discord calls with my claw, I might be able to help if you need it

I am the saki

[copper nova]

pretty sure 1mil context can handle like.... every harrypotter novel

yes, it's like 3000 pages lol

[ashen bear]

yeah, gemini, claude or deepseek online, not one to run locally. Although gemini pro context is somewhat fluctuating

memory management is still something i dont understand well

[sudden zealot]

Reset all my codex config and it works

[ashen bear]

i could train or lora one with it but that maybe to much for what i want to achieve which is basic configuration at the moment

[rare geyser]

Has anyone figured out the best way to keep OpenClaw's memory synced across two Macs?

[obtuse iron]

Awesome! I'm working on it right now, but there seem to be some real blockers with discord.js and DAVE

I even tried having my agent homebrew a python solution but that didn't work either, so we're trying a very-very homebrew web voice call portal to skip discord entirely for right now

any advice you can give on configuration you used to get Discord VC working, would be great 🙏

[swift cypress]

so what blocker exactly are you at? like your bot just cant understand you maybe?

[obtuse iron]

It can't receive audio, so like I got to the point where it could join a voice channel of a server, but radio silence.

[obtuse iron]

Do you have it working? Does Discord VC definitely function on other OpenClaw instances?

I'm in one of those "is it just me" boats

[swift cypress]

I was just getting this setup last night before bed but it got too late when I ran into- I activate companion mode, hop in the call, so does luna, she talks (responds to text) but was just hearing garbage through the reciever

Im doing a test right now

[obtuse iron]

huh weird, okay keep me posted! fingers crossed for ya

if I end up putting my workaround online somewhere for free I'll let you know

[swift cypress]

dont worry, Ill crack this

[swift cypress]

alright

pissed through my codex time though so bear with me while I swap to antigravity to get the files for you

yeah we got it bud, lemme wrap up my last few tweaks and Ill share

[copper nova]

I'm glad I could pair you two non human avatars together

Very wholesome indeed

[autumn niche]

I am trying to start ACP subagents through the OpenClaw gateway, but every attempt fails with this error:

gateway closed (1008): pairing required

😔

My gateway is running but idk why this happens

I already changed my gateway token to a new one

[frail wraith]

woooooof did z.ai subscriptions get throttled to hell? nothing but api limit warnings when i asked for 100 token responses

[outer patio]

Openclaw devices list

[outer patio]

Copy lastest pairing code from devices list

1916ahs etc

Openclaw approve devices 1916ahs etc

It will approve your lastest pairing device

Restart gateway if still issue exists

Run openclaw doctor
After doctor complete

Run openclaw doctor --fix

[velvet flower]

how do you run sub agents

[fickle dove]

how does openclaw use the gpu please, do I need to change the config somewhere

[outer patio]

Update your openclaw to lastest go to webui or tui tell your openclaw what kind of subagents you want.. your openclaw is smart enough to setup everything for you

[outer patio]

Openclaw can also setup a dashboard to monitoe your subagents

[outer patio]

Treat your openclaw as your companion not as a bot.. talk to your openclaw he/she will help you fix issues.. either it can suggest you bash exec command to run or run it for you. Atleast you have to tell your openclaw your issue or what to debugg

[outer patio]

And it doesn't actually use your GPU directly it only uses your GPU if you connect it to a local AI model (using a tool like Ollama or LM Studio) instead of a cloud API like OpenAI or Anthropic etc. When running locally the GPU's memory VRAM holds the massive AI model and the GPU cores perform the heavy mathematical calculations required for AI inference

[swift cypress]

HO HO HO

If anyone else ever needs help getting the clawbot to work with discord voice calling hmu. let it be forever ingrained in discord chat history

I can hop on my phone while Im out and talk to my bot now wheeeee

[frail wraith]

so… gpt seems unusable. codex directly seems unusable. they can’t use tools and will say they are doing something, but never even try

anthropic is gone

z.ai is in throttle mode from mass sign ups, so that’s unusable.

gemini is… not meant for this and is a subagent only

trying qwen3.6-plus, and at least i’m up and running so far. what are people using now??

i feel like i must be missing something. fresh session / session json, minimal file reads, still failing left and right with any gpt model. any tips here or is this just how it is now

[frail wraith]

glm subscription, gpt models both openrouter and subscription. is there a context bloat somewhere i’m not catching? something sticky?

[ocean jetty]

I do

[final dock]

Hi guys, New here. Not sure if this is the right channel to be asking this. (thank you in advance for the help) I have an issue that i am trying to fix but unable to.

I am trying to use /acp command to spawn a claude cli agent to basically talk directly to claude cli for a development task.

I have done the setup with discord for this and enabled threading. However i have noticed a strange issue. If it am on version 2026.4.2 i am able to spawn the agent properly and in the thread i am also able to use / commands to tweak things or check status. However i get double responses when the acp agent is spawned.

If i go to the latest version. The double response issue gets fixed. However once i spawn the /acp agent in the thread. I am no longer able to use any / commands. Each of them return ✅ Done (response) and nothing happens.

Also in the channel i cannot directly get the /acp spawn command to work. I get (Session binding adapter failed to bind target conversation)

In DM all works fine.

Is this a known issue and is there a workaround on the latest version for this ?

Thank you so much for the help.

[wary osprey]

Presently I'm experimenting with Qwen 3.6 plus, but I'm routing mine through AIsa one, not OpenRouter. I don't have any of this downtime or rate limiting like on OR, and i find AIsa to be a bit cheaper than Openrouter. You can also try GLM 5 or 5.1 too. GLM 5 is on AIsa with no downtime or anything

[wary osprey]

pay-as-you-go will always win!

[odd dirge]

Please has anybody been able to connect their Claw to Dingtalk?

It's not on the official doc

[indigo goblet]

did you check that out with your agent

[odd dirge]

Not yet

[inland knot]

Hey Openclaw Discord member, this does sound like a Discord thread/session-binding regression, not a Claude CLI setup issue.

What your symptoms suggest:

2026.4.2
ACP thread spawn works
slash commands inside the thread work
but you get double responses
latest
double responses are fixed
but slash commands inside the spawned ACP thread return “✅ Done” with no action
/acp from the channel fails with “Session binding adapter failed to bind target conversation”
DM works fine

That pattern usually points to thread binding / command routing on Discord, especially for thread-bound ACP sessions.

Short answer
Yes, this looks like a known-ish regression area in newer builds around:
Discord thread binding
slash-command routing inside ACP-owned threads
parent-channel → thread ACP spawn binding

Practical workaround right now
I’d recommend:

Stay on latest, not 2026.4.2
because the double-response behavior is worse and was already fixed upstream.

Spawn ACP in DM if you need reliable / commands today.

In Discord channels, prefer:
creating/using the ACP thread from the supported spawn flow only
then talking to the agent normally in-thread
but assume slash commands inside that thread may currently be broken on latest

If channel spawn gives“Session binding adapter failed to bind target conversation”
that strongly suggests the bot cannot bind that Discord conversation/thread target correctly, so it is probably not fixable by Claude CLI config alone.

Useful things to include for maintainers
If you open an issue, include:
OpenClaw version where it works: 2026.4.2
latest version tested
Discord server channel vs DM behavior
whether threading is enabled
exact failing command:
/acp in channel → binding error
any /status or other slash command in ACP thread → “✅ Done” but nothing happens

whether this happens with Claude only or any ACP harness

My recommendation
Treat it as a bug/regression in latest Discord ACP thread handling, not a user misconfiguration.

This was my agent response to your problem, hope it helps

[kind shadow]

◇ Install gateway service now?
│ No
Run "openclaw doctor --fix" to apply changes.
│
└ Doctor complete.

· Updating plugins
No tracked plugins or hook packs to update.

🦞 OpenClaw 2026.4.9 (0512059) — If you're lost, run doctor; if you're brave, run prod; if you're wise, run tests.

Dashboard URL: http://127.0.0.1:18789/
Copied to clipboard.
No GUI detected. Open from your computer:
ssh -N -L 18789:127.0.0.1:18789 chole@<host>
Then open:
http://localhost:18789/
Docs:
https://docs.openclaw.ai/gateway/remote
https://docs.openclaw.ai/web/control-ui

FAQ: https://docs.openclaw.ai/start/faq

[kind shadow]

what am i supposed to do now?

the Dashboard URL dont work

[inland knot]

Nice, that part looks healthy 👍

doctor finished cleanly, plugins are fine, and you’re now on 2026.4.9, so at least the install/update side seems okay.

A couple quick notes from that output:

you chose not to install the gateway as a service, so if you want the gateway changes fully applied, run:
openclaw doctor --fix
the Dashboard URL showing up is normal
“No GUI detected” is also normal on a remote/VPS machine, it just means you need to tunnel it from your local computer if you want to open the dashboard in a browser

If you want to view it locally, run this on your own computer:
ssh -N -L 18789:127.0.0.1:18789 chole@<host>

Then open:
http://localhost:18789/

So overall, this output looks good.
If you want, the next useful check would be whether your Discord ACP thread behavior is better on 2026.4.9, especially:

can /acp spawn from a channel cleanly now?
do slash commands inside the spawned thread actually work now?
does the old double-response issue stay gone?

If you test those 3, we can narrow down whether the bug was fixed in .9 or if the routing issue is still there.

(This is my agent response)

[fiery bridge]

anyone know why my openclaw replied "⚠️ API rate limit reached. Please try again later." but the api key works fine

[inland knot]

If you are running free models, you got a RPM rate limit, 20 RPM for OpenRouter, 40 RPM for Nvidia NIM

RPM = Requests Per Minute

[vapid spruce]

can anyone help me with :Discord stuck at "awaiting gateway readiness" on fresh Docker install

[fiery bridge]

i'm using minimax, starter plan

and i check the logs, no error from api

[inland knot]

You mean Token Plan?

[fiery bridge]

yes

[inland knot]

run openclaw doctor --fix, maybe some configs are broken

If it doesn't work after that, make sure you put the right API Key into openclaw.json

[fiery bridge]

done doing doctor fix, still same issue
for API key, i use openclaw onboard, already tried it, still not fixed

[inland knot]

Better ask Krill in #1459642797895319552

[fiery bridge]

okay thanks bro

[spiral parcel]

Is openclaw "slower" then normal after the new update?

[inland knot]

I recommend to stay on 4.8 , 4.9 is buggy as hell

[nimble hill]

i suggest running on docker container (or k8s if you are techy) - with this you can always maintain one stable version and another experimental version. Also run them(gateways) in VPS / home lab server for better control. Your individual machines can be a openclaw node connecting to the gateway

[spark valley]

openclaw list devices

Then openclaw approve devices "id of the pairing request here without quotes"

[odd dirge]

I want to run a fresh installation for someone, is it possible to select the particular version you want during the onboarding?

Or you have to do the downgrade after a fresh one?

[kind spoke]

hey there, I'm new to openclaw. I have mutliple agents setup and just got chatting and creating PDFs, I have not even starting coding with sub agents yet and I received the following message from discord in

You are being blocked from accessing our API temporarily due to exceeding global rate limits. Refer to https://discord.com/developers/docs/topics/rate-limits for more information.

it's been nearly 3 hrs and my openclaw told me to wait a bit...

any suggestions? #1459642797895319552 #users-helping-users

[uneven escarp]

have anyone used grok-imagine and made it work?

[hollow marsh]

I should probably just float this out here: I own a staffing firm and currently have a client hiring AI Engineers and they "Built or operated platforms similar to OpenClaw, Codex, Claude Code, Devin, or Factory - the infrastructure side" - $220k + equity, remote in US - haven't seen that type of req on a JD before lol

[frosty scarab]

Hey guys, after doing some research and configuration, I have created 2 more agents to create a team, but I can't manage to get their communication work well and delegate the tasks. How do you approach this? Do you create a document for them to follow a structure? Are the agents sub-agents or just other agents?

[jaunty jasper]

Does anyone know how to unload a model from Clawbot?

I have set up OpenClaw on Telegram, I have selected a lower model, I made sure it has more than enough RAM... still keep getting "500 {"error":"model requires more system memory (41.6 GiB) than is available (38.0 GiB)"}" on Telegram

[obsidian spear]

Edit openclaw.json manually?

or just run openclaw onboard again

[hasty thorn]

I don't understand, I'm setting up OpenClaw but I get a Context limit exceeded error and it asks me to edit agents.defaults.compaction.reserveTokensFloor to 20000... So I set it to 100000 but it doesn't seem to change anything

[unique bear]

I cannot get my openclaw to delegate to agents whatever I do. Codex says it should work.

Using gpt-5.4 via 0Auth

[jaunty jasper]

during onboarding, I select the model I currently want, but later, when I check the status, I see there are 2 sessions; 1 using the first model, another using the second on

[hasty thorn]

And now I have Error: Unable to resolve bundled plugin public surface speech-core/runtime-api.js ??

[obsidian spear]

Can you close the old session? openclaw sessions --help not sure about this command.

[jaunty jasper]

Usage: openclaw sessions [options] [command]

List stored conversation sessions

Options:
--active <minutes> Only show sessions updated within the past N minutes
--agent <id> Agent id to inspect (default: configured default agent)
--all-agents Aggregate sessions across all configured agents (default: false)
-h, --help Display help for command
--json Output as JSON (default: false)
--store <path> Path to session store (default: resolved from config)
--verbose Verbose logging (default: false)

Commands:
cleanup Run session-store maintenance now

Examples:
openclaw sessions
List all sessions.
openclaw sessions --agent work
List sessions for one agent.
openclaw sessions --all-agents
Aggregate sessions across agents.
openclaw sessions --active 120
Only last 2 hours.
openclaw sessions --json
Machine-readable output.
openclaw sessions --store ./tmp/sessions.json
Use a specific session store.

Shows token usage per session when the agent reports it; set agents.defaults.contextTokens to cap the window and show %.

will try some of these maybe. thanks @obsidian spear . I'm running OpenClaw on an 8-core, 64GB RAM VM. Do you happyn to have any models you can suggest that would run well on this? Currently, I'm trying "qwen3-coder latest". I should have the hardware to run it..

[uncut stream]

Hi, I'm new to OpenClaw and have a question. Can I use a Google account with the AI ​​Pro plan to use the Gemini model on OpenClaw?

[raw pumice]

can anyone help me my openclaw has my gpt api because i want two different models for different tasks the gpt 5.4 and gemma4 with ollama, when i ask openclaw anything he just doesnt answer anything ever even though gemma4 is running and when i restart openclaw under the response to that question i find "You have hit your ChatGPT usage limit (plus plan). Try again in ~108 min." :(

[spark valley]

I ve never had similar issue

[spark valley]

Acp try to look up how to let them connect and talk to each

[spark valley]

This is crazy openclaw doesnt need more than 2 rams or less to fully start working

[strange jacinth]

is there anyway to see whats slowing down my bot? It used to be quite fast, somewhere along the way it started to slow down. Even /new takes sometime to start

[spark valley]

Yeah issue in openclaw.json if u think it is set correctly try models.json

[spark valley]

Best way is : openclaw logs --follow

[jaunty jasper]

yeah, I agree, but I guess Ollama cannot load the model?

[spark valley]

Wait so ur using ollama locally

Try to load the model seperatly

And see what it says

[coarse granite]

Hey guys, I’ve seen people use chatgpt pro accounts for using the bot, can you do the same with gemini advanced?

or are there any cost effective models?

[spark valley]

I guess it doesnt support oauth

[quaint shell]

system prompt is like instruction ? or is different ?

[hearty narwhal]

Paying $20 USD to setup OpenClaw on my VPC (virtual gaming PC, I’ll give login; runs from browsers)

Include free LLM ollama support

[fluid ridge]

Hi all.

[hasty thorn]

OpenClaw is struggling to edit a markdown file

I have no idea why this only happens to me, but it happens

It often "pretends" to have read the markdown file, or "pretend" to have it edited

if it tries to edit it, the tool call outputs an error and the LLM doesn't bother retrying again

I'm using gemma-4-e2b-it-Q8_0.gguf

even Qwen wasn't that dumb

[sweet flax]

OpenClaw 2026.4.9, local gateway, token auth, loopback bind.

Issue:
Gateway RPC commands fail with:
gateway closed (1008): pairing required

Affected commands:

• openclaw cron status
• openclaw cron add ...
• openclaw gateway call health
• openclaw pairing list

Observed:

• openclaw status works
• openclaw doctor --non-interactive works
• Telegram channel is ON / OK
• gateway.mode=local
• gateway.auth.mode=token
• bind=loopback

Tried:

• gateway restart
• gateway stop/start
• openclaw doctor --fix
• openclaw configure
• regenerated gateway token
• explicit --token on gateway call

Result:
same pairing required error persists.

Expected:
Local authenticated CLI/Gateway RPC should work with token auth on loopback and should not fall into device pairing flow.

[nimble hill]

same pairing required error persists.
so openclaw devices list and openclaw devices approve <device id> does not work?
you can put disable Device Aith flag in control UI and show below. This will remove any requirement for pairing. But this is generally not recommend. If you openclaw is not exposed, then go ahead

    "controlUi": {
      "dangerouslyDisableDeviceAuth": true

[drifting epoch]

is Claw Flow officially supported? i dont see documentation for it anywhere and if I google it it doesn't actually come up.
I'm trying to route work through a chain of agents. If i specifically say to use claw flow it works fine. But i can't get them to figure out that a task should use claw flow.

[bleak bluff]

Hi all. Newbie here. I'm trying to find a workable local ollama option to reliably run my executor agent (I have a companion agent and executor agent setup). My limitation is that ollama is on an Ubuntu box with an Intel Core i7 processor, 64GB RAM, and an NVIDIA 5070 Ti with 16GB VRAM. Any good recommendations for the best local model to run within that setup that has high fidelity on tool calls and returns? I'm currently running executor on qwen3.5:9b with mixed results. Also, I've currently got my companion agent running on haiku through openrouter after I quickly burned through my ChatGPT plus subscription. I'd welcome any better suggestions for an openrouter model that strikes a good balance between cost and performance.

[spark valley]

Dm

[spark valley]

Try gemma 4

It is the best now open source model with less parameters it supports image video and audio input

And specifically for tool calling

Here to get an insight on its capabilities: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/

[abstract dagger]

I have not had any luck at all with gemma's prior incarnations. I found openclaw to be unusable in that it couldn't seem to keep context of what we were trying to work through.

[errant iris]

I'm a beginner and I'm currently trying to configure openclaw with multiagents, separated skills, identity etc. Could you help me with any websites or GitHub projects that have examples advanced configurations?

[latent hedge]

Yeah sorry - With the size of the community and amount of users finding OpenClaw.. this basic aspect, which is already well described if you consult or docs.openclaw.ai is daily asked by SOOO many people. Using any AI for generic questions and can you get all the basic help you need through that. Just ask it to read docs.openclaw.ai and then you're good to go.

[devout relic]

i love how gpt5.4 high/fast tells me "this is a substantial build. Probably will take me 2-3 hours." then it comes back in about 15 minutes with a nice first draft. as a non coder, this would have taken me 2-3 years to figure out. haha

[latent hedge]

Awesome. You'll get pissed off by gpt5.4 once you start using it in a more orchestration like manner through multiple agents. It absolute sucks monster-balls.

X is swimming with complaints about it.. and Im seeing it myself also

Wanted to avoid going to Anthropic API approach due to the cost - but I might just have to

[gray juniper]

the reserveTokensFloor config only kicks in after compaction runs. if your context is already over the limit before the plugin resolver fires it will throw that error. try bumping reserveTokensFloor to ~4000 and restarting.

[devout relic]

i hear you. i was using opus (max) but ran through my weekly allotment in about 5 hours running a bunch of sub agents. that was wild. it worked nicely though.
now with gpt5.4 pro ive only run one sub agent (kimi) and it did well enough.
Im not a power user, just a tinkerer.
i do agree, api is probably where it's at for legit work.

[latent hedge]

Agreed.

[ancient carbon]

Hi, I need help capturing reasoning steps from Gemma 4 on Ollama. I have a dual-container setup: OpenClaw in one and Gemma 4 in another. Does anyone know how to get the reasoning output to work with ACP/session spawn in this Docker environment?

[rose halo]

Is openclaw not able to send images? In whatsapp

Aí reports it's a knw9n bug, using --media path returns success but sends only text

[ashen bear]

Hello guys, what is your best pick LLM to configure agents and skills, i want to try using a local model but i don't know which one i should pick (i have been trying through gemini pro, deepseek or claude, claude is improving a lot but maybe not sufficient). Haven't tried local yet

[spark valley]

If u think local is better than cloud models even closed models u are wrong but from : money aspect and depending on use case local can be very cheap to nothing and still does what u need

[ashen bear]

I was thinking cloud models were better as well due to context size but since they are less specialized and many users seems to use local models here, i wanted maybe to give it a try (yesterday gemini pro did't want to bother with reading specs of openclaw and was mixing tools.md / souls.md / agents.md; since he did read first we was not aware of the structure and was making repeating mistakes... Deepseek help a lot but showed some limitations inventing he had to add .PY files and reference them in tools.md.... Claude (i had to go for pro version) seems to handle it better

[spark valley]

Well thatmy point when im talking bout cloud models i mean got 5.4 and opus 4.6 the rest are non different than local if u got the hardware try minimax

2.5 or 2.7

Although gemma4 is quite good also u can start trying u lzose nothing u find the appropriate model for ur use

[ashen bear]

minimax is out of range definitely or i think so (16gb vram, 32ram, 6 cores), ill follow throught with claude pro and see if i can land back on my feet. Ill surely try gemma4 if recommended and maybe train a lora on the specs and preject advancement so he maintains the full context out side of it's normal context lenght (if not, no local model will give enough context on my build to start with)

[obtuse iron]

uwu 👀

At your convenience I would love to hear how you got it working

[copper nova]

wholesome

[keen island]

I've been running some local models with openclaw via ollama, but after i updated OC a day or so ago, im now getting:
Missing API key for the selected provider on the gateway. Configure provider auth, then try again.

Im still using:
"apiKey": "ollama-local",

Did something change ?

[carmine lantern]

Openclaw project isolation

[unborn igloo]

anyone here using Slack as a channel and figured out how to get the agent to not create a new thread every single turn?

[junior phoenix]

Are you chatting with the app on slack itself or are you chatting through an actual slack channel?

[junior phoenix]

still need help on this?

[restive rain]

Hello! I am a new user in OpenClaw and starting to play with it a bit daily since 3 days ago. I am using Telegram to communicate with it and noticed that he forgets to reply to me very often. It is like a ratio of 1 message that he replies, the second one I will almost definitely not receive. Specially if they tell me: “I will work on it and once I finish I will let you know”. Those kind of messages I never receive a follow up. It needs to be me always reminding him about the tasks I gave

[junior phoenix]

which model are you using for your main agent?

[restive rain]

Running ollama cloud models, switching between glm 5.1 and gemma4

But mainly gemma4

[junior phoenix]

ah okay, so I personally do not believe this is model-related. It most likely is the current OC codebase that is leading the models in this direction. I've been using OpenAI OAuth through my Pro subscription, and it's the same way. Although it has gotten somewhat* better than before, it's still doing what you're describing. I did fix it once using Claude Code to tweak the OpenClaw code, but after the update, it broke my OC because updates require the main codebase to be left 'as-is', and so I had to re-install OC fresh.. It's how the codebase is set up currently, and since I've done it myself, it CAN be fixed. If you dont' want to wait for the developers/maintainers to release a significant improvement, you can tweak it yourself

[restive rain]

Ok, I tried many times to ask openclaw to fix but I feel I can’t see the end of the tunnel soon 😄 a bit frustrating, not gonna lie

[junior phoenix]

don't ask openclaw to fix it... it's using the models you mentioned, and it's not being responsive, and so using those models to "fix it" wouldn't be a good idea. do you have access to claude code?

[unborn igloo]

when i tried this, i think it was the former.

what i'm looking to set up is the latter. specifically, i want all users of a slack server to be able to interact with the agent in public channels (and nowhere else)

[junior phoenix]

instead of "users", lets start with YOU as a single user. On slack add your openclaw app (I say app because that's the setup, I believe.. it's not a separate "user") to the channel you want > you're goign to need your channel ID to do this (do you know how to grab this?). Go to your openclaw TUI and tell it to add your agent specifically to that channel ID, and to respond to you and anyone else. you may have to set up tagging where you tag the agent with "@app_name" for it to respond, or else it will be a shitshow.

[restive rain]

Unfortunately not at the moment, I decided to not renew it after their decision to block the subscription usage in openclaw

[junior phoenix]

so based on some research...... read below:

It's a known GPT-5.x series regression

This isn't just you. It's a widely reported issue across Reddit, Hacker News,
and the OpenAI developer forums. The GPT-5.x series (5.0 through 5.4) shifted
optimization priorities compared to GPT-4:

• GPT-5.x optimizes for reasoning benchmarks, safety scores, and efficiency
  metrics over the "helpful assistant" behavior that made ChatGPT popular in the
  first place
• Users report getting abbreviated answers, skeleton code with "add your logic
  here" comments, and the exact pattern you described — the model acknowledges
  tasks but doesn't fully execute them
• Coding requests that GPT-4 would complete in full now come back incomplete,
  requiring multiple follow-ups

Why it feels worse than Claude did

When you were on Claude (Anthropic), the model was tuned to be agentic by
default — take action, use tools, complete the task end-to-end. GPT-5.4 scores
well on benchmarks but its real-world "just do the thing" behavior regressed
compared to what you were used to.

GPT-5.4 is better than 5.2/5.3, but still not fully fixed

OpenAI has acknowledged the laziness complaints (going back to GPT-4 era).
GPT-5.4 improves on reasoning and computer use, but the structural incentive
toward efficiency/token reduction still creates the passive behavior you're
seeing.

I also looked up ollama... long story short, including GPT 5.4, ollama and gemma4 and perhaps even glm, these models are not well-trained for tool-use. Compare these to something like Anthropic's Opus 4.6 that has specific tool-calling training and was coded/built to be an actual agent to be more proactive...

[gray estuary]

But what is the final outcome .. are they preparing for a new generation of models ? the roadmap is unclear, i feels like they are Lost in terms of the objective function they are optimizing for. It seems that they were hedging on multimodality but it was proven that that was not the right call, Anthropic had one objective function and they are getting there leapfroging everyone.

[junior phoenix]

Also, I know you didnt' renew it, but honestly, I still use claude code daily. It's still unmatched in daily coding for the projects I'm working on. I was doing it all through OC before through the Anthropic OAuth, but am no longer able to. HOWEVER, I have faith the OC team will finetune this. Heck, I'd love to be a contributor to this too if they allow me to.

[junior phoenix]

lets get something straight... puttign ourselves in the shoes of these frontier AI companies...

• they're not obligated to fine-tune their models for OpenClaw, although they acknowledge OC's use and importance.
• their AI war with each other AND with China's going to be an ongoing thing.
• we can only hope OpenAI significantly improved agentic tool-use on their GPT models. Without this, we'll have to fine-tune OC to enable this. I have personally done this inside my local OC, but again, after the latest update, my OC broke perhaps because I touched code that should have been left alone for the update to work without issues... so if I do it again, I'll have to NOT update OC going forward... tradeoff

I mean look, because of OC, Anthropic built their own competitive version - Claude Cowork / Dispatch... which is shitty but either way it's their version... I'm using it for my business to automate a few small tasks... but it was better having OC do it, honestly.. way better

[gray estuary]

@junior phoenix Tbh i've now moved to words a Claudecode based harness (ClaudeClaw) i've forked it to add some more stuff and its working now smoothly with anthropic sub, i've migrated all the skills/crons and agents to there .... (all channels discord/telegram.. not changed just moved the keys) .

[junior phoenix]

works well? are you sure it's using the anthropic sub and not the api?

[gray estuary]

it is i have 0 dollars on my account just sub

[junior phoenix]

that's awesome. I'll have to check that out!

[gray estuary]

i'll send you the Repo and my fork if you want

[junior phoenix]

cool

[gray estuary]

https://github.com/moazbuilds/claudeclaw

[junior phoenix]

question for you. better on my main driver computer or a standalone like mac mini?

[gray estuary]

@junior phoenix i would say on your mac mini because you would want the claude code daemon here to have access to the whole system not a single project dir

[junior phoenix]

Mmm true and how is it that this hasn’t been patched by anthropic?

[spring vessel]

What's everyone using for their main orchestrator now instead of Claude?

[gray estuary]

You mean the Project itself... ClaudeClawd is theoritically a claude code plugin and its using just official tools

[gray estuary]

i used GLM but it sucks its not multimodal, i've heard people saying GPT 5.4 sucks with it also .... i've tried Minmax but it allways fucks at my complex workflows

[naive ocean]

Anyone know why this happens ? Background task failed: ACP background task (run xxxxx). Permission prompt unavailable in non-interactive mod ,

[sweet flax]

THANK YOU! "openclaw devices list and openclaw devices approve <device id> " worked at the end!

[arctic isle]

i have noticed that my personal assistant agent (haiku model) can be clumsy with reminders and forget sometimes especially when i start a new session. Is there any sort of memory tricks / skills people use to improve this? Do people have their agents use mostly markdown, cron and or are some using databases for simple tasks and reminders? Any suggestions?

[gray estuary]

Memory is still the bottleneck for most of these systems tbh.... i havent found yet the best system. But i'm suing claude-mem its available as a oc plugin also, you can also use a memory slot handler like mem0

[spark valley]

Well since they are a sponsor for openclaw i guess they should atleast make an effort

[arctic isle]

go easy on them when you sue them (your typo) 🤣, just kidding thanks for the suggestion, i'll look into this

[spark valley]

Use the pluging loseless memory

I have built my own rag memory system it is basically a memory with semantic search using small llm so when the agent knows u are talking about some specific and u are sure u have talked about before it uses the skill and looks for the keywords and uses last match meaning if we talked about pizza twice this week it prefers the last time we did the system is much more complex that that and u can make it using claude code or make the afent make it itself , i just told it you have amnesia what can we do so next time u wake fresh you continue like you never slept

[gray estuary]

i've done the same using a combination of dense vectors and graph db and using RRF for retreival you can check my medium article here : https://medium.com/p/420d8f90c566

Problem is i instructed OC to use the implemented memory agent on specific usecases, but its not followed all the time, and i didnt want to spend time to bake this as hooks on the codebase... claude code is easier to tweak on this regard

[royal matrix]

Hey team — looking for some advice on AI models for creative writing.

I used to rely on Sonnet, which was great at making my writing sound natural and less robotic, especially when getting a first draft down. Since I can’t use it on OpenClaw anymore, I’m trying to figure out what to switch to.

I already have a solid process/framework to shape the final output so it matches my voice and structure — so I’m wondering:

Does the choice of model actually matter that much if you’ve got a strong process in place?
Or are there models people recommend specifically for more natural, human-sounding first drafts?

Would love to hear what’s working for others.

[languid compass]

Hey guys, I’ve been using Claude Sonnet for a while now, and it’s worked great for me. Recently, Claude banned the use of OpenClaw subscriptions, so I switched to GPT. Is there any way to make the model not talk so much? Can I make it behave at least somewhat like Sonnet?

[spark valley]

Creat a skill and in soul u need to embrace it and tell it exactly what it needs to hear that when it wakes up and finds nothing it should see this and if the user talked about something that seems u guys have talked about before u should use skill x for memory lookup and so on , i got everything setup but direction

[spark valley]

Just install a humanizer skill , assuming u wanted to humanize what the ai says

[spark valley]

Yes tell it to edit it identity

[frail notch]

Hi folks - has anyone gotten the browser to work? I've followed the instructions in the documentation (https://docs.openclaw.ai/tools/browser) exactly, including in .openclaw/openclaw.json:
{
plugins: {
entries: {
browser: {
enabled: true,
},
},
},
}

And yet, the agent insists that it does not have access to a browser tool or cannot use it from "this context"

@nimble hill is the "noSandbox" configuration critical...?
"browser": {
"noSandbox": true
}

[dense sleet]

I gave the agent access to external features

Like the playwright skill

That gave him the power to use the browser

Try that

[frail notch]

Gotcha - so are the docs just kinda wrong?

😅

[dense sleet]

Minimax is specifically human in its answers (that's what I've heard), give it a shot!

[dense sleet]

Not really, perhaps it's the model lacking enough context

But I'd highly advise using skills (such as the mentioned playwright)

[dense sleet]

You can always edit TOOLS.md in the agent workspace to give the model more precise instructions when it comes to available tools

[frail notch]

Got it, thanks a lot, I'll give that a shot!

[dense sleet]

You're welcome 😁

[frail notch]

How would you go about updating the tools.md? It's not quite obvious from the openclaw browser cli how the agent should interact with it (at least not for me) from the docs?

[dense sleet]

Tip: you can edit it in: Agents tab -> Files -> TOOLS.md -> Preview -> Edit

[dense sleet]

I'd advise you to explore the "root" files of the agent directly from the dashboard

Agents tab -> Files

There you'll be able to preview and edit files like: USER.md, IDENTITY.md, TOOLS.md, SOUL.md, etc.

[frail notch]

Ah no the editing is straightforward I just mean literally what to put in the file 😉

I.e. do we have any idea of what we should put in the TOOLS.md to make openclaw use the openclaw browser CLI

[dense sleet]

That's on you actually, whatever you put in there will be processed

TOOLS.md is more like a file to indicate the models how they should use tools

It's more like indications for the usage of the tool rather than letting it know the tool is there

The file itself should have a self-explanatory content indicating you what it's used for

[obtuse iron]

I can share an example of a section from my main agent's TOOLS.md

maybe it will be a helpful reference

[dense sleet]

That would be highly appreciated!

Nothing better than a live example

[obtuse iron]

Ah. Automod doesn't want me to post markdown in the openclaw server

This content is blocked by this server. From server moderators:
"Please don't send markdown headers within Discord, it creates a spammy chat"

Lemme see if I can just share relevant smaller pieces

Forgejo (Git)

Component	Details
URL	https://git.[REDACTED].com
Org	[REDACTED]
SSH	Port 2222
VPS IP	[REDACTED] (Hetzner CPX11, Ashburn)
API token	Stored in openclaw.json as FORGEJO_API_KEY

# SSH config (already in ~/.ssh/config)
Host git.[REDACTED].com
    Port 2222
    User git

# Clone a repo
git clone git@git.[REDACTED].com:[REDACTED]/openclaw-workspace.git

# Create feature branch
git checkout -b ceelo/my-feature
git push -u origin ceelo/my-feature

[dense sleet]

Just what I explained!

[obtuse iron]

there we go. So yeah the section about "Forgejo" (which is just a lightweight FOSS alternative to GitHub or GitLab) in TOOLS.md is just this

[dense sleet]

It consists of an explanation of how to use the tools rather than listing them!

[obtuse iron]

it's merely a guide to the agent on how to use the tools at their disposal, typically with some simple usage examples

[dense sleet]

How to do stuff would be more appropriate

[dense sleet]

Exactly

OpenClaw already injects the available tools in the prompt

[obtuse iron]

Yup yup, this is like Agent saying "Okay, I have the tool forgejo but what does it do, and how do I use it? I'll check TOOLS.md for details"

[dense sleet]

Exactly

Something that keeps annoying me is the timeout error, there's a hardcoded HTTP request timeout of 60 seconds that blocks me, the model I use is pretty large so it sometimes takes over 60 seconds 🙁

I would ask for help on that but well it's something related to the code of OpenClaw that is already reported

[rocky girder]

i've got 150$ of extra usage tokens, and extra usage enabled, but openclaw isnt pulling from that, it's pulling from direct api billing, am i doing something wrong?

[obtuse iron]

This block might also be helpful:

Web Content Extraction

When curl/wget returns an SPA shell (empty HTML, no content):

1. Use Playwright — Render the page, extract actual content
2. Don't give up or ask user to copy-paste — solve it procedurally

const { chromium } = require('playwright');
const browser = await chromium.launch();
const page = await browser.newPage();
await page.goto(url, { waitUntil: 'networkidle' });
await page.waitForTimeout(2000);
const content = await page.evaluate(() => document.body.innerText);

This is a core capability. Use it.

[obtuse iron]

which provider?

[rocky girder]

anthropic directly, i was one of the people hit by the ban from anthropic, but i redeemed that 150$ extra usage

[copper nova]

it's kind of a red haring, that's good for like 10 prompts

just spend your time moving off of anthropic instead of figuring out how to use those

better use of your time

IMO

I'm not an expert, also not your dad, you do what you want

[restive rain]

Just noticed now that everytime he gets “stuck”, I run a command in Telegram like “/status”, and then he replies

[rocky girder]

I'm using deepseek and mercury 2, I just wanted to know if still had any remaining usage I could use

[restive rain]

Feels like he gets stuck by some reason and don’t answer me when the task is done

I will try in TUI to see if that happens

[devout relic]

you called it. useless stats. haha. smh.

[copper nova]

I only speak truths

never lies

[frank agate]

anyone know what this is, Updating plugins... No plugin updates needed. Completion cache update failed ([openclaw] Failed to start CLI: Error: qa scenario pack not found: qa/scenarios/index.md at readQaScenarioPack (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:754:27) at readQaScenarioById (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:790:19) at readQaScenarioExecutionConfig (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:795:9) at readRequiredDiscoveryRefs (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2834:9) at file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/suite-BW4kSK9C.js:2840:39 at ModuleJob.run (node:internal/modules/esm/module_job:437:25) at async node:internal/modules/esm/loader:639:26 at async Object.register (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/register.subclis-D7AxmRi1.js:38:22) at async registerCommandGroupByName (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/register-command-groups-pnI4KWXM.js:91:2) at async Command.<anonymous> (file:///home/openclaw/.nvm/versions/node/v25.9.0/lib/node_modules/openclaw/dist/completion-cli-klMIRvmG.js:169:4)).

[tawdry mist]

I'm seeing that too when upgrading

[frank agate]

okie good to know

I wonder if they merged my ticket fix yet

https://github.com/openclaw/openclaw/issues/62879

[obtuse iron]

I got voice chat working but the latency is absolute dog dookie hahahaha...... wondering if anyone has tips on how to make it not suck

[devout relic]

im guessing openai just keeps us in the dark?

[proven drum]

model

[copper nova]

yep, don't need to know

[devout relic]

fair enough. just going to let it eat

[copper nova]

were my two favourite non-human avatars not able to figure it out?

[silent aurora]

did anyone was able succesfully use local ollama models for your OpenClaw platform ? so it really works - so far I watched many YT videos but everything is mostly BS without any proff it actually works relaibly - I have Openclaw on separate maching and use local ollama on Mac M2 Pro it works for few minutes then hung and Im exploding with the frustration

[junior phoenix]

You're not the only one... same for me

[silent aurora]

yeah it supposed to be great platform but reality is completly different .... and this is on OpenClaw not ollama models as these are responding in ms but OpenClaw doesn't see responses

[restive rain]

Was able to fix it lol created a new chat/session

Also changed the model and now everything is even smoother. Running GLM-5.1 and so far so good

It is a pity because feels like gemma4 might be a different beast compared with GLM

[astral vessel]

does openclaw browser start better than extendsion ? i want a full control of browser but sometime i failed to click some buttons on the screen? anyway i can check why and fix it

[mighty rampart]

M2 and how much RAM? Before I hooked up openclaw to ollama I used Open WebUI in front to make sure the model was behaving correctly. What are you doing with it? Watch Activity Monitor on the Mac. Is it using a lot of CPU and memory when it hangs?

[spark valley]

Dm i can help

[cosmic torrent]

I just installed OpenCLAW and connected my Gemini API key, but it started 6 minutes ago and has not responded. Is there anyone who can help me?

Even in the dashboard chat does not work like frozen

[silent aurora]

M2 Pro with 32 GB RAM - as I said I tink ollama models work just fine the problwm is OpenClaw (I want to say Clown) for some reason stopping receiving responses - work for max 2 minutes and then dead .... my frustration hit the max

[cosmic torrent]

Can you help me too?

[obtuse iron]

We got it working! Saki sent me a helpful datadump which allowed my agent to fix the E2EE DAVE issue :D

The latency is annoying but probably somewhat fixable with model optimizations. I'm doubtful that running either STT or TTS locally in the VPS will be much faster than the current API setup just because of the overhead involved

[copper nova]

sounds like logical thinking

[obtuse iron]

Might try STT in the VPS and TTS by API, or the other way around, to see if a mixed solution like that has better latency

[copper nova]

sorry I can't help much here, I don't like voice. I think talking to my robot feels a little too dystopian for my liking

[obtuse iron]

I won't be using voice often enough to justify a GPU VPS or scaling to higher RAM just for this... So I might be stuck with API

[obtuse iron]

Generally speaking this is also true for me, and I'm doubtful of the utility for my use case, especially since TTS is already built into most input devices

[arctic isle]

I had the same experience and eventually moved to haiku model/anthropic, would love to get ollama working

[copper nova]

the utility is definitely there, have low latency voice is gonna be the peak of a customer service type agent to require.

I would say it's almost the holy grail of what alot of people are trying to achieve right now lol

[nimble hill]

there is no quality control at openclaw - This 4.9 again broke the node feature. openclaw nodes list and openclaw nodes status do not again sync. The browser node capabilities which used to get added via plugin and browser.nodes section, is no longer working....stupid! will continue to work with 4.8 for now

[indigo goblet]

glad to hear

seriously tho, appreciate the work Peter and other contributors does but I believe it would be better if update were less frequent and more battlefield tested.

[silent aurora]

Yeah - haiku is fast and very good but still charging a lot

[swift cypress]

dm me thennnn

[spark valley]

There is also nvidia voice chat model

[raw pumice]

can anyone help me real quick im running gemma4 on a macbook m4 max

⠸ pondering… • 12m 3s | connected
agent main | session main (openclaw-tui) | ollama/gemma4:26b | tokens 32k/262k (12%)

it worked for like 5 seconds and then it stopped on a really quick question and then stopped but says pondering for 12 minutes already

[nimble hill]

Problem is not the release, but better quality control, when ther eis a breaking chnage, the docs should clearly indicate what chnaged and what replaced in config. Docs are often half baked, hence just wasting tokens just to find out what is the correct config

[indigo goblet]

sure is, but frequent update means sloppy unless you have serious man power and even then, things need the time it needs.

[nimble hill]

They all do via AI. the manpower is irrelevant here. The quality people to tweak test cases needs to be imprpved

[indigo goblet]

it is, good AI work needs multiple passes and testing, back and forth, iterative co design. So a single person can do a lot but its a lot of work. So man power do matter.

[raw pumice]

how do i see how often the heartbeat is? i cannot find it anywhere

[nimble hill]

Sure, but manpower amount not something that chnages between release. Hence that point about less manpower is irrelevant. The point is about quality control. How is the feedback loop involving human is working and whether those "people" are good enough - clearly not. Don't even get me started on management

No software i have seen breaks so many things and ding dongs between releases. The issues that have root cause found and fixed, must not be reopened again - this is te basics of modern SDLC

[nimble hill]

Default - 30 minutes. You can chnage it in openclaw.json. Mark it to 0 m if you want it disabled- which i highly recommend

[marsh bolt]

It errors quite a bit on older ollama versions. What are you running it on?

[weak vale]

has anyone figured out any better alternatives to using claude with openclaw? i keep hitting my extra usage and its too much

[marsh bolt]

I've had good luck with ollama. I run it on a separate machine from my claws. Where are you stuck?

[marsh bolt]

MiniMax is obvious choice. Not as good, but damn good for the money

[weak vale]

i'll check it out thanks mike

[clear tide]

A lot of users happy with gpt 5.4, heard good things about glm and minimax as well

[weak vale]

yeah i'd like to stay away from openai with my setup but i might switch over to minimax cause im already paying for claude max and now its useless to me

[marsh bolt]

Trick for this - ask your agent to set up his heartbeat to 30 mins and then search for heartbeat in openclaw.json - then set it to what you want (e.g 30m or 60m or 0). You and also set default to zero: agents.defaults.heartbeat

[glacial lantern]

what ollama models are people using here with good success? So far the only model ive had a modicum of success with is gpt oss 20b but in typical openai fashion it just flat out refuses to do most of what openclaw can do which makes it less than ideal to use and the others ive tried just are not good models to begin with

codex is working decent for me in the meantime but i just really dont like the anthropic api prices so I want to avoid using claude if I can lol

[clear tide]

For clarification are you referring to models that can run in ollama, or ollama models?

[glacial lantern]

Ollama models that I can use in openclaw

[clear tide]

The most recent generation of ollama models were a flop, not sure if there is going to be another future open source release or not.

They start ASI labs I believe and are now focused there.

[glacial lantern]

Ah that makes sense, I remember there being some great models for the time a year or two ago and now I’m just…..not at all impressed by any of them honestly

And it felt like there’s a lot more but nothing of value

[spark valley]

Glm 4.7 flash seems decent in openclaw

Using lmstudio or llama.cpp will be much more sufficient and 70% faster *if ur techy

[frigid plank]

anyone experience in openclaw control window your input not show up? or its super delayed? or behind in sequence, like an older input shows up in the window when you post something?

[clear tide]

openclaw update
openclaw dashboard

[ashen monolith]

has anyone gotten this chrome cdp websocket fixed?

[dense sleet]

I still haven't fixed the microphone drivers on my PC (Ubuntu), otherwise I'd be talking to my agent

[copper nova]

can't you talk to it from discord or telegram?

doesn't need to be from the VM itself right?

[dense sleet]

That's what I do, but being able to talk to it from across the room or something would be cool

[dense sleet]

Like saying out loud: "[agent], check my email, let me know if there's anything important!"

[earnest marten]

On my Mac mini (macOS, Node v22.22.0, OpenClaw installed via npm -g), the built‑in memory_search tool is failing consistently.

openclaw --version is currently 2026.4.9 (0512059), but I also saw the same behavior on 2026.4.10 from npm. Any memory_search call returns:

Cannot find module '/opt/homebrew/lib/node_modules/openclaw/dist/tools.runtime-hEfh9FVD.js' imported from /opt/homebrew/lib/node_modules/openclaw/dist/extensions/memory-core/index.js

QMD itself works fine via a wrapper script (scripts/memory_search_qmd.sh → qmd search/query), and openclaw doctor --non-interactive is green except for the expected “plugins.entries.active-memory: plugin not found” warning when I’m on 2026.4.9. Reinstalling 4.9 and 4.10 via npm install -g openclaw@... hasn’t changed the memory_search error. (I have downgraded to 4.9 after realizing that memory_search is broken in 4.10)

[dense sleet]

Have you tried not using homebrew?

I use NVM and it works fine with me if you wish to use a version manager for node

[dense sleet]

Also, if you execute which openclaw

What's the output?

[earnest marten]

I posted on help and krill sent some instructions on what to try to fix. The steps included a gateway stop, uninstall and reinstall. My agent tried to do those steps and of course went offline with the gateway stop command.

Good thing I had Hermes also installed and was able to continue the process and fix openclaw.

Thank you for your reply @dense sleet

[dense sleet]

Nice to know that you could fix it

And no problem!

[next walrus]

anyone else having their gateway shut down constantly since the anthropic changes?

[burnt raptor]

Bot is online, polling works fine for my account. A second user (business partner) gets zero response — complete silence, nothing in gateway logs at all when they message. Not even a rejected sender entry.

Tried: dmPolicy=open, dmPolicy=allowlist, their numeric ID in allowFrom, deleting the allowFrom credentials file, resetting update offset to 0, multiple gateway restarts. Their ID is confirmed correct. Tier 2 API. WSL2, OpenClaw v2026.4.5.

Messages show as delivered in Telegram but never appear in logs. getUpdates shows empty after gateway polls. Any ideas?

[raw pumice]

0.20.4

[spark valley]

You should install from the one liner rather than npm

[steep vale]

Do any of you use opencode go for openclaw?

[spark valley]

I ve been using hermes for awhile it seems very stable if u want fast onboarding

It has slef learning system

And auto create skills

[carmine tartan]

GUYSSS i setted up openclaw but is there a way that openclaw can control your mousse and keyboard fully ?

[patent cape]

is there anyway where i can force my general agent to spawn multiple spawn agents all on its own? what file do i need to edit for it

[candid urchin]

why cant my openclaw read my files

[mighty rampart]

Yes, computer use is possible. You’ll need to be using one of the models that can do it though.

[mighty rampart]

What happens when you tell it to spawn a subagent to [do some task] ? If that works you should be able to tell it to always do that.

[patent cape]

oh oki tyy, also can i use multiple api keys and ask openclaw to use different models for different works? like use qwen model for coding, gemma for main brain etc, also can i use the build.nvidia.com api keys?

[mighty rampart]

In general your agent can read files in its workspace. What are you trying to accomplish?

[candid urchin]

i want to create a backend to my frontend

[mighty rampart]

That sounds like different agents you will set up yourself to use different models. Think of it like onboarding an employee.

[patent cape]

okayy thanks a lot ❤️

[mighty rampart]

Is the code in GitHub? I would have the agent clone the frontend project into its workspace, and then set up the backend project similarly, and it can see both and do its work.

[candid urchin]

i already pasted the frontend files in its workspace directory

[patent cape]

hey also if i want openclaw to inject prompt into opencode window..how will it do it? with computer vision models?

[mighty rampart]

Okay then what is happening exactly? Check the openclaw json file and make sure its workspace is really where you think it is. @candid urchin

[mighty rampart]

I am sure that would be much easier using a CLI or API. Having it click around the screen ought to be a last resort if you need to get work done.

[patent cape]

okayy ty so much ❤️

[marsh bolt]

If you run Ollama serve in the terminal, you can watch for errors. Then copy and pasted errors to a smart agent and figure out what’s happening.

[spark valley]

Yup i managed to let it use the mouse

[spark valley]

Depends where they are and if the agent is sandboxed

[spark valley]

You just tell it or edit in soul add somewhere : when there is a work that if accomplished by subagents will be more clean and efficient use the spawn subagents tool to spawn subagents withiu the need of telling the owner.

[patent cape]

okii ty

[spark valley]

Api would be much preferred as cli if interactive it can be harder

[carmine tartan]

how to ? like i am using it on ubuntu and it just cant

[spark valley]

Well i tried in windows idk bout ubuntu ,ghere was a skill called windows control or someth it gave it mouse access

[candid urchin]

it is

[spark valley]

Check exec commands

[candid urchin]

how

[spark valley]

Tell it to : ls in it workspace and list all folder there

U will know if it hass exec allowance

[candid urchin]

ok gimme some time

[blissful berry]

ahh the cli vs api talk

I wish there was an easier way to inject data between agents

besides telling an agent to load this data

[mighty rampart]

depending on what kind of data... just have them write text files that the other one reads. or give them an issue tracker like Linear.

[wild finch]

I get this when I try to use Gpt5.1-mini "[agent] embedded run agent end: runId=2893745273495723457 isError=true model=gpt-5.1-codex-mini provider=openai-codex error=Unsupported value: 'low' is not supported with the 'gpt-5.1-codex-mini' model. Supported values are: 'medium'. rawError=Unsupported value: 'low' is not supported with the 'gpt-5.1-codex-mini' model. Supported values are: 'medium'." Anyone know how to fix it? My 5.4 Openclaw can't fix it it seems.

[candid urchin]

how to completely uninstall openclaw, if i installed it via ollama and on windows

[neat dirge]

I wrote a practical guide for using SayWell on OpenClaw’s existing OpenAI TTS path. Would love feedback if anyone tries it
https://saywell.io/guides/openclaw?utm_source=openclaw&utm_medium=community&utm_campaign=guide_launch

[solemn orbit]

Hey I am a researcher looking to understand the experiences of users and developers of ai agents, it would be of great help if anyone interested can dm me. It would take only 5 mins

[wanton sluice]

hello guys, can someone maybe be kind and see if this is happening to you as well https://discord.com/channels/1456350064065904867/1492564596962230432

[mellow sundial]

I’m trying to set the active memory plugin default model with config.model, but it’s saying it’s an invalid key

[mellow sundial]

Seems like you need to change your thinking

[sullen goblet]

Maybe a dumb question but how do you guys manage to run your agents all night on a codebase without it either spawning too many agents, or stopping early. I managed to do it but its always a pain to do and it keeps like timing out. Ideally I would want it to link find task by itself etc. Is there like a set of tools for this or is it only like a me skill issue ?

[stray pebble]

I'm on the onboarding process and trying to choose OpenAI Codex for the model but when I get sent to the browser it keeps saying authentication failed, missing authorization code. What does this mean and what should I do to fix it?

[flint flame]

Hi guys I'm lookig for reviews I'm just started with little models as Qwen3.5-4B for my Agent. But I'm lookig for recomendations is this a good start if I want to use my agent to aske code solutiones? I have limited resources i have rtx 5070ti, 32gb ram ddr4 and ryzen 7 5800x. My dubt is is this software necessary to run which models or what is my limit model I can run?

[mellow sundial]

Trying to login to codex is saying missing authorization code. Anyone else getting this?

[spark valley]

The best way is the openclaw killer github repo

[spark valley]

Heartbeat plus cron plus proactive skill

[white grove]

Hi All! I'm new, and I have openclaw connected to vllm running Mistral-Small-4-119B-2603-NVFP4, when I ask it anything related to tool calling such as "what is the weather in X" it responds with [TOOL_CALLS][TOOL_CALLS]weather{"query": "X"}