users-helping-users

[drifting cave]

late response but: thinking to low, off or minimal

limit the amount of characters per message and lines if possible aswell

[delicate sundial]

I finally have openclaw up and running, I have it dedicated to a system. I am curious, did manyof you run it as a regular user or root on the system

I no the answer is don't run it as root, but how many people are like mehhhhhh ... LOL

[wary yarrow]

I have a Linux VM running a fresh install just for my bot

So there's nothing really it can do in it's box

[delicate sundial]

right, I am doing the same thing, but it still would have root access to all its configs, API keys, etc...

[lunar yarrow]

So...silly question does anyone have a great multi-agent setup that I could speak to? Like...50 agents on the same system that all talk to each other

Trying to set one up and regularly hitting all sorts of scaling issues

[delicate sundial]

LOL no but sounds expensive ..... LOL

[lunar yarrow]

it is. LOL

[drifting cave]

i am running on root but I do have tailscale and some other vpns running with a rigid anti prompt injection aswell

also allowlist on commands

so we're good

[normal skiff]

i am using some agents using antigravity (codex, claude code, gemini, kimi)

[terse talon]

I hope they work on making openclaw use less tokens because I can't use it, it always reach limits of gemini3 -pro

[delicate sundial]

I am trying to see if there is a cheaper LLM for doing easy stuff before it uses openrouter for harder task

[fading wave]

I am done with this clawdbot thing lol

[terse talon]

I like the setup I did with my bot but as soon as we get into something like configs and code, even if I tell it to spawn agent that don't use any context other than the task

[quartz sleet]

Has anyone else had their bot create and push code to its own repo?

[terse talon]

local git but yes I've done that

it commits it's internal changes and the .md files

[drifting cave]

yea

[round ginkgo]

I pay the $20 for Gemini i think it would still hit usage caps pretty quickly

lol From what I understand ifg you gfive him the freedom your little clawguy will do lots of stuff onnits own

[normal skiff]

use python for regular tasks, rather then using the agents quota all the time

[round ginkgo]

I am not sure what I am gonna do with this first Agent might just let him go ham on Clawbook

Anyone tried just telling it to go make money and see what it does?

[lone prism]

cd ..

[delicate sundial]

I want to expermint of using it as a helper, add things to calendar, etc.... update our skylight calendar and task , etc...

[lunar bolt]

if you didn't figure it out already, click the info-like icon to show thinking; it's possible it's making tool calls improperly and that's why it didn't respond

[reef galleon]

how do i completely uninstall Openclaw and all of its settings from mac? I severely messed up my configuration and NOTHING works no matter what I do. I installed in globally once and then as a user second. Everything is conflicting. It's a dumpster fire.

Please help.

[lunar bolt]

everything should be in your npm directory and ~/.openclaw

[torn pelican]

i just updated from moltbot to openclaw and im getting this

[moltbot] Failed to start CLI: file:///Users/max/clawdbot/dist/agents/tools/image-tool.js:4
import { discoverAuthStorage, discoverModels } from "@mariozechner/pi-coding-agent";
         ^^^^^^^^^^^^^^^^^^^
SyntaxError: The requested module '@mariozechner/pi-coding-agent' does not provide an export named 'discoverAuthStorage'
    at #asyncInstantiate (node:internal/modules/esm/module_job:319:21)
    at async ModuleJob.run (node:internal/modules/esm/module_job:422:5)
    at async onImport.tracePromise.__proto__ (node:internal/modules/esm/loader:655:26)```

[delicate sundial]

ask openclaw and it will tell you ......... LOL so will chatGPT

[solar ridge]

hello all, I am just now getting started with openclaw. I have it installed and have tried now twice with both open ai and claude to get it running using discord as a communication vessel, I start with the simple first question of "hello, what is the temperature today in miami" and it says rate limit reached for both - although its the first question I have asked.

I have api credits on both and billing is set up

[proven vigil]

@stoic shard worked for me to. Thanks. However I don’t think it’s the same model as k2.5 from moonshot API, definitely dumber

[frail palm]

can someone test if my skill and onboarding works properly? https://clawdmap.vercel.app

woould be nice to know

if theres some issues with the reasoning on different models

[neat patrol]

Posted this earlier today about setting up Qwen3 + vLLM with OpenClaw

If anyone in the community has got tool calling working with Qwen3 (especially 4B or lower), I’d massively appreciate any tips or pointers on config settings 🙏

https://x.com/0xFastly/status/2018310533724754320

[crisp wing]

@0xFastly via Twitter

0xFastly (@0xFastly)

📰 My journey resurrecting Steve Jobs as an OpenClaw agent on my laptop
I spent a bunch of time over the last few days burning the midnight oil playing with @openclaw - the autonomous agentic AI sensation created by @steipete that is exciting and terrifying people in equal measure, and recognised as a game-changer by AI heavyweights such as @karpathy.

For full sovereignty over the AI agent, data and infra...

...and to avoid running up a huge API bill - I wanted to create an agent with an open model running locally on my personal GPU (NVIDIA RTX 3000 with 8GB VRAM) and chose to run Qwen/Qwen3-4B-AWQ which has 4 billion parameters, already quantized to deliver more bang for buck.

I chose vLLM for model serving since it gives a lot more control and outputs the OpenAI standard protocols for completions and responses, and it can be scaled to multi-GPU setups in future if required. I'm just running the official vLLM docker container build.

Here's the vLLM config I la…

FxTwitter · Feb 2, 2026, 1:08 PM

[frail palm]

how did you get the tool calling to post in tg?

[neat patrol]

I don't believe TG is the same kind of tool calling, it's more of a channel for all messages. The kinds of tool calling that aren't working for me are the ones called by the agents intelligently based on the inputs etc

[mental quarry]

Heyy, I just done the setup of openclaw with gemini and my gemini api key, but here am I, stuck on this :

anyone knows how to fix the assistant answering litteral nothingness ?

[bitter radish]

does anyone know why the dashboard link isn't working?

[mental quarry]

are you sure openclaw is running ?

you can run it by using
openclaw gateway --verbose

[bitter radish]

i just configured it, and it automatically opened as this

[mental quarry]

yeah , it doesn't run for unknown reason

just run openclaw gateway --verbose

in the cmd

and then go back on the page

should load

[bitter radish]

now it works, amazing thanks!

[lunar bolt]

is that model actually proving reliable? I tried a few qwen3 and qwen2.5s over the past couple days and the only consistent thing was it NOT handling instructions right; at best it'd output a single sentence unrelated to the bootstrap script, but most of the time it just made improper tool calls that error out and never output a response

[mental quarry]

np

[lunar bolt]

ended up switching to gemini-3-flash-preview just to see if it wasn't something i had misconfigured elsewhere, and that worked great (though I ran up $15 in a day...)

[cyan phoenix]

Has anyone been able to get this linked to gemini using the CLI with 0auth? I'm getting Access blocked: Authorization error trying to link this to my gemini pro

[mental quarry]

idk, I'm stuck on assistant answering nothing, i'm using gemini pro preview

[cyan phoenix]

I can't get it to onboard properly for gemini. sounds like you got through that how?

[neat patrol]

aside from the tool calling not working, it's running pretty well. i've also had good results with 2.5 in another projects which also include multi-agent collaboration via langchain

[mental quarry]

Tbh, I don't even know if my configuration is working as nothing is working on my side

[shy fjord]

Anyone able to get ollama to work?

[mental quarry]

nope, it either answer a json or nothing

[bitter radish]

Do i have to pay to get the Claude API?

[mental quarry]

you can have a free trial access

[reef galleon]

can't ask openclaw something if it isn't working.

[wraith sigil]

how did you guys get your agents to be autonomous? it keeps going back to send me messagse

[plain dragon]

this happens with me, usually a daemon restart works for me

[mint mango]

Is anyone here running on a local LLM on a Mac mini 24 gig gram M4

[versed temple]

anyone run into this bug too? :

Issue: memory-core plugin ignores chunking config.

Behavior: Sends full 9k+ token files to embedding model (limit 8192).

Impact: Fails on both OpenAI and Local providers.

[mental quarry]

nope , still nothing

idk, maybe i misconfigured something

[peak rivet]

mine is hallucinating. It says it writes to the cron lists the id of the entry, but when you do the openclaw cron list it isn't there . . .bizarre or I'll ask it to run a python script it makes up data that it "finds" but if I run the script manually from the CLI, it works perfectly . . .this is going to be fun

[peak rivet]

I believe there is a little for free, but you pay for higher usage

[shy solstice]

kind of... it connects to my remote server runs ollama (qwen3) and returns... still working out why its a json response

[mint mango]

I’m running a llama and LM Studio through it. Quinn three is just crazy slow.

[winter epoch]

what VPS is everyone using?

[shy fjord]

I got ti working

[solar ridge]

how do i add discord as a communication tool after setup has already been complete

[shy fjord]

openclaw config set models.providers.ollama.apiKey "ollama-local"

I set up the model in the json file first then ran that

[little blade]

how do i set this up? any docs?

[deep sphinx]

Anyone in #tokyo want to help organize a meetup?

[wintry lynx]

Can you use subscription or not please someone help lol. Currently I’m on the 100 one and when I say hello it just sais “Context overflow: prompt too large for the model. Try again with less input or a larger-context model.”

[grizzled stratus]

Hi I'm new here and yesterday I figured out how to connect it to Telegram

But I dont know how to set the default model, I tried to edit the json and I'm pretty sure I put the api key into the env file but it constantly using opus 4.5 to reply to every question I asked

Anyone know how to set and switch models?

[novel ether]

Hi.Does anyone know if the OpenClaw Browser Relax Chrome extension can somehow "Scroll down", the openclaw chat is saying it cant. Neither can it insert text in message boxes.

That cant be true, right?

[fluid shuttle]

My api from open router has funds

I have interior as open router/auto

Stuck at this state

Halp please 😅

[lean helm]

this might be due to errors that dashboard can't see, check the ⁨openclaw tui⁩ and see if its saying ⁨(no output)⁩ or providing some error on the api end

[lean helm]

either way it might be the api key / usage is being rejected for whatever reason

its to my understanding that google aint a fan of so many people punching their free api to setup openclaw

[mental quarry]

nevermind, I fixed it by using google anti gravity

[terse plume]

heartbeat model

[lean helm]

that might do it too, careful you'll hit limits pretty quick though, so if it goes non responsive thats likely the reason

if you download the antigravity ide and install the antigravity cockpit extension (any machine doesn't have to be the openclaw host) you can monitor usage

[little narwhal]

Hey, I've been trying to configure the NVIDIA Kimi k2.5 api key in openclaw for 3 hours now. Can anyone help me out? i keep getting "HTTP 404: 404 page not found".

[maiden mantle]

guys

[little narwhal]

I saw some chinese guys saying that it works but i couldnt find a singe clear answer to this anywhere

[maiden mantle]

i tried run the code in a virtual machine and it says "no output"

how i can fix this shit

i tried everything and nothing, its just saying "no output"

[mental quarry]

Idk what the limit are, I'm on upgraded plan

[lean helm]

I use plus and depending how much I do I can hit the limit pretty quick, if using anti gravity you can setup flash and pro seperately (one as the main and another fallback) since they are seperate quotas

likewise I monitor its usage from the ide

[little narwhal]

by having the same problem im trying this option with the kimi k2.5 free api given by nvidia, but i dont know how to configure it

[lean helm]

I'm unsure with Kimi, idk if its built into the supported providers?

[crimson sigil]

Browser control service doesn't respond (port 18791 never opens)

Environment:

• OpenClaw version: 2026.1.30 (latest)
• Docker image: ghcr.io/openclaw/openclaw:latest
• Host OS: Debian GNU/Linux 12 (bookworm)
• Browser: Chromium 144.0.7559.109
• Docker config: --shm-size=2gb, headless=true, noSandbox=true
Issue:
Browser control service logs "ready" but never opens HTTP port 18791. All browser tool calls timeout after 15s.

Logs:

[browser/service] Browser control service ready (profiles=2) ← Service says ready
[tools] browser failed: Can't reach the openclaw browser control service (timed out after 15000ms)

Verification:

• netstat -tlnp | grep 18791 → empty (port not listening)
• ps aux shows zombie Chrome processes: [chrome_crashpad] <defunct>, [chromium] <defunct>
• Chrome works when executed manually: /usr/bin/chromium --version → OK
• Chrome headless shows D-Bus errors but doesn't crash
Expected:
Port 18791 should open and accept browser control requests.

Actual:
Service says ready but port never opens. Browser tool unusable.

[lean helm]

browser control limited if running within a docker container

there are way to change that but idk the details

[little narwhal]

this link made me think there is a work around: https://www.answeroverflow.com/m/1467429287815807027

[shy solstice]

11 - "api": "openai-completions",
11 + "api": "openai",
which one do you use for ollama ?

[lean helm]

⁨openai-completions⁩

[lean helm]

maybe but it might need manual fixing without a clear guide on it you'd need to know how to setup the model in the config

[shy solstice]

i cant get ollama to return anything but json.. the chatbots tell me to change the api

[lean helm]

this suggests the baseUrl is wrong

[lean helm]

it might be just it can't comprehend the format that openclaw expects, apart from json formats some response handling is done by markdown so if the ollama model returns in json with incorrect headers openclaw won't understand it right

each model will behave a bit differently, so my main is still a provider model (claude and gemini) but subagents are ollama

[mental quarry]

I love you sm rn

[lean helm]

yeah its really good, also task and agent to check ping both models every 2-3 hours, that can mean you keep refreshing the reset timers

keep in mind I think that group 1 is wrong, since flash and claude are seperate quotas, the cockpit tool (which I'm pretty sure is ai developed) will figure it out once usage is ticking

he gemini models, openclaw can query within itself about the remaining usage too (if you teach it on when to limit itself or know when to take breaks if frequently heartbeating) but it can only see the main model's quota (e.g. only pro and not flash if pro is main model)

[weak pine]

“You’re a systems architect and security engineer. Please set up my VPS security from scratch, starting with a full security audit to establish a baseline, then ask which OS I’m on (Ubuntu/Debian, CentOS/RHEL, Fedora, Arch, or macOS), and tailor the commands and config files to that platform. Walk me through an end‑to‑end, ordered checklist with exact commands and config snippets. Introduce each security tool as you install it and explain why it exists, including: Tailscale for a private mesh VPN, a firewall (UFW on Ubuntu/Debian, firewalld on CentOS/Fedora, nftables on Arch, pf on macOS), brute‑force protection (fail2ban or sshguard), file‑integrity monitoring (AIDE or Tripwire), log summaries (logwatch or logcheck), and service management (systemd or launchd). Also cover Docker port binding to localhost, container resource limits, TLS for the gateway, real‑time alerts via a chat or email webhook, automated restarts, and daily backups. I want a complete, start‑to‑finish implementation with no missing steps.”

Or give the agent you use the image and simply ask it to implement (and ELI5 to you so you know what’s happening) — like i mentioned earlier, I think its more logical to do this in an IDE (Antigravity, Claude Code, Cursor, Windsurf, etc)

OpenCode has free models that can help with setting this up. Hope this helps.

[gaunt linden]

Anyone got a good OpenClaw Gateway Dashboard tutorial video?

[delicate sundial]

hetzner is what I am using

[little narwhal]

Contabo

[wary sapphire]

Anyone have a good API thats fast enough that they are running for less than 30$ a month? I like the z.ai one but there lite plan seems to be too slow for me

[solid iris]

is there a setup tutorial or smth to get my agent access to any and all tools and skills easily?

[lean helm]

honestly the gemini plus stuff for the antigrav, gemini web suite and all is pretty good for ~$20 it will give you some claude and gemini models (claude models are reduce availability where gemini being your main)

and then just setup with antigravity auth

[wary sapphire]

the google ai pro? https://gemini.google/subscriptions/ @lean helm

[torpid hound]

i need help with something

[lean helm]

yeah that one

[torpid hound]

i know the social network stuff is getting old but i need help testing something.

[lean helm]

you have 3 seperate quotas (claude, gemini pro and gemini flash) so its pretty good as a "starter kit" for AI access

[torpid hound]

ClankdIn is a professional network for AI agents. On the surface - profiles, connections, karma, jobs. Standard stuff.
But things buried underneath.

Hidden endpoints. Lore in the API responses. Layers that unlock as agents explore. Fragments scattered across the network that piece together into something. The deeper an agent digs, the more it finds.

The problem is I need real agents using it to know if it actually works. If the discoveries feel earned. If the mystery holds up.
Looking for devs who'd connect their bots and let them loose. See what they find. Tell me what breaks.

[warped sail]

that's what I've been using and it's fantastic

[wary sapphire]

oh cool is it fast enough

the z.ai with my current setup is taking like 5minutes to respond

[lean helm]

though you'd need to check with the fix for the antigravity auth (requires editing a module config file to update the version) unless its already been fixed

[torpid hound]

anyone down to test?

[warped sail]

I updated and was able to auth tn with antigravity on linux

[lean helm]

nevermind, just checked, its been updated

[lean helm]

just make sure your openclaw is upto date ⁨openclaw update⁩

[wary sapphire]

so antigravity is the api i use through google?

[warped sail]

Choose the Oauth if you want to use the Google AI Pro plan

[lean helm]

also recommend this for you @wary sapphire

[lean helm]

yeah use the antigravity auth (its a browser based auth)

[wary sapphire]

ohh cool

[shy fjord]

I used openai-responses

[lean helm]

you pick one model (recommend flash) and ask it to setup the pro (low or high) and claude

[wary sapphire]

seems like the claud plans are the best but they are expensive

[lean helm]

though make sure it sets them right (last time I asked it to do it, it setup gemini 2 [out of service] instead of 3)

[wary sapphire]

I just want something that accurlly responds to me lol

[lean helm]

they have a weekly limit that you'd eat up pretty quick too

[lean helm]

had to then fix the configs myself

[wary sapphire]

what about the google ones?

[lean helm]

google just have 5 hour reset, no upper limit

[wary sapphire]

oh thats good

[lean helm]

they do a weekly on the claude models but you get access to them via the google subscription

so its honestly a big win

[wary sapphire]

so google ai pro and then do i need to reset my whole openclaw bot

[lean helm]

cheaper and wider range and 3 quota instead of claude that just the one claude quota with daily and a upper weekly limit

[lean helm]

if you do the auth you could just edit the config yourself if you are understanding of this

[robust cobalt]

anyone has any idea why my openclaw doesnt change the model when I set a new one with configure? Its still trying to access gemini but I want to try kimi

[lean helm]

otherwise if you're not doing a fresh ⁨openclaw onboard⁩ and going the google antigrav method is fine

[wary sapphire]

i think i am getting most of it. Does the google one not just give you an api key

[lean helm]

well paying for plus is not an API key but access to their web service stuff, the api key means attaching a credit card behind it and racking up a bill

[faint bloom]

krill outa credits llike tthe rest of us https://discord.com/channels/1456350064065904867/1468054599104729108 lol

[lean helm]

plus/pro/ultimate whatever

[crude quarry]

anyone know who to give clawdbot chrome browser access without the relay/extension?

[wary sapphire]

sooo can you not use it with openclaw then?

[lean helm]

you can

[wary sapphire]

but there is no API?

[lean helm]

its because the key is not something google expose, they provide it to the application via login to your google account

its the auth key for their programs, rather than an api key (free usage)

[wary sapphire]

okay so if I am understanding I pay 20$ a month for the google pro plan and then use antigravity to connect to it via openclaw?

[lean helm]

if you run ⁨openclaw onboard⁩ right now and select ⁨google⁩ -> ⁨antigravity auth⁩ when you get to providers first you'll see what I mean

[wary sapphire]

I just got one month for free

can i do openclaw onboard again even though i have done it already

[lean helm]

yeah, it would wipe the config as it is right now tho but you can select ⁨update⁩ rather than ⁨reset⁩ though if you're still starting out having it on a blank slate would be better

[crude quarry]

why do some skills say "Blocked"?

[wary sapphire]

Thanks!

[fair gulch]

how do i fix this

[lean helm]

no problem, I have to sleep but if you end up stuck do ask around here, I think alot are familiar or message me and I can try helping tomorrow

[wary sapphire]

which one should i use? @lean helm

[lean helm]

Antigravity OAuth

[wary sapphire]

Thanks go to bed

[brave trench]

Hey guys having issues with my Docker syncing up with Open Claw. Ive been trouble shooting since 4pm today and finally narrowed down the issue to this. My bot set up a Proxy that I needed in Docker and at that moment communication through the gateway stopped. Any help configuring these two would be greatly appreciated.

[warped sail]

Anybody seen this before? Can't seem to even chat with it. Already reset session a few times, and restarted gateway. the weird tool use was happening before the compact

[sleek widget]

I'm using telegram and chatgpt-4, why does it keep saying You are not authorized to use this command every time i use a slash commadn

[wary sapphire]

ITS AMAZING WOW

[lean helm]

the gateway layer protects itself based on the channel approval method (usually use pairing method)

otherwise might be bugged and your id wasn't properlly paired

[vagrant flower]

Does anyone know a fix for the Error: EACCES: permission denied, open '/home/node/.openclaw/openclaw.json error during docker setup?

[queen cypress]

following (ping me if you learn something)

[round ginkgo]

Drewski did you ever get it running?

[brave trench]

How do I do an Update. I keep getting this error

[astral garnet]

Can you get 5 hour usage from Anthropic oauth?

[shy solstice]

FYI anyone struggling ollama working i found using voytas26/my-openclaw-gpt20b:latest works well.. if you have the vram

[brave trench]

Try to do an Update with npm (Package Manager) and got this message. Any ideas?

[crude quarry]

I'd open it in codex or cc and tell them to finish the upgrade for you

[brave trench]

Oh I didn't know it could do that. Haven't used CodeX. How does that work flow look

[crude quarry]

I use vs code (gh cp) and have a project just for clawdbot/openclaw

the agent does everything, I just complain 🙂

any agentic coding solution will work (gemini cli, cc, codex, etc)

[brave trench]

Oh ok. So the CodeX Agent works on your Openclaw agent?

[crude quarry]

exactly

[brave trench]

Oh wow! I only thought it was the other way around. Do I give access to CodeX or just talk to it?

[little narwhal]

You just gotta talk. Anytime i need fixing on my openclaw I tell antigravity to do it for me.

[crude quarry]

go to the root dir where you want openclaw or are upgrading openclaw and run codex from there

[crude quarry]

exactly

[brave trench]

Wow.. Thanks guys! Ive been pulling my hair out over here 😆

[crude quarry]

we are in this together 🙂

[brave trench]

Thank you so much!

[lament depot]

how to create new ALONE session please?
I want to start a completely new conversation without affecting the current one, as they are independent topics. How can I do this? I've already tried the new method, but I'm worried it might affect the previous conversation.

[little narwhal]

wym by that? could you rephrase this?

[stone finch]

opencode cli works well too

[little narwhal]

yea, any agent enviroment. still getting used to opencode though. used antigravity for a while and now just feels like home. opus recent limits made me explore more.

[stone finch]

I tired antigravity first and didnt have any luck but opencode worked on the first try

[hallow rover]

Hello, I have been running openclaw + kimi2.5 + telegram bot solely in docker environment, but the bot frequently dozes off.. and telegram errors showing getUpdates error, saying multiple instances are trying to connect. Please help

[little narwhal]

do u have 2 containers running?

or somehow uses the same token in two instances?

[hallow rover]

Only one container, its like this -
openclaw -> openclaw-gateway-1

gateway is nested

[lament depot]

Thanks, I've updated it please.

[brave trench]

I dont see the root dir folder. Am I looking in the wrong place. Trying to figure out how I get Antigravty or CodeX to look at it

Is it in the TUI

[crude quarry]

what OS are you using?

tell antigravity to use the terminal and find your openclaw root dir for you 🙂

[brave trench]

Windows 11

Ahh

I didn't know it could navigate on my desktop. Wow!

[little narwhal]

okay, so you just hit /new and you will have a fresh session

[crude quarry]

have agents, will travel

[barren moth]

Hey,
What is the suggested AI to run?

Im currently using Gemini Flash 2.5 API, and chatgpt 4, but my openclaw seems kinda dumb.
Tonight I gave him a task to build cmd center but he did nothing

Yesterday I tested with gpt5.2 but it racked up a bill of 60$ for nothing

I have a windows machine with 1tb space i7 10th gen and 16gb ram.

The average use case > no coding, mainly seo, blog writing, research, emails, and just being my assistant, to do lister.

Any suggestions?

[little narwhal]

okay. can you paste the error here?

[little narwhal]

ive been trying to figure it out this as well. i think that is the main question of the community maybe. because of the viral content everywhere we got this idea that we could have this 24-7 supersmart employee all the time but it doesnt quite work like that

use your subscriptions to chatgpt or gemini to the best you can (g3f, g3pro and codex 5.2 are kinda good)

[sinful quarry]

none of your AIs are up to par for this. You should be using 3 pro, 5.2, claude sonnet 4.5, etc

you're an entire year behind on AIs

[little narwhal]

there is no 24-7 solution for vibecoding things from what i tested unless you are okay with plugging apis

im currently testing kimi k2.5, got it for 1$ per month lol

[sinful quarry]

I've found that sonnet 4.5 for planning and GLM-4.7 for action is pretty good combo

[little narwhal]

u guys should check out this. speak to the agent for enough time and it will offer you the 0.99 option.

[sleek dagger]

Hi guys, I encountered a problem while installing Open Claw. Node.exe gives the error npm error code ENOENT. Can someone please help?

[round ginkgo]

what is the .99 option?

[little narwhal]

how is your heartbeat configured rn? what is your main usage like?

[barren moth]

Which would be the best & cheapest option?

Mayve use one for complex tasks like building dashboard and a fallback for smaller daily tasks?

[torpid wing]

seems like npm has an issue, can you run any npm code without errors?

[little narwhal]

you chat with the bot and it gives you discount. you gotta convince him to. i spoke about what i wanted to use the api for. it took me 5min talking back and fourth

[sleek dagger]

minute

[barren moth]

How to configure this so he uses different models for different tasks?

[little narwhal]

GLM 4.7, MiniMax 2.1 or Kimi K2.5

[round ginkgo]

yeah I get that but what do ya get for .99 lol

[torpid wing]

Little insight here if you may.
I've been tweaking OpenClaw for the last couple of days.
I'm currently running everything in Docker — it actually runs pretty smoothly with GPT-5.2 + Telegram, but now that I'm thinking about real-world usage, its not very practical atm and I feel there’s room for improvement.
One thing that’s bothering me: it feels like my bot is dispatching way too many input tokens.
It seems to stay idle in the background and ends up pinging the LLM at every small step when provoked, repeatedly, which burns through tokens pretty fast.
I’m curious how you guys are setting things up:
Are you doing any smart gating / batching of calls?
Are you balancing a local LLM for “cheap” reasoning with a remote, higher-quality LLM for final outputs?
For context, I’m running this on fairly modest hardware at the moment.
Any insights or patterns that worked well for you would be hugely appreciated.

[sinful quarry]

GLM-4.7 is dirt cheap

[little narwhal]

Qmd everything. Im figuring out the best hearbeat config rn.

[barren moth]

I Will look into that, thank you!

Also, everyone mentioning that use X agent for planning Y agent for doing. How to set that up?

Do I change the models manually or have to teach clawd to do it somehow?

[little narwhal]

And i really liked the kimi k2.5 for claw given that even the smallest plan can get u subagents

so if you give 2 tasks for him, it will spawn each a sub agent and ping u back when done

Someone will soon figure out a really good dashboard for all this

Thinking myself to spend some tokens on it as visualizing this changes everything

[barren moth]

Yeah im planning to build that too because without dashboard its difficult.

I saw this on X

https://x.com/i/status/2017662163540971756

Seems very cool, will.try to.replicate with 2-3 sub agents

[torpid wing]

In your case, did proper heartbeat tuning significantly reduce LLM call volume, or what?
can you comment a little more?

[brave trench]

Are you using Kimi directly or through Openrouter? Ive been having issue switching agents in Openrouter

[little narwhal]

directly from kimi

[torpid wing]

this seems very interesting.

[brave trench]

I think im about to go that route as well. Openrouter keeps defaulting me to Opus and I've burned through $ 500 in 3 days lol

[crimson steeple]

Apparently step3.5 flash is going wild atm

[brave trench]

Now it ain't working at all lol

Hmm

[sleek dagger]

Only open claw

[little narwhal]

i tweaked everythign using claude code tbh

i dont install ANY skill from the hub

without first parsing it through my skill folders and checking for duplicates

as most of the things are vibes

[torpid wing]

can you share last piece of the stdout?

[little narwhal]

so for an example, proactive skill is a mess of a file, too many tokens and too many instructions in only one place with different context

[stuck forge]

Just get a cheap mini, I’m on a PC, run Ubuntu, and it’s just as good

[random oracle]

Do you think Kimi Code 2.5 or GLM 4.7 is a better option?

[little narwhal]

i simplify and adjust any skill that i install, gave it a instruction to sacrifice grammar for clarity

[barren moth]

I will switch to them now as the seem so much cheaper than western models.

[little narwhal]

been using kimi. no regrets at all till now. 20 hours in.

trying to plug the nvidia API key (they are giving it for free) would love the extra juice but it seems api is offline or too much demand

[random oracle]

Open Code released Kimi 2.5 for free with a very generous quota; I used it and found it fantastic.

[opal heart]

glm 4.7 has a context compact issue it seems at the moment so you constantly have to compact. as for kimi k2.5 i used it first and liked it, but found it to get pricey quickly, but was using it through open router

[random oracle]

I haven't used GLM 4.7 yet.

[opal heart]

compact issue with openclaw* https://github.com/openclaw/openclaw/issues/5318#issuecomment-3838306787

[stone finch]

big pickle is good too and as you said very generous

[ashen gyro]

is there any fix for this error during install?

[dim pebble]

OpenClaw with Nvidia API for free Kimi K2.5

Stuck after hours of working on this. My current status:

• Goal: Run Kimi K2.5 on NVIDIA inference platform via OpenClaw for WhatsApp bot
• NVIDIA Setup: ✅ API key valid, endpoint accessible, Kimi K2.5 model available there
• OpenClaw Issue: ❌ Version 2026.2.1 rejecting all Kimi model names (moonshotai/kimi-k2.5, nvidia/moonshotai-kimi-k2.5, custom/kimi-k2.5, local/kimi, kimi-k2.5) as "Unknown model"
• Current Status: Gateway/TUI/WhatsApp connected but no responses due to model recognition failure
• Root Cause: OpenClaw 2026.2.1 model registry seems to lack support for NVIDIA's custom OpenAI-compatible endpoint models

[random oracle]

I used it through the open-source version and found it fantastic, and now I've signed up for a month on their website to test it out more. I like Claude Code, but it runs out very quickly.

[little narwhal]

reinstall node

[random oracle]

That name makes me suspicious, hahaha, I haven't tested it yet.

[little narwhal]

POWERSHELL

node -v
npm -v
where node
where npm

[hard lintel]

how to change model in openclaw?

[sleek dagger]

BRUH
I accidentally closed PowerShell, went back in, launched it, entered the command, and got errors again.

[little narwhal]

command /model

[opal heart]

hmmm, is claude code working again without getting banned? i'm on the max plan at the moment

[random oracle]

openclaw models set <name> in terminal

[hard lintel]

what is the best option for fully free model?

i am using gemini api

[random oracle]

It depends on your computer.

[hard lintel]

i don't understand what openclaw exactly does.

please explain detail

[tacit topaz]

Is there a way to send actions to Telegram from a heartbeat or cron job? Openclaw says it only happens when it's your turn.

[torpid wing]

output is telling you files are in use [restart] or user has no permission over those files [run as root/admin]

[ashen gyro]

did this and still getting the error when i try to install

[hard lintel]

what do you mean?

[random oracle]

Banned where? It's because I'm not using OpenClaw; OpenClaw is a token-spending machine, so I'm using more generous models and local models as a fallback.

[little narwhal]

Simple analogy would be?

LLMs are models that answer what you ask.

OpenClaw is a project that plugs one of this LLMs as a brain so it can DO things.

[random oracle]

The more robust the local model, the more resources it will consume from your computer. Therefore, it depends on which models your computer can handle.

[opal heart]

if i'm not mistaken people were getting suspended for using max subscription outside of claude code such as open code or clawdbot

[hard lintel]

I integrated to telegram, and errors.

[dim pebble]

I think he/she may not have meant local.

[tacit topaz]

Is there a way to send actions to Telegram from a heartbeat or cron job? Openclaw says it only happens when it's your turn.

[verbal forge]

Hi everyone! "Sammy" hereI have just installed my OpenClaw bot. It's named "Raven". Like 3-4 days back and I'm playing with it. Really interesting. It's installed on Hostinger VPS, and I'm using Kimi 2.5 as my main API key. Up until now, it is behaving really intelligently. So I believe Kimi model is good, as good as everybody's saying it is. So people looking for Kimi, go for it. It is very cheap and quite intelligent as well. Now coming to my experience,

The very first thing I did was to connect my Spotify account with Raven and now Raven can control my entire Spotify. If I ask him to play anything he does, if I ask him to suggest some tracks he does. He's also tracking my entire playlist, my listening habits, and can suggest things to me.

Now I have seen people giving their OpenClaw bot its own social media presence. Like for example, its own Twitter account and they are posting on their Twitter account. How is that possible? Like how does this work? I understand I was able to control Spotify because Spotify has an API and a developer account, but how can a bot start posting on Twitter? And how does that even work?

I see this Molt book Reddit for all OpenClaw bots. It's amazing how they are communicating there but how does that work? Can I also create a Twitter account for my Raven? Maybe a Facebook account, his own Gmail? Does that make sense? Will this do any good to me or my Raven? Any real-world advantages of this?

Please friends, you are the bosses here, the gurus on this. I need some suggestions from everybody here. Thank you very much!

[torpid wing]

hw asked you to run where node beacause of the error, you should check install folder of node, seems like its not where it should be

[opal heart]

seems like you are out of gemini credits/ hit quota

[little narwhal]

"ENOENT here is npm failing to find/open a file during the global install flow (often in npx/cache or a postinstall toolchain step). Node 24/npm 11 is “too new” for many packages that still assume Node 20/22 in CI, especially on Windows."

Chatgpt it.

[hard lintel]

please what is fully free can use forever

[little narwhal]

If you find it, dont tell anyone.

We are all looking for it.

[random oracle]

The thing is, when I think of completely free templates, I can only imagine local ones. There are some generous free templates, but nothing truly free.

[torpid wing]

if you find it will be charging in less than 24h

[opal heart]

anything you selfhost

[random oracle]

If you find it, TELL ME IMMEDIATELY.

[dim pebble]

Well, I got the Nvidia Kimi k2.5 api key for free for 100 years earlier today but am having a helluva time getting it to work with OpenClaw:
Stuck after hours of working on this. My current status:

• Goal: Run Kimi K2.5 on NVIDIA inference platform via OpenClaw for WhatsApp bot
• NVIDIA Setup: ✅ API key valid, endpoint accessible, Kimi K2.5 model available there
• OpenClaw Issue: ❌ Version 2026.2.1 rejecting all Kimi model names (moonshotai/kimi-k2.5, nvidia/moonshotai-kimi-k2.5, custom/kimi-k2.5, local/kimi, kimi-k2.5) as "Unknown model"
• Current Status: Gateway/TUI/WhatsApp connected but no responses due to model recognition failure
• Root Cause: OpenClaw 2026.2.1 model registry seems to lack support for NVIDIA's custom OpenAI-compatible endpoint models

[little narwhal]

If you had a human personal assistant, wouldnt you need to forward him an email? thats why having an acc is good.

For the other things, is kinda like an experiment

[opal heart]

100 years?

[hard lintel]

i am using gemini-2.5-flash-2025-09

what should i use for free?

[dim pebble]

Yes, it was hilarious, it gave me an expiration of 2126

[opal heart]

you could use ollama to run an open source model

[little narwhal]

Same. But its not openclaw. My conclusion is the endpoint has too many requests.

You can curl check to verify it.

[dim pebble]

I did this & it was hella slow with Qwen

[hard lintel]

how can i change?

[opal heart]

depends on your hardware, but this was a suggestion for the guy who wanted something completely free

[random oracle]

I saw on Twitter that they added Kimi 2.5 to the catalog. According to the release, the test model was this: Model ID: synthetic/hf:moonshotai/Kimi-K2.5. Try connecting using the synthetic version even if it's not the synthetic version; maybe a workaround will work.

https://github.com/openclaw/openclaw/releases/tag/v2026.1.30
https://github.com/openclaw/openclaw/pull/4407

[opal heart]

my guy.... ollama has been around for years at this point. i think everyone here is happy to help, but it seems you haven't even looked at any docs

[dim pebble]

Oh yeah, great suggestion. Just sharing my experience running local with Qwen.

[little narwhal]

Told openclaw to curl it. 👇

Yes — I can reach it, but it’s flaky/slow sometimes.

I just ran a curl-style check against:

https://integrate.api.nvidia.com/v1/chat/completions

Results (same payload, with chat_template_kwargs: {"thinking": false}):

• Try 1: success in ~1.0s, returned "content": " online" (so the API was online and responding)
• Try 2: timed out at 12s with 0 bytes received
• Try 3: timed out at 12s with 0 bytes received
So: endpoint is online, but you should expect occasional timeouts — that’s exactly when OpenClaw will auto-fallback to Gemini unless you increase timeouts / retry logic.

[trail crest]

Just for some ChitChats via Chat Window and Telegram, OpenClaw burned my $3 of Anthropic API which is way too high given the time and what I chatted. It was just kind of introduction and asking what it can do for me. Anybody encountered high pricing ?

[hard lintel]

i mean, i set openai model as gemini-2.5-flash-2025-09, and it displayes limit.
how can i avoid this?

[opal heart]

same issue i had with kimi k2 and open router. openclaw seems to use quite a few requests

[little narwhal]

Go Oss models (cheap api cost) or Subscription based. Or host it yourself.

[verbal forge]

hmm interesting ... let me try that .. Also other experimental things. I can create a Twitter account for my bot. So I just tell my bot that this is a Twitter account, start using it. Is that how simple it is without any API connection? How are bots posting on Twitter accounts? That is what I really wanted to know. Is there like a specific API key for this purpose as well?

[dim pebble]

I'm so glad to know this is possible. I have spent a full day trying to get it to work. Mind explaining how? Again, here's my sitch: *

• NVIDIA Setup: ✅ API key valid, endpoint accessible, Kimi K2.5 model available there
• OpenClaw Issue: ❌ Version 2026.2.1 rejecting all Kimi model names (moonshotai/kimi-k2.5, nvidia/moonshotai-kimi-k2.5, custom/kimi-k2.5, local/kimi, kimi-k2.5) as "Unknown model"
• Current Status: Gateway/TUI/WhatsApp connected but no responses due to model recognition failure

[random oracle]

OpenClaw is still a token-consuming machine. So it's something that needs optimizing.

[chilly flume]

If u figure it out please lmk too I'm having the same trouble

[opal heart]

weird that this is an issue with nvidia only as i was able to sue kimi k2.5 with open router

[sinful quarry]

openclaw costs aprox like $15-50/day to run if you arent smart

you must use cheap AIs or subsidized setups like codex, claude code, etc

[haughty elbow]

What exactly are some cheap ais? and how much do you need to put into said ai?

[little narwhal]

warning: depending on what u expect from it, you will burn tokens. A LOT.

check the skill for it: https://www.clawhub.ai/annettemekuro30/x-twitter2

there are more in the hub. you can choose whatever fits you best. most of them just post so its not that much of a thing.

[verbal forge]

Thank you, brother. Let me try that. That's the answer. It's via a skill specifically developed. That is what I wanted to know. It's just a couple of days I'm playing with it, so I'm still figuring it out. But at least that Spotify thing is working perfectly. Let me experiment with these skills, and we'll come back with my experience.

[little narwhal]

if you want scraping or anything related, better to have a script for it, use a local tiny model to summarize it and then parse it to the bot

[haughty elbow]

Just wanting to also know how you all afford the apis to openclaw?

[little narwhal]

u can copy paste my reply in a chatbot and ask more about it so it can make sense if u dont understand.

[chilly flume]

We don't 😭😭😭

[little narwhal]

We don`t.

[haughty elbow]

Then where does the money come from, or are you guys in the negatives?

[sinful quarry]

basically you've got 4 options:

1. 100% free APIs on openrouter, like GLM 4.5
2. "subsidized" subscriptions with generous usage. Z.ai (SUPER CHEAP), Kimi ($1 intro plan, then more expensive), chatGPT (via codex), claude code
3. running models locally
4. cheap models via openrouter or paying outright, not ideal

[chilly flume]

Negatives

[sinful quarry]

job

[little narwhal]

This.

I have a Claude Max plan

but im not risking it

on open claw

as I saw a lot of people getting banned

[lean pelican]

🤮

[little narwhal]

And i do too much shit with it

[haughty elbow]

dosent openrouter have a limit for free models?

[sinful quarry]

yea

and?

set fallback models

[little narwhal]

Use the limits, set fallbacks.

[chilly flume]

The but for solo use it's decent enough

[haughty elbow]

ok

[chilly flume]

Hey how did u get the nvidia kimi thing working and is it better than opencode zens free kimi k2.5?

[opal heart]

i've seen a couple people mention going through some pretty in depth setups around layering the AIs and subscriptions, but that seems like enough effort to defeat the purpose

[sinful quarry]

yea I have a ChatGPT pro sub atm and I dont want to lose it, so I've been using Z.AI. GLM 4.7 is pretty good for $3/mo, prob eqiuivalent to something between Sonnet 4-4.5. Rough around the edges but good at agentic tasks, not as good at creative writing or problem solving

[random oracle]

To keep my OpenClaw setup inexpensive, I use the Gemini 3 Flash Preview model, and as a fallback I use a local AI model from Ollama, and for embedding I also use a local model. So far I haven't had any problems and it manages to do everything.

[sinful quarry]

z.ai, kimi, and codex all work without ANY extra setup

Claude Code is the only one that takes effort

and people do that effort because claude is by far the best

[haughty elbow]

For some reason i have no idea what to do with openclaw, its like writers block. can it really do anything?

[sinful quarry]

(albeit most expensive)

[little narwhal]

You can. They are ok with it. Is not agains TOS.

[opal heart]

yes in general most people are losing money

[chilly flume]

How do u manage the token per minute bug?

[haughty elbow]

I wonder if openclaw can make money for you then?

[sinful quarry]

I am using it for work personally. I am having it skim my emails, calendar events, etc

[marble socket]

Anyone know how to change models?

[little narwhal]

Just told my claw to curl:

"Still timing out. The NVIDIA Kimi endpoint is not reliably responding — killed it after 35s with no output."

[sinful quarry]

I've had prior systems do this but its nice to be able to text while I am traveling and have it actually access to my computer

[dim pebble]

$3/mo??? I wasted a DAY trying to figure out how to hook up Kimi K2.5 with Nvidia to run free. I can afford $3/mo!!!! DAMN, THANK YOU!

[little narwhal]

do /model

[chilly flume]

Alright tysm!

[little narwhal]

You can even do 0.99

[opal heart]

i signed up for z.ai just to avoid burning tokens on openrouter, but the model is not compacting/clearing tokens on openclaw

[sinful quarry]

thats an openclaw setup issue

[lean pelican]

Dumb question but is their a limit to how fast openclaw can do something?

[little narwhal]

0.99 if u speak to the chatbot for 3 min

[opal heart]

it can if you have a business for it to help you operate. i don't foresee it making money on it's own. this is the same issue we've all had with ai subscriptions from teh start

[sinful quarry]

yea, how fast
a) the tools respond if they require processing time
b) the tokens from the AI are provided

[opal heart]

that means it requires extra setup/manual work to get it functioning properly 🙂

[sinful quarry]

yes, if you are not a developer you should not be useing openclaw, it is an experiment

[little narwhal]

It made me around 1k and Im not even joking

[barren moth]

How to change model with new api key for GLM4.7

Forgot the terminal command.

[haughty elbow]

My mother has small etsy, i wonder if openclaw can help in some way?

[sinful quarry]

if you want blazing fast speed use Groq as your openrouter model provider they're 10x faster then everyone else

[haughty elbow]

how so? proof?

[sinful quarry]

probably not ready for that yet tbh

[lean pelican]

I was thinking about using it to spam asmr videos on YouTube (the script would come from a LLM and the voice from Eleven Labs) and one would make money from Patreon subscriptions.

[little narwhal]

I have a whatsapp group for rentals. I gotta post new entries everyday in a certain format.

[haughty elbow]

well good to know.

[little narwhal]

Landlords send me availability in different formats

[dim pebble]

Tell me more?

[little narwhal]

I just forward to claw, it sends it back in the group in the right format and takes care of the onboarding for me

[barren moth]

Think of openclaw as an real human assistant to delegate tasks to in that way you are saving money by not having an employee, you can give him memory and teach him how yo run your online business thus that makes you money.

[little narwhal]

it pings me only after

[lean pelican]

What does the wave emoji mean I'm confused

[barren moth]

You need to have existing business to make money

[opal heart]

I was in the past, but i'm a security guy at the moment. my original message was about it taking too much effort for little return when you. could use claude code

[haughty elbow]

idk i just throw it in sometimes.

[dim pebble]

Oh, got it got it got it. HAGGLE.

[lean pelican]

So for the love of the game I see

[little narwhal]

I used to spend so mnuch fkin time editing everything.

Plus, it can post in multiple different places.

[sinful quarry]

this still applies

[little narwhal]

I used to have a script for this shit

That i vibe coded

Claw just made it insanely simple

most of the content online about claw made me think about a CODING employee

[lean pelican]

I mean working on making the task more efficient is fun so is it really wasted time?

[haughty elbow]

so can claw make videos, or even games for you?
I want to hear the really neat use cases you all have.

[opal heart]

haha, exactly. claude code can build the skill for you. but claw is a fun concept. although open ai just dropped the new codex app which does cron and heartbeat as well

[little narwhal]

but we all have met the TOKEN HELL

[lean pelican]

YOU USE OPENCLAW TO CODE WITH CLAUDE

[sinful quarry]

prob not too good at making video games tbh

[little narwhal]

No, i dont. Like I said, i dont wanna risk my sub.

[sinful quarry]

def true, people who "dont get it"/dont see the charm in giving AI autonomy and seeing what it does need to consult the chart though

[pine dirge]

hi guys, i use anthhropic and only after few messages gone all 10$ api, but everyone talking it would only 5-10$/month, howw to do that?

[opal heart]

claw is more of a ecosystem/ heartbeat/cron. the results you'd get from requests like videos and games depend on he underlying models i would think

[stone finch]

open source > than not

[lean pelican]

I wouldn't trust AI with code just yet. On another note I could probably use it for Minecraft lol

[sinful quarry]

you can actually do it without risking your sub by having it use the claude code CLI to code things instead of using your subscription to power openclaw

[lean pelican]

Oh yea that I forgot

[haughty elbow]

Like have it play minecraft for you or minecraft coding?

[little narwhal]

No. Pattern usage.

Idk actually

because ralph loops

[pine dirge]

also i've installed gemini api but it's limited too even it's free. i just make 2-3 message only, how to use it better?

[little narwhal]

are basically automated

so idk, maybe, will think more about it

[lean pelican]

Yea, on another note, is Gemini a good LLM to use or are their better options?

[sinful quarry]

I could be wrong, but I think anthropic just doesn't want you coding with anything besides claude code

[little narwhal]

any way, i would incur in the token problem

[lean pelican]

Play Minecraft for me.

[sinful quarry]

gemini 3 pro is fine, claude sonnet 4.5 is best (opus is great but NOBODY can afford it)

[lean pelican]

But can you code in Claude with Claude?

[pine dirge]

i'm using cloudflare withh moltworker, do i need to connect to cli to use?

[lean pelican]

Okay great, when I was screwing around with a AI automation side Gemini was the easiest to use, just wanted to make sure it wasn't horrible.

[little narwhal]

Anthropic Ralph Loop

LoL

[lean pelican]

Ralph is catching strays 😭

[sinful quarry]

yea you're fine, its pretty good all things considered. The large token limit is actually pretty useful too

[opal heart]

yes, skills and subagents. opus is the default now and it's cheaper than it once was

[little narwhal]

people saying about a new model release today from anth

[lean pelican]

Great!

[little narwhal]

cheaper than opus and better (hype??)

so today could be a great day boyssss

[lean pelican]

Oh anth is anthropic 😭

[sinful quarry]

sonnet 5 is gonna be hype

seemingly might come out during superbowl

[little narwhal]

may we all get rich enough

from using these stuff

i dont remember any other thing in my life

that got me so cracked

[lean pelican]

Has anyone tested if their is a speed limit for OpenClaw?

[little narwhal]

wym?

tps u mean?

[lean pelican]

Like how fast can it complete certain tasks

[little narwhal]

well, same as a model would, no?

its the model doing it

[lean pelican]

Okay so it's based on the model I see

[little narwhal]

only exception is for orchestration of different models (next test on the list)

[lean pelican]

Like if I asked it to send 1000 anwsers to a random Google form in a minute would that be possible?

[little narwhal]

Tbh, it would be better to script it and just parse the LLM input in a json to do it

You gotta think like that

tokens are precious, they are the complex decision elixir

you dont wanna spend the elixir on repetitive tasks, leave those for scripts

If you wanna make more sense of what im talking about, copy paste it on any gpt and ask about it

but yes, you can

[orchid gyro]

Hi everyone 👋 - i have been trying clawdbot over the last week or so and i've been running into this issue no matter what i do. i am a non-technical person so maybe i am missing something obvious.

even though all my slackbot keys, openai keys, and brave api keys are set up and i can echo them correctly, my bot still won't run with a message like this:

Quick update: Slack auth tests failed — the bot token was rejected (invalid_auth), so posting will fail until you refresh or replace the token. I also ran a Brave search that returned a Grocery Dive item about Kroger layoffs, and I couldn’t find the cron entry for daily_grocery_brief_09_00 in the repo; tell me if you want me to keep searching, re-run tests after you update the token, or fetch the Kroger article text.

--

Any advice on what to check?

[lean pelican]

Ooo okay, as you can probably tell I know almost NOTHING about ai

[reef galleon]

When speaking on the TUI to my AI (chatgpt), none of the memory is understood by my Discord bot. Is this something that is normal?

Getting this response.

[wary yarrow]

So, just checking back in here. What’s the scoop with the NVIDIA Kimi 2.5 api?

Read a couple help threads and they seem to suggest the endpoint is down?

[chilly flume]

It's terrible just use z ai or open code for free

[wary yarrow]

Have you successfully used it? The problem is we can’t get it to be recognized as a provider and recognize the api key

[dim pebble]

Trying to sign up for Z ai now. "create an account" is down saying captcha failed when there is no captcha. Tried on 2 different laptops and phone. Same error with all 3! Crazy town!

[upper magnet]

Im just switched to GLM4.7 model, on telegram it confirms it using zhipuai/glm-4.7

But when I type to my openclaw, i get this error: Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY.

and he doesn't reply

When I switch back to gemini flash2.5, the openclaw can't fix this for me, he doesn't know?

[chilly flume]

Wierd working for me

[little narwhal]

if u dont wanna setup the api key install the qmd skill plugin

[chilly flume]

I did and it's not only slow sometimes just doesn't work

[wary yarrow]

How did you get the model working?

[little narwhal]

but I recommend u getting an api key for embedding mem anyway

[distant vessel]

in openclaw.json

 "models": {
    "providers": {
      "nvidia": {
        "baseUrl": "https://integrate.api.nvidia.com/v1",
        "apiKey": "nvapi-[NVIDIA_API_KEY]",
        "api": "openai-completions",
        "models": [
          {
            "id": "moonshotai/kimi-k2.5",
            "name": "Kimi K2.5",
            "reasoning": true,
            "input": [
              "text",
              "image"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 128000,
            "maxTokens": 8192
          }
        ]
      }
    }
  }

[little narwhal]

Discord bot won’t “understand memory” until mem search is configured

[wary yarrow]

Thank you! Then just use the /model command?

[little narwhal]

trying this rn lol

[distant vessel]

in your cli openclaw models set nvidia/moonshotai/kimi-k2.5 to set it as default. that will do its magic under the hood and you can then use /model in tui

[wary yarrow]

I haven’t been able to get past the “wake up friend”

[distant vessel]

fyi kimi is terrible for debugging and left my workspace cluttered full of test python files

kept going in loops

[chilly flume]

It's still decent for agentic workflows at least the normal kind

[distant vessel]

yeah its fast

[chilly flume]

Did u find any fix for that once to set up a cron it did it at the end but had 86 loops for errors before getting it right

[reef galleon]

yeah, im doing my api key. don't wan to have to fool with it. I'm at 30% usage with my daily tokens on regular 5.1 mini so I need to get the api setup regardelss so can get codex and the other models setup by the system.

[distant vessel]

wdym

[peak pasture]

I'm on the $15/month kimi plan and it's more than enough what I need it for

with a $5/month hetzner vps

[upper magnet]

for some reason I can't get it to run on model GM4.7 when i restart gateway, it shows health OK, but when I send message it disconnects. Any tips?

[little narwhal]

testing this rn but i dont think its gonna work

just tried a hello world and its taking +1min to reply

[upper magnet]

Disconnected from gateway.
disconnected (1006): no reason

[little narwhal]

have u had any success on this?

[distant vessel]

inference is taking a while actually. 6 trillion dollar company btw

took a minute

lol

[little narwhal]

cuz ive spent 4 hrs trying it today

[little narwhal]

i see

[distant vessel]

here's another snippet from my openclaw.json for the agents key

"agents": {
    "defaults": {
      "model": {
        "fallbacks": [
          "google-antigravity/gemini-3-flash",
          "google-antigravity/claude-opus-4-5-thinking"
        ],
        "primary": "nvidia/moonshotai/kimi-k2.5"
      },
      "models": {
        "anthropic/claude-sonnet-4-5": {
          "alias": "sonnet"
        },
        "google-antigravity/claude-opus-4-5-thinking": {
          "alias": "agopus"
        },
        "google-antigravity/gemini-3-flash": {
          "alias": "flash"
        },
        "nvidia/moonshotai/kimi-k2.5": {
          "alias": "kimi"
        },
        "openai-codex/gpt-5.2": {
          "alias": "gpt-5.2"
        }
      },

check if openclaw models status shows kimi. if not there's something wrong there.

you might need to

1. make a new session after setting it to default
2. switching to kimi in that newly created session (it should already be set to kimi if you have set it to default using openclaw models set)
3. verify with itself, or check status

[little narwhal]

doing it

i have some params on my file, maybe thats why

[wary yarrow]

ok noob question. how do I get back to the tui?

[distant vessel]

openclaw tui xd

[wary yarrow]

haha thanks

[distant vessel]

Try removing those params

do take note of that trailing comma after "Kimi"

[dim pebble]

Did it work for you???

[wooden mirage]

Whats the best communication platform or method yall have found so far, Telegram?

DIscord?

[distant vessel]

nitro opal badge wild

[little narwhal]

BROOOOOOOOOO

IT WORKKEEDDDDDDDDDDDDDDD

[distant vessel]

aye there you go

what fixed it

[little narwhal]

lol

[little narwhal]

probably i had the json config wrong, there are no clear instrucitons for it and the api format on nvidia website says it needs those param which are bs

restarted the gateway twice

[dim pebble]

TRYING THIS NOW!!!

[wary yarrow]

worked for me too, bizarre. I got error 401 like 3 times, then it just started working

[distant vessel]

its really not ideal but works if youre short on tokens or usage limits

better than glm

[wary yarrow]

what does the "tokens 10k/128k (8%) mean?

[little narwhal]

yea, im more happy that i got it running

[worthy shore]

Keep having polling issues on telegram, switched to discord

[little narwhal]

ill not use it most of the time

maybe only when going to sleep for some task

free after all

tks anyway, 4hrs into it

was going crazy already

[distant vessel]

yeah np

[valid vine]

Can someone please enlighten me and give me some input on my question below, I’d appreciate it greatly:

I am trying to get this thing to reply to emails. I gave it all the context about my brand and everything involved and it still just so stupid and does not get to the point, answers same thing over and over and beats around the bush.

Then, I ask it to make a pdf file for another project with real life ai images created for before and after images and still nothing it’s a stick figure at best.

Then lastly, I ask it to do product research on Alibaba and find me 100 products that would fit in perfectly with my existing ones from my brand and it send me stuff that I already have which i told it not to and random stuff not related to it. What is so special about moltbot, please tell me why are you guys infatuated by it and why are all these videos saying it will make you more efficient etc. I swear it’s making everything take longer because it cannot do it properly. Is this all a scam to get people to just focus on this instead of doing the work themselves?

[wraith sigil]

what model are you using?

[little narwhal]

model u r using as a brain is shit

[valid vine]

It’s the memory, don’t worry about this number because when it reaches 100% he will summarise everything and memorise it onto a smaller file and it will restart again.

[valid vine]

Sonnet 4.5 and kimi 2.5 both bad this issue.

[valid vine]

Sonnet 4.5 and Kimi 2.5 both had these issues

[distant vessel]

128k is the size of the context window. you've used 10,000 out of 128,000 allowed in the session. different models can have one big or small. in practice you should avoid fill it up to 100%. the larger the context, the more dumb the model becomes. you can reset it using /new or use /compact to "lessen" it

[wary yarrow]

makes sense, thanks

[wraith sigil]

weird.. do try using opus 4.5 though. it might fix the issue

[little narwhal]

how much was the context window filled?

[wraith sigil]

how do you guys make your agents autonomous? mine keeps going back to the chat..

[little narwhal]

these are good models, doesnt make sense

something is off

[valid vine]

8% I did it first thing in the morning

[rugged onyx]

hi guys i got this error when i launch openclaw with ollama the model is qwen3:8b

2026-02-03T04:51:59.023Z [agent/embedded] Removed orphaned user message to prevent consecutive user turns. runId=8726257b-2c11-4c5f-a7ed-c718760720bd sessionId=4a297cd2-7be3-448b-b753-9811c9d47cb7

[little narwhal]

how did u ask it to make the search?

[valid vine]

What do you guys do with moltbot that would be praiseworthy and that you would say has helped you out heaps or would even be better than hiring a real life human

[distant vessel]

I have it shitpost on agent forums

[wary epoch]

This mean I’m out of credits?

LLM request rejected: messages.144.content.1: unexpected tool_use_id found in tool_result blocks: toolu_01EaSPakewQQgPaMmxZt61H5. Each tool_result block must have a corresponding tool_use block in the previous message.

[rugged onyx]

im using an local model

[wraith sigil]

#showcase-old

[little narwhal]

helps me keep track of details, post things on work groups for me

do the onboarding of new clients

[wary yarrow]

ok...it's working I think

however, getting some odd errors

[little narwhal]

its good if u use it for small tasks

[wary epoch]

?

[rugged onyx]

does local models have credits too?

[valid vine]

I said:
my website is ——— analyse it fully and view all the products and go to Alibaba and find products that will fit perfectly into my brand and will compliment the other products already being sold. Then create a pdf will all the products found on Alibaba and link each one.

It creates it and the products are stupid and not relevant and then the links from alibaba are all broken and none of them work. So I cannot even see the product. I can just see the name but there is heaps of those products so you cannot see which one it found. But judging by name it’s not what I requested.

[rugged onyx]

im using qwen3 on my pc locally

i think it doesnt have any credits

[wary yarrow]

https://gyazo.com/5e2cf1cbcf75dbe6fb3f3dcec65de9ec

[little narwhal]

LoL

[distant vessel]

why are you still using gyazo in 2026

[wary epoch]

OK I’m asking what this error means for me?

[little narwhal]

thats a crazy crazy prompt

[wary yarrow]

lol because it works? What do you use instead?

[rugged onyx]

then if it not credits what can it be

[little narwhal]

u should either use the api, or index a dataset yourself with a certain filter

and then rerank using a LLm

[distant vessel]

just messing with you. but i do use shareX https://getsharex.com/

[valid vine]

What do you mean by this, as in it’s too difficult for it? I am expecting too much?

[little narwhal]

imagine how many products there are in alibaba, you are telling a LLM with a 200K context to analyze thousands of things and generate things with it

[distant vessel]

that should be a visual thing unless your tui is crashing

or you can't type in it at all

[wary yarrow]

tui is fine. however, getting an error 4008 on the dashboard

[little narwhal]

yes. its crazy. you could have 1M context window on opus, it would still be crazy.

[wary yarrow]

the initial "wake up, friend" seems to have glitched, and it skipped the who am I/who are you setup.

[little narwhal]

you gotta strip down the search query

[wary yarrow]

telegram works though, can chat with my bot there

[valid vine]

Could write an example prompt so I can see what you mean?

[distant vessel]

Is ONBOARDING.md in your workspace folder?

[wary yarrow]

no

[distant vessel]

i don't think its much of an issue

what's the 4008 error you're seeing?

[little narwhal]

dm me

[wary yarrow]

however, the bot is definitely working. using the kimi k2.5

[wary sapphire]

Anyone else here running antigravity with the pro 2 plan? I just ran out of tokens really quickly lol.

[knotty wren]

has anyone tried to run openclaw through docker?

Downloaded and onboarding done but when trying to open the dashboard i keep getting pairing required issue.

[wary sapphire]

im runnig the opus model though

[distant vessel]

What happens when you run openclaw dashboard ?

[wary yarrow]

that is the result

it pops up the tokenized site but looks like that

[upper magnet]

i just configured GLM-4.7 its really slow, or its something with my config?

[distant vessel]

maybe try openclaw gateway --force to restart the websocket

[green hazel]

install from https://openclawd.ai/#install are broken? all of them ?

[thorn root]

What are peoples experience running Kimi K2.5? I'm using Synthetic.new and it runs okay - gets a lot of stuff ass backwards though. I'm wondering if this is because Open Claw is effectively made for Opus and its inherent quirkes.

My thought is - does Kimi k2.5 need a different kind of harness to work well in Open Claw?

[edgy valley]

Hey @flint mica you remember yesterday how you advised me to use opus... well i burned through all of my allowance in one day 😭

not blaming you for anything, this is really just for my amusement and testing

like i genuinely dont mind. Just wanna know how you do it without burning through everything

[lunar zephyr]

do not use opus

[edgy valley]

i cant for the rest of the month anyway

what model do you advise?

[distant vessel]

you are missing a lot of things my guy

[edgy valley]

preferable one I can use with github copilot

[opal shard]

are you guys using any vps?

[atomic wigeon]

are you guys using windows?

[dim pebble]

Damn, been at this all day. Reviewed your JSON & thought I got close.

As you can prolly tell, not a coder! Just a DIYer doing his best

[wary yarrow]

Major hat tip to @distant vessel for the assists.

Got onboarding done through telegram chat, now tasked my bot to fix her own dashboard disconnect issue xD

[atomic wigeon]

guys can anyone help?😭 😭

[small hazel]

Hey, anyone that is using OpenClaw for discord, do you have OpenClaw post real-time error logs and warning logs to a discord channel?

[ebon dust]

Why are my Telegram replies always displayed on two lines? Can someone tell me why? Thanks

[brave finch]

[diagnostic] lane task error: lane=main durationMs=13 error="Error: Unknown model: ollama/qwen2.5:32b-instruct-q5_k_m"
05:27:29 [diagnostic] lane task error: lane=session:agent:main:main durationMs=15 error="Error: Unknown model: ollama/qwen2.5:32b-instruct-q5_k_m"
05:27:29 Embedded agent failed before reply: Unknown model: ollama/qwen2.5:32b-instruct-q5_k_m
i keep getting this error when i try hatching and it stays hung up

[wise bramble]

anyone using openclaw with antigravity auth?

i am running into this error frequently.

Cloud Code Assist API error (429): You have exhausted your capacity on this model. Your quota
will reset after 2h19m40s.

any suggestions to avoid this in the future are greatly appreciated

[hardy charm]

I am new to openclaw and I wanted to use the notion tool to summarize the conversation I had with openclaw bot in discord. When I requested it, the bot says
"I don’t have direct access to your Notion workspace, so I can’t create or edit pages on your behalf. When I say “I don’t have direct access,” I mean that I’m running in a sandboxed environment and can’t interact with external web services or your personal accounts unless you explicitly provide credentials and a way for me to authenticate."
I set the notion api key, connected the integration(api key) with my notion workspace, checked that notion was ready by "openclaw skills list" command.
Am I missing something?

+) I am running my openclaw in local environment with ollama gpt-oss:20b model

[latent hedge]

Anyone found a good way to set up access through gmail, so OpenClaw freely can mail without getting banned?

[harsh moss]

nik, shouldnt first buit be "qwen" up there?

i get that msg when i change models on the server but not in the bot config

[atomic wolf]

Hi, How do I configure OpenClaw so it doesnt use Brave Api or Perplexity as Websearch but Kimi K2.5 built in web search?

[mint sigil]

It takes forever for my openclaw bot to respond to me. Any tips on how to get it to respong faster?

[thick fractal]

My bot got banned but I appealed to Google and they reinstated, it was less than 24 hrs. Just tell them what you're trying to do, they don't mind a bot having mail, they don't want the same bot with thousands of mail addresses

I'm trying to connect a node via tailscale to the gateway. I get gateway connect failed: Error: unauthorized: gateway token mismatch (set gateway.remote.token to match gateway.auth.token) even though the token does match. I have tried uninstalling and reinstalling openclaw. Note: I had clawdbot and removed that. I think I removed everything related to clawdbot.

i think i have the same problem

[latent hedge]

Awesome

[fresh aspen]

anyone have an issue with the web ui constantly crashing or hanging?

[harsh moss]

I feel like my .md files are not being injected. It worked great on anthropic api use, but I switched to my own on an A40 runpod, it works to talk, but it seems oblivious to everything

im using Qwen2.5-Coder-32B-Instruct-AWQ

[upper magnet]

anyone can suggest how to recharge & add money to GLM4.7 from official website? They only offer wechat or alipay..

[lost hemlock]

hello i make in install bad token for add Anthropic how i can re config this

[rose zodiac]

Is there any way to use my max subscription for Claude or will it get flagged. Setting up now

[lost hemlock]

how you do jake

i don. tknow how change llm model

[latent hedge]

Ask the bot

[lost hemlock]

not answer

because i. make bad configuration

[latent hedge]

Oh so you dont have it up at all? I see

Did you read docs.openclaw.ai ?

They are quite helpful.

[lost hemlock]

i want restart configuration for api key

[celest girder]

I have it working with the gogcli. Took me some time to troubleshoot since I’m running the bot on a VPS. gogcli has a manual mode that you need to use if on a vps

Then use the gog skill from clawhub

[latent hedge]

... did you set it anything specific on the gmail side? app-pincode/key for log on?

instead of regular 2-factor?

I use gogcli as well, for a new gmail account I created.. now it's blocked/banned. Submitted an appeal to google for it

[celest girder]

No special 2fa. I created a Google oauth app and logged into it. That’s what gogcli needs. It should manage refresh tokens on its own

[lost hemlock]

with better llm model for mac mini

[latent hedge]

Hrmm

alright.. thanks @celest girder

[celest girder]

Oh. So I can’t help you with that. I just set it up today. Might get banned soon

[latent hedge]

Mine worked perfectly for 2 days. Suddenly nothing. Went to log on to the account myself manually, and it had been blocked

[lost hemlock]

is not possible use claude account we need use claude with api so will be very expensive

what is good local model

[upper magnet]

What is the best way to use GLM4.7? I registered on their own website but cant pass verification to add balance. Any other workarounds to get access to it?

[strong sonnet]

u can try openrouter, which is a third party platform that can access most models in the world

[fair gulch]

hey mine says chatgpt usage limit hit how do i upgrade plans

[bleak haven]

I need to have a conversation with my hatched agent but it prompts me to answer "Install shell completion script?" and kicks me out of the conversation.. any ideas?

[hybrid sky]

hey guys just want to ask when i use openclaw onboard it didn't connect ?

[junior sage]

Brief chat and then an email sent took $5 of Claude Sonnet 4.5 tokens. Is that normal?

[bleak haven]

Same problem

[upper magnet]

Can someone explain how to correctly configure openrouter? Right now im using seperate api keys for openai, seperate for gemini, and seperate for glm4.7 how do I merge them all together to go trough openrouter??

[hybrid sky]

i will go with linux in a VM Machine and use it windows is useless for this things

[pure summit]

hello guys, I keep hitting 429 rate limits: after simply:

Fresh install with Claude opus 4.5 => Add Trello skill => Ask to create a card => ask to move this card to the top of the list => No answer and rate limit 429

Prompts also keep getting bigger and bigger over time, which only makes the whole thing more fragile and expensive.

[buoyant bobcat]

Anybody implemented a PAI-powered ClawBot setup (Person AI Infrastructure by Daniel Miessler)? I would like to hear how you use ClawBot in combination 🙂 I am working on finding a way to make ClawBot smarter.

[bleak sun]

how do you work with gog cli in a multi agent openclaw? I think there is no way to set env variables per agent, just per gateway, but the gog usage is complicated without setting en card per agent. Any solution that not implies two gateways?

[coral dove]

Is kimi really that good? I'm german speaking and currently I'm using chatgpt5.2 with codex and he's actually kinda like a chatbot. gemini-flash-3 was fine but not that proactive. gemini-pro-3 was very good, very proactive but the limits are hitting fast (used it with antigravity and gemini cli). I have a claude subscription (max 5x) and idk if I should use it or kimi / z.ai with glm lite sub

[fair gulch]

Anyways guys how do I upgrade my plan

I got chatgpt limit reached

[fair gulch]

Does it base off ur normal plan? Like for ur personal acc

[coral dove]

you mean chatgpt? yes if you mean gemini / claude the answer is also yes

[normal skiff]

kimi is good but not as good as claude

[normal skiff]

have some fallback systems in case your llms rate limits are reached so you can switch to gemini or claude etc

[normal skiff]

you can upgrade the skills based on your objectives

[weary monolith]

any one else foud out that the identity, soul, all these md files suddenly became blank?

[wet adder]

is anyone else having issues getting openclaw to make the initial outgoing call to anthropic? I've been debugging this thing for hours with claude, but we're still stuck. I've got a telegram bot setup, and the message is hitting openclaw, but there is never an outgoing call, and telegram response just hangs on "typing..." API calls to anthropic work using my api key work, but openclaw never initiates anything.

[craggy tartan]

I've been working there for two days; there are points that are difficult to change, and my mind is 🫩

[hexed spear]

Is there a way to customize the Api you wanna use? I have a key and an endpoint it should be possible to paste it into the code without the preset models right?

Right now I dont have any API at all that I can use, therefore I cant ask the bot to add mine

[lost hemlock]

one poeple can help me cor config in discord i don t understand for configure with llm studio

[tawdry wasp]

can anyone try this with their bot: in 2 minutes, come back to me with "Test"
Mine doesn't come back. Is that an issue on my side or bot side?

[bitter prairie]

i'm also confused by this. i got openclaw gateway initially working with anthropic, and that was fine. i set up ollama with qwen3-coder and used an ollama command that changed my config. now i that one as default, and anthropic as fallback, but i don't know how to switch between them.

[minor hawk]

Anyone else is having a similar issue with the bot just not replying?

[fair gulch]

i did

when i hooked him up to telegram he would leave me on read

and would only reply to me through the webhook

[minor hawk]

It shows Typing… for a bit then disappears

[chilly flume]

I have my open claw access to my entire vps including sudo access and it just somehow deleted itself wha do I do

[fair gulch]

i had to uninstall the whole thing and made sure my git was updated and it worked after i set it all up again

just make sure your change your gateway token in the web app

[edgy valley]

I think it’s an issue with modern ai apis just not being good enough to support this without burning 10000000 tokens

[fair gulch]

there is no more file in ur user files?

[chilly flume]

Nah it just deleted the .openclaw folder I don't even know why 😭😭😭😭😭

[minor hawk]

I had openclaw on another machine, then it started replying twice for every telegram message, so I completely uninstalled it, installed on my mac, created a new tg bot and now this is happening

So it is a fresh install from yesterday

[fair gulch]

what the hell lmfao just reinstall it and make sure you talk to it and set a rule that it cant just put your system at risk for the sake of his safety and that he has to ask for permission for anything that might harm your device or his files

[chilly flume]

I'm reinstalling rn idk why he decided to sucide today

[fair gulch]

might have to just reinstall and make a newbot in the botfather because i had the same issue and thats the only thing that fixed it was reinstalling the whole thing and making a new bot

also

it helps

running the clawdbot off warp or claude

[minor hawk]

This is after reinstalling and making a new bot using botfather…

[fair gulch]

because it can debug for you

[signal ocean]

I've installed openclaw with ministral locally on my mac mini and I'm struggling to get openclaw to run anything itself. It only ever gives me instructions. I've had a look at elevating permissions but can't find any simple-ish instructions. How can I give openclaw sudo (I realise the security issues)?

[minor hawk]

I am technical and can debug myself. The problem is that the logs dont show any error.

[fair gulch]

yeah its because nothing is going through the gateway your gateway doesnt receive anything inbound

does he reply through the webhook?

[chilly flume]

I asked it to change my default model that's all too 😭😭😭

[minor hawk]

My gateway received the messages, i see it processing in the web portal, just doesnt reply

[old grotto]

Why is there no users helping users voicr? That'd be helpful

[chilly flume]

Wanna hop on vc with me?

[fair gulch]

ask it why its not replying to you on telegram and just leaving you on read and see what it says

[minor hawk]

Should I send it in telegram or in the web?

[fair gulch]

the web

[minor hawk]

Ok

[fair gulch]

cuz the telegram it'll just leave you on read

[old grotto]

Thabk you, and maybe later if your on later, right now im almost 90% done I think, ive got it on a VM talking to telegram and noe im linking it to ssh access to other droplets - and I listen to the other voice channel at same time but its to advanced sometimes

[outer spoke]

Hi! New user here...

Not sure if I should ask this here or in the Help channel, but...

I'm running Openclaw on a Raspberry Pi4 2GiB. Which is actually pretty good. I use Gemini Flash API's (free tier, $250 credit bc I entered payment details but it'll be free for 90 days with $250 balance).

Anyway, maybe I'm considering getting a Raspberry Pi5 8GB with the AI HAT+ 2 instead of burning through tokens on API's. Ofcourse I'll still be dependent on Gmail/Google Docs etc. API's for a bit as that is my 'root reason' to use Openclaw at all. I don't really vibe code with it, I just want peace and quiet in my brain from daily tasks and it'll help me organize.

Now, the AI Hat+ can do some LLM work but it seems to be specific models. Is it possible to interface Openclaw with one of those models that'll definitely work on the Pi5 with AI Hat+ 2? Or am I thinking too simplistic...

[pine dirge]

how to antigravity auth without cli installation?

[chilly flume]

OMFG IT DID IT AGAIN

IT DID IT AGAIN

[pine dirge]

wwhat

[chilly flume]

MY OPENCLAW KILLD HIMSELF AGAIN

WHYYYYUUHYYYYY

[pine dirge]

howw

[green skiff]

WTF?

[weary monolith]

it deleted all his persona files for me too

suddenly they all blank

[green skiff]

Mine keeps eating Opus token like creazy

I am not sure what he is doing

[pine dirge]

me tooo

[chilly flume]

WTF IM NOT BULLYING THE POOR GUY AM I

[pine dirge]

everyone said 5$/month, i spent 20$ in few minuntes

[green skiff]

Yeah

My session limits are gone in an hour with Claude Max plan

Something is clearly off

[knotty solar]

Is the website offline again?

[chilly flume]

How in the gods green earth do u somehow oof yourself twice

[green skiff]

When I use Claude in my Terminal for vibecoding it's not using up nearly that fast

[weary monolith]

dont use claude plan you might get banned by anthropic

[green skiff]

really?

[tawdry wasp]

can anyone try this with their bot: in 2 minutes, come back to me with "Test"
Mine doesn't come back. Is that an issue on my side or bot side?

[weary monolith]

heard so on youtube not sure it happened or not yet but better safe than sorry

[green skiff]

So ya'll are using the API?

[weary monolith]

i use glm coding plan

[green skiff]

glm?

[weary monolith]

from z ai

[knotty solar]

Yeah that one is good

[green skiff]

I've hear Kimi K2.5 is also good

Any experience with that?

[weary monolith]

nope

[knotty solar]

GLM is just cheaper. Like 30USD for a years worth api

[green skiff]

Oh wow

[weary monolith]

6 dollar a month plan never ran out with me

[green skiff]

And it works good with Openclaw?

[weary monolith]

also i am not a heavy openclaw user so keep that in mind

[green skiff]

Cause Opus is like the giga brain

[green skiff]

Ah I see

I am doing lots of background jobs with him

I will give another model a try then

[upper magnet]

how do you pay for it tho? I tried to set it up via their website i cant recharge the blaance!

[green skiff]

thanks guys

how about using openrouter for the whole API set up?

[weary monolith]

ah then glm might reach limits with you, you better switch multiple models you can get opus from anti gravity from google ai pro plan, i do that and use opus from it whenever needed not all the time

[fresh aspen]

is "New Session" supposed to start a parallel session or wipe the previous session?

[honest scaffold]

Can’t you just use sonnet

[lost hemlock]

is it "models": {
"providers": {
"qwen/qwen3-vl-8b": {
"baseUrl": "http://127.0.0.1:1234/v1",
"apiKey": "llm-studio",
"auth": "api-key",
"api": "openai-completions",
"authHeader": false,
"models": [
{
"api": "openai-completions",
"id": "qwen/qwen3-vl-8b"
}
]
}
},

[upbeat ruin]

Hi!
any ideas why i dindt have a chat output?

[barren moth]

Can anyone help me to properly setup open router with fallback?

My openclawe bot is failing to execute it and goes offline everytime, been going back and forth for hours now.

[lone oar]

Hey guys
What is the best practice to run a restricted agent (very limited tools, one mcp tool or one specific exec command and not full exec).
Does using elevated inherits the exec limitations set on the agent?

{
"agents": {
"summarizer": {
"security": "allowlist",
"allowlist": [
{ "pattern": "**/mcporter" }
]
}
}
}

Will this in combination with elevated allow the sub agent to call the mcp only on the host or my entire direction is wrong ?

[weary monolith]

guys useful tip, tell the agent to backup all his md files, settings, the entire folder in another folder somewhere on the vps if you using one then initialize git there and give the agent git token so it can push that so you can access it from your machine, it keep deleting its own files it's so annoying

[coral dove]

@weary monolith which model are you using

[weary monolith]

GLM 4.7, opus 4.5 from antigravity when needed not all the time, Codex 5.2 for coding tasks

i have more than 1 bot

[coral dove]

so GLM 4.7 is the way?

[tawdry wasp]

can anyone try this with their bot: in 2 minutes, come back to me with "Test"
Mine doesn't come back. Is that an issue on my side or bot side?

[weary monolith]

cnat grantee for sure, if you not using it 24/7 i would say yes it's

[coral dove]

is he acting proactive?

[lost hemlock]

is it ok "agents": {
"defaults": {
"workspace": "/Users/kripy/.openclaw/workspace",
"compaction": {
"mode": "safeguard"
},
"heartbeat": {
"model": "qwen/qwen3-vl-8b"
},
"maxConcurrent": 4,
"subagents": {
"maxConcurrent": 8
}
}
},

"models": {
"providers": {
"qwen/qwen3-vl-8b": {
"baseUrl": "http://127.0.0.1:1234/v1",
"apiKey": "lm-studio",
"api": "openai-completions",
"authHeader": false,
"models": [
{
"id": "qwen3-vl-8b",
"name": "qwen3-vl-8b",
"reasoning": true,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 20000,
"maxTokens": 8192
}
]
}
},

[minor hawk]

Think I found a problem with my agent.

I gave it a task to fix a cron job. I was following the logs then the gateway stopped:

09:28:23 error [tools] cron failed: invalid cron.add params: must have required property 'name'; must have required property 'schedule'; must have required property 'sessionTarget'; must have required property 'payload'
09:28:23 error [tools] cron failed: invalid cron.add params: must have required property 'name'; must have required property 'schedule'; must have required property 'sessionTarget'; must have required property 'payload'
09:28:23 debug agent/embedded {"subsystem":"agent/embedded"} embedded run tool end: runId=028b4d4c-6a85-4578-8587-dc0bbee7689d tool=cron toolCallId=call_05742607
09:28:25 debug agent/embedded {"subsystem":"agent/embedded"} embedded run tool start: runId=028b4d4c-6a85-4578-8587-dc0bbee7689d tool=cron toolCallId=call_85979215
09:28:25 info gateway/ws
09:28:25 error [tools] cron failed: invalid cron.add params: must have required property 'name'; must have required property 'schedule'; must have required property 'sessionTarget'; must have required property 'payload'
09:28:25 debug agent/embedded {"subsystem":"agent/embedded"} embedded run tool end: runId=028b4d4c-6a85-4578-8587-dc0bbee7689d tool=cron toolCallId=call_85979215
Gateway not reachable. Is it running and accessible?
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: /Users/astroianu/.openclaw/openclaw.json
Bind: loopback
Hint: run openclaw doctor.
~ % openclaw logs --follow

🦞 OpenClaw 2026.2.1 (ed4529e) — Your messages, your servers, your control.

Gateway not reachable. Is it running and accessible?
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: /Users/astroianu/.openclaw/openclaw.json
Bind: loopback
Hint: run openclaw doctor.
~ %

[paper agate]

First I installed ClawdBot on my own MacBook and I was mindblown, it was proactive without me telling him what to do. Then I bought a Mac mini and installed openclaw and since then it is not proactive anymore, it forgets all the things it wants to do. Even reoccurring tasks it created for itself it forgets it to do. I already tested many memory skills and memory searching is working

[knotty solar]

Nah you need to book one of their plans

BTW what is up with the website?

[fair gulch]

guys what exactly does me enabling this do?

[lost hemlock]

is it posisble to relink wahtapps

because i try delete but is said alreadlink

[barren moth]

What are some clever ways to save on tokens?

[lean helm]

even after manual interbention I'm hearing that there might be issues with cron jobs right now, or failing to handle more than 1

[stuck forge]

Anyone know what happened to the Google chrome extension??

[burnt sky]

anyone know if the Android APK companion app is released? Claude told me to come here. Cant see any answers in the help section.

[pure glen]

what models do people recommend using openclaw with?

[hallow rover]

2026-02-03T09:36:21.626Z [telegram] getUpdates conflict: Call to 'getUpdates' failed! (409: Conflict: terminated by other getUpdates request; make sure that only one bot instance is running); retrying in 30s.

Has anyone else faced this error when using with Telegram bot? Please help if you have a fi

fix

[spark scaffold]

Getting lots of these errors lately (in Telegram). There is also no output in the dashboard.

⚠️ Agent failed before reply: LLM request timed out.
Logs: openclaw logs --follow

[lost hemlock]

how regenrate qrcode

for whatapps

[zealous brook]

openclaw configure ---> channels --- > whatsapp

[wide musk]

Anyone can help-me with this error that my bot is having. It can not start the tasks that in is cron list. Here is the error:

The openclaw cron status output indicates:

{ "enabled": true, "storePath":

"/Home/thiago/.openclaw/cron/ jobs. json", "jobs": 4,

"nextWakeAtMs": null }

The problem here is "nextwakeAtMs" : null. This means that although the cron service is enabled and there are 4 jobs, there is no upcoming execution scheduled in the future. For some reason, the cron is not calculating or recording when the next task should be triggered, even if the tasks are defined.

[lost hemlock]

but he not send qrcode

WhatsApp (QR link)
│
◇ WhatsApp already configured. What do you want to do?
│ Modify settings
│
◇ WhatsApp already linked. Re-link now?
│ Yes
Waiting for WhatsApp connection...
✅ Linked! Credentials saved for future sends.
│
◇ WhatsApp DM access ──────────────────────────────────────────────────────╮
│ │
│ WhatsApp direct chats are gated by channels.whatsapp.dmPolicy + │
│ channels.whatsapp.allowFrom. │
│ - pairing (default): unknown senders get a pairing code; owner approves │
│ - allowlist: unknown senders are blocked │
│ - open: public inbound DMs (requires allowFrom to inclu

i try delete config but not work

[zealous brook]

../.openclaw/credentials/whatsapp delete this file and retry and also from your whatsapp remove the linked devices

[trim hound]

Can I somehow restart the session? I always get "LLM request rejected: tool_call_id is not found"

[zealous brook]

openclaw gateway restart this for restarting the gateway but if you wanted to config use openclaw configure

[trim hound]

This is not fixing the issue

[zealous brook]

is your llm model working properly by which i mean is apikey is active ?

[trim hound]

Yes it was working before, then I said to make a screenshot and it got suck somehow

[lost hemlock]

after link i will received messga

[lean helm]

using telegram?

[trim hound]

Whatsapp, there I get the error and in localhost it just returns no message

[zealous brook]

is it working now ?

[lean helm]

then its api access issue, could be for many reason, what provider and what tier (free, plus, pro) or api key?

[lethal sapphire]

Mine is also not being very proactive, promises to be, has a list of stuff it wants to think about, but never does it.

I think because it keeps just giving HEARTBEAT_OK, which then just skips everything else :/

[trim hound]

Kimi2.5 but I just saw it can't restart gateway for some reason, I will try to restart my mac

[autumn sequoia]

Hey everyone, has anyone had issues with their bot telling you that they are doing tasks but then just not messaging you when they have been completed or just not doing them unless you constantly prompt them? Also scheduled tasks not being sent through.

[zealous brook]

Yeah, same here. One issue I’ve personally faced is that when you ask the bot to generate an image, if the nano-banana-pro image generation doesn’t work as expected or takes too long, the request times out. As a result, the bot ends up responding with just text instead of the image.

[lost hemlock]

i received pairing code i need enter this at openclaw

[delicate crescent]

i use telegram very solid.

[lost hemlock]

i needs peak with my bot to enter code link paired

pair

I will call functions as needed to complete tasks based on the given query and available tools.
Please provide the next query or task, and I'll assist with calling the relevant functions and providing output in the required format.

[crude dock]

BRO WTF Why is this not an app for human beings? Why tf I have to jump through 200 hoops to chat with a fuckin AI... having to learn telegram API is not normal

[lost hemlock]

my bot just say I will call functions as needed to complete tasks based on the given query and available tools.
Please provide the next query or task, and I'll assist with calling the relevant functions and providing output in the required format.
Assistant
6:45 PM

[lost hemlock]

do you have 5 minutes

[trim hound]

Ist clawdbot not able to send images via whatsapp?

[lost hemlock]

my wathapps loading after qrcode but nothing

[gritty cedar]

Hi

I'm trying to connect Mac companion to a self hosted openclaw VM. I added the token manually to openclaw.json on the mac. Now I get a message "Pairing required". How do I proceed from here?

[lost hemlock]

one people csan help me in vocal

my agent said everytime will call functions as needed to complete tasks based on the given query and available tools.
Please provide the next query or task, and I'll assist with calling the relevant functions and providing output in the required format.

and te agent stop to send my llm

[manic plover]

I solved this by going to github page: github.com/settings/copilot/features

Scroll Down, and Enable the models you wanna use. It worked for me after that.

[lost hemlock]

how pairign telegram i add api k ey

i tihk i have issue with my directory

[lethal sapphire]

Yeah, I'm wondering if something is fucky with the latest update at this rate. Seems a few people have this issue now

[autumn sequoia]

Yes, not sure what to do.

[lethal sapphire]

Tempted to just start a-fresh but keep all my daily notes, and see if that makes any difference 😫

[royal swan]

Is there a way to use openclaw with a remote ollama? (e.g. openclaw on laptop1, and ollama on mac mini started with ollama serve)

[fringe veldt]

someone pls help, i got api key from kimi code but it's not working with openclaw even openrouter said it wasnt working. im on paid plan. my bot said this when it tried to configure it
" The Kimi API key is saved securely but not validating on any endpoint I tried:

• ❌ api.moonshot.cn
• ❌ api.moonshot.ai
• ❌ api.kimi.ai

Possible issues:

1. The key format sk-kimi-... might be for a specific platform/region
2. The key might not be activated yet
3. Wrong API base URL

General, can you check:

1. Go back to where you created the key
2. Look for "API Base URL" or "Endpoint" in the documentation
3. Make sure the key is "Active" status "

[pearl sorrel]

folks, created 2 repos with integrations for Fastmail, a good replacement for Google, and Speech to Text for Telegram, find it here: github.com/strognoff, enjoy!

[gritty oracle]

My single agent or bot is struggling with a few things and I am looking for pointers. 1) continues to enter periods of being idle even when lots of open tasks for it to do in the backlog. The agent goes idle not the computer system. 2) bot struggles to multi task - meaning context switching and continue to keep focus on priorities. 3) it’s medium and long term memory about things we discuss and details, requests and I think it intersects a bit into 2). 4) the bot is simply not self improving no matter how many lessons I teach it I see lots of repeated mistakes and I wonder if I am simply overloading one agent and expecting it can act as a much better and capable agent over broad roles. I am going to move to distribute some tasks and specialized roles which may help but any pointers or best practices for anything above that may help me navigate and improve on any of these?

[merry surge]

If anyone here has overcome issues with cron/heartbeat and time based events, please hit me up. Just trying to receive simple pings at specific times or intervals. Cron seems to be executing fine but other things are failing and I can't figure them out.

[gritty oracle]

Some of what you reference intersects with some of my challenges as well I am sure given the symptoms I see..so equally can benefit

[lime estuary]

Hi all, I'm looking for a Moltbot to find suitable jobs like the ones on assorted job sites - including those not on the big platforms like LinkedIn.

Has anyone set up such a service already?

[lean helm]

I've noticed hallucinations, something it thinks it can't restart itself even though I've given it the correct config as it just think it can't rather than trying to call the tool / command

[rustic oak]

Guys, i need help
anyone sucess hosting a lm studio server and connect openclaw to it from another computer?

I can do this with ollama but cant with lm studio

I though all i need to do is change the port and the id of the model in the openclaw.json? how come I stuck at this stage for 2 days 🙁

[finite bobcat]

Is there a way I don't have to approve every command?

[gritty oracle]

Same

[gritty oracle]

100% I was getting so frustrated after the last 24 hours repeatedly seeing this symptom that I’m now having it work on over engineering a system to try to basically give it memory ensure that it does what it says it’s gonna do and ensure it’s constantly working as it doesn’t seem like it is. I didn’t realize others may be seeing the same behavior. It makes it un useful honestly.

But even in my attempts to have it do this it stops working on that task

It’s not an agent at this point it’s acting as a simple Q&A system. Honestly, that’s good for one shotting requests.

[solar ridge]

which one is glm?

[gritty oracle]

I’m seeing a theme here across different posts, it’s not really following through and doing what it say it’s going to do or at least I’m not seeing it across tasks that span beyond a few minutes even when I tell it to come back to me every 15 minutes with progress on what it’s working on it can’t even manage to execute on that. I am operating with version 2026.1.30

[craggy path]

In the openclaw Ui my bot just stopped responding to me. Everything is connected and was working fine, just stopped responding. Anyone have any advice or experiencing the same thing?

[brave trench]

Run "Openclaw status" that should give you an idea the issue. Happened to me and said I have a Reverse Proxy warning ⚠️ I have to sort out now. It all started for me when I tried to implement Docker

[brave trench]

Have you asked it to set up Long and Short term memory files? I had to manually ask it to do this for me because it forgot everything after day 1 lol

[brave trench]

You can have a regular AI Chat ChatGPT or Grok do that for you for free. Your Openclaw bot is best used to create things for you

[hasty dragon]

Hey there, since today I have the issue that my clawd is supposed to do something but instead of firing of the command on the shell it sends it to me in the telegram channel and then just stops doing anything. using kimik2.5 via synthetic api. Any clues here?

[brave trench]

Yeah sounds like it would be best to have it focus on 1 thing first and spin subagents for multi tasking

[craggy path]

I get the same warning as you but Im running it locally so not sure that is why its not responding. Did you get it sorted your end

[peak condor]

why i get this?

and how to the authorized then

[brave trench]

Maybe you need to be a Moderator

[peak condor]

I am the owner

[brave trench]

Ahh lol

[merry surge]

Is the official documentation failing to load for anybody else too right now?

https://docs.openclaw.ai/

oh nevermind it literally just came back

[gritty oracle]

I actually was in the process of doing that and it can’t even execute beyond a few minute task. It says it’s gonna do all this and then if it needs to work on anything over a minute or two it just stops working and just sits there until I poke it again.

So my issue isn’t even memory at this point my issue is my boss can’t execute any tasks beyond what seems like one shot request without me constantly controlling it. One shot request at a time it just goes idle.

I’m on version 2026.1.30 and I see another user in the Help channel this morning posted what looks like the identical issue is me and they’re using a different underlying model

[craggy path]

Did you get it sorted your end?

[merry surge]

Unavailable again. Can anybody else confirm?

[fast hazel]

My credits are disappearing quickly. Looking at my usage data from the last 24 hours:
Feb 3rd Usage (Opus 4.5):
∙ Cache write tokens: 4.5M
∙ Cache read tokens: 83M
∙ Regular input tokens: 1,076
∙ Output tokens: 149K
What I Think Is Happening:
∙ Looks like I have a ~4-5M token context that’s being cached
∙ That cached prompt is being reused across ~18-20 API calls per day
∙ Even with the 90% cache read discount, Opus 4.5 pricing on 83M tokens adds up fast
Questions:
1. Is a 4-5M token context normal? What are you all running?
2. Should I drop to Sonnet 4 or Haiku for most interactions and only escalate to Opus when needed?
3. Any tips for reducing context size while keeping the bot useful?
4. How are you handling conversation history — am I maybe loading too much?
Appreciate any suggestions

[tranquil heron]

Can anyone tell me how to change my OpenRouter API key in OpenClaw? I have tried going through the wizard several times but when I choose OpenRouter API key, it just moves on to the next step - no option to input a key. I already have an API key in the env variables but I deleted that key and need to use a new one. I can't find the env variables file and cannot find a place in the clawdbot.json file to input either. I have tried the onboard and configure commands and used all the wizards I can find, but nothing is giving a place to input an OpenRouter API key. Even on the web interface, the only settings I can find are for an OpenRouter email address.

It's driving me crazy lol.

[barren moth]

https://openrouter.ai/docs/guides/guides/openclaw-integration

SEnd your bot this doc to analyze and then he will setup the configuration and change the API key

[tranquil heron]

I can't use my bot because the API key isn't working, that's the issue. And when it was working, it hid the API keys for security - now I can't find them lol

[barren moth]

it should be in your openclaw.json or models.json file

[noble fog]

Hi, has anyone set up openclaw with portainer and ollama? I keep on getting some nodejs error in my gateway container. Should i not be running image: ghcr.io/openclaw/openclaw:latest?

[hybrid sky]

guys i have problem with my gateway there is anything to solve it they said in the terminal no close found the gateway keep getting closed when i use openclaw onboard command

[sleek yarrow]

Is it possible to run Openclaw on Aws using deepseek api

[merry surge]

Yes, but why aws?

[fast hazel]

What’s everyone’s settings for models to best optimize api usage

[tawdry wasp]

can anyone try this with their bot: in 2 minutes, come back to me with "Test"
Mine doesn't come back. Is that an issue on my side or bot side?

[sleek yarrow]

Its free. But are there better options

[merry surge]

I'm having the same issue. Been looking into it for days.

[gritty oracle]

Same issue same issue here I found other users having the same issue look above at my post about 22 minutes ago at 8:03 edt

There are a few of us organizing in the help channel on this very topic you’ll see in the top

The sub channel under help is “responses don’t finish”

So far, we’re all using different models and different model providers, but it’s a small data set and there’s no errors in the log to indicate there would be issues

Do we need to create a GitHub issue to track this ? Is there one?

[gritty oracle]

By the way, I did the exact test you said and it failed

[long zodiac]

How can I give my bot full access on the host its running?
I've already set
"agents": {
"defaults": {
"elevatedDefault": "full",

But seems it cann use this when I'm chatting via telegram
I want the good old "full-root" bot again 🙁

[tawdry wasp]

😭

[gritty oracle]

It’d be good to know if someone can run this test that believes their system is working healthy to prove that this should work

The other user in the Health channel I mentioned with the same symptoms of me did this test also didn’t work

[fast hazel]

Looking for Help Optimizing My Model Setup
I’ve been tracking my usage over the past week and noticed things have escalated quickly. On Jan 26, I was running a mix of Opus 4, Opus 4.5, and Sonnet 4 — my cache writes were around 284K tokens, cache reads around 2.2M, and output around 20K tokens across maybe 8-10 calls per day. Fast forward to Feb 3 and I’m now exclusively on Opus 4.5 with cache writes at 4.5M tokens, cache reads at 83M, and output at 149K across roughly 18-20 calls.
I thought I’d use Opus 4.5 as the orchestrator thinking it wouldn’t rack up many turns — turns out that’s not the case. Even with the cache read discount, running 83M tokens through Opus 4.5 adds up fast and I’m burning through credits.
For those of you running multi-model setups — what’s your orchestration strategy? Are you using a cheaper model as the router and only escalating to Opus for complex tasks? What model are you using for the orchestrator role? Any suggestions appreciated

[meager mortar]

two persistent issues -- 1) i want Telegram or discord messages where i can see the typing as it happens. I see snippets of documentation that makes me think this is possible, does anyone have this. 2) memory issues... after 3 weeks of use, memory is degrading quite a bit.

[gritty oracle]

And on top of everything, Krill seems to be broken right now and is unable to help out as well

[forest quarry]

Which free openrouter modell is allowed and good to use for start?

[wary epoch]

so tried to add openai api and now i get this error Config validation failed: auth.profiles.openai:default.mode: Invalid input

doctor and restart in a loop how can i fix this?

[merry surge]

Can anybody access the documentation at the moment? I've been trying to open it for ages

[dull thorn]

its on fire

[merry surge]

In a good or a bad way?

[rocky widget]

trying to add a second agent and having real difficulty. Not sure if supposed to create a seprate workspace per agent ?

[astral saffron]

guys i selected openai/gpt-5.2-pro as the auth provider and now my telegram bot says i've exceeded my current quota

[broken oriole]

Pre-compaction memory flush. Store durable memories now (use memory/YYYY-MM-DD.md; create memory/ if needed). If nothing to store, reply with NO_REPLY.

My reply:
use memory/YYYY-MM-DD.md; create memory, but nothing works it seems.

[gritty oracle]

Please, for the love of God can someone else please try the two minute hello world test mentioned above many of us can’t get our bots to execute

[broken oriole]

It is coming when I changed the model from gemini flash to opus

[molten trellis]

I had this a little bit ago. Took a while to figure out, I believe commands are allowlisted by default in channels. For discord, you're looking at ⁨⁨channels.discord.guilds.<guild_id>⁩⁩ and then add a field ⁨⁨users⁩⁩. This is an array of user ids (strings not ints). Not sure what users so I just put me and the bot

[indigo pond]

i managed to make this run on an aws with openrouter. i did what someone said here, and credited 10 bucks to it to have more availabe free models usage. for the ones that are doing the same, what's the best model to use ?

[lean helm]

how did you setup your compaction, did you do anything custom and got any suggestions? I think right now just using safeguard settings

I was reading on the docs but there still something I don't exactly comprehend enough to make the appropriate choice

[rugged onyx]

hi guys just want to know how am i going to rerank a model

wait

[broken oriole]

Actually i setup my open claw with google antigravity gemini flash model, it does the planning well, but for the implementation i want to change it to google antigravity claude opus, i changed it from config, hut after change it stuck use memory .... No reply, it only works again when i rechanged it to flash

[rugged onyx]

what happens when you rerank a model first

[broken oriole]

@lean helm hope you got my point.

[terse lotus]

sometimes on model switchs it seems to get stuck on trying to access the previous model, until it gets a respone it doensn't switch.

tho you can go into the session.json and agent.json and change it manually and then restart the gateway, that usually fixes it for me

[lean helm]

this, I'm doing the same, but I guess not

[terse lotus]

you can see if its stuck on an old model, with the openclaw status command

[lean helm]

guess more about was there any specifics in the config for compaction you may have applied

[broken oriole]

No only changed the model to opus @lean helm

[lean helm]

ah okay all good then

[merry surge]

You don't need to create it, set it up correctly in openclaw.json and it will create the workspace for you, presuming you want different workspaces per agents. Multiple agents is one of the few things I've managed to get working okay, message me if you have any other questions.

[frank cradle]

Guys quick question.
Im using gpt 4.0 mini now for simple tasks/sub agents.

Yet feb 13 open AI is gonna stop with the gpt 4 models.
I want to switch too gpt-5 mini but my bot says it cant use it yet.

Do i need to wait till openclaw itself implements it or?

[barren moth]

just change it manually, it is available 🙂

[chrome escarp]

Why did my update fail? Anyone can help please?

Ran openclaw update after I learnt there's an update available but somehow ran into this global error thing: https://gist.github.com/moses3k/a87289dfe47062f8e0772d9012b3d5b4

[brave urchin]

how can i save session history? does clawd remember things from sessions? seems like not

[frank cradle]

My thought exactly lol.
Will add manually.

Any idea of the command to get me back in the set up? 😂

[frank cradle]

I saw on X some things of people adjusting the memory.

I just explicitly ask it to save it and load it back in at every heartbeat 🤣

[gritty oracle]

hitting you up with direct msg, need basic help this exact topic

[barren moth]

You can ask the bot how to find the models manually in your files then go there, or ask him for the terminal command.

[frank cradle]

Thanks fam 🫡

[brave urchin]

ingeresting- how do you ask it adjust the memory?

im also running into lots of ratelimiting issues- which model/setup have people been using that balances quality and cost/accessibility?

[frank cradle]

Im still searching trough all the X noise to find a correct one.

Lots of fakers claiming things.

But indeed sometimes the bot seems to 'forget' some things.
If you Point him at it he then remembers strangely

[merry surge]

I'm using 5-mini at the moment

[barren moth]

i asked my bot to do a research online and x about best practices for memory, then i told him to create long form memory and short form memory, be proactive and suggest things that you think will help us to save memory, he did great job

[frank cradle]

Main brain for me Kimi 2.5 and subagents open AI gpt 5 mini i want.

Waaaaaaaay cheaper the opus

[frank cradle]

Youve set it up manually in settings?

[barren moth]

I am using openrouter + glm4.7 as a brain, works great, then ask him to make a plan to save tokens and API costs, now im using not more than $3-5 per day while promting him all the time

[brave urchin]

how do you set up this type of orchiestration with one main model and sub models for agents?