users-helping-users

[tawny thistle]

previously was on earlier version + another different pr to solve the 4.6 bug, but that combo had thinking sig issue

I think 9th feb version

then I updated few hours ago to 2026.2.12 + this new pr

this new pr is supposed to 'Add Opus 4.6 support and fix thinking.signature error'

not sure if its the new update, or this pr that causes the issue

[lusty knot]

what's the best not so token hungry google model?

[tawny thistle]

api key config as in openclaw.json?

mind to clarify what do you mean?

but my keys are in auth-profiles.json, not in openclaw.json

is it not like that?

probably not config issue though?

[livid vigil]

guys what is the fix for whatsapp link , i link it with QR code and my phone also shows openclaw in device list . but when i try to send messaage i get error No active WhatsApp Web listener

Waiting for WhatsApp connection...
✅ Linked! Credentials saved for future sends.
meee@kclaw:~$ openclaw message send --channel whatsapp --account meee --target "[phone]" --message "hello from openclaw"

🦞 OpenClaw 2026.2.9 (33c75cb) — I'm like tmux: confusing at first, then suddenly you can't live without me.

│
◇
Error: Error: No active WhatsApp Web listener (account: meee). Start the gateway, then link WhatsApp with: openclaw channels login --channel whatsapp --account meee.

[tawny thistle]

since if I change the primary model, it works

[willow jasper]

is clawhub no default skill in the config wizard anymore?

[pseudo fjord]

have you checked the update already

Does opus 4.6 work already with AG

Would make my day if i knew this

[candid river]

Bro

Google banned me

Google banning people left and right

If you aren't banned, stop using the antigravity/gemini-cli auth plugins

[pseudo fjord]

I got multiple accounts

[pseudo fjord]

Wdym with this

[candid river]

Why all ya'lls got multiple accounts

[pseudo fjord]

Well

[candid river]

everyone I warn is like "iss chill, i got multiple accounts"

😭

I only had my 1

[pseudo fjord]

Cant disclose

But it would work if i have more accounts you think

Or just general issues

[candid river]

No, just saying a lot of people be having multiple accounts

which surprised me

[candid river]

Well, it's like, you have 10 account, 2 get banned, so what. 8 still work

But it's still risky, cause google has been cracking down on this

Soon all of them may get banned yk

[dense river]

anyone know what the 85/400k means? i wasnt aware that i only had 400k tokens to work with

│ agent:main:discord:channel:1466…                                                                                            │ group  │ just now │ gpt-5-mini             │ 85k/400k (21%)  │

[livid vigil]

in such cases always do : openclaw gateway restart

[unborn remnant]

Hi, I want to add a friend to a telegram chat with my bot, but I don't get a connection. The bot is not answering. I put the Telegram ID in the "Allow From" section, but still no connection. Any hints for me 👯

[livid vigil]

@do you know how to get that whatsapp group id . i would like openclaw to be inside whatsapp group , as now whenever someone send me message he get message , OpenClaw: access not configured. pairing code .

[unborn remnant]

A side question on your whatsapp connection. Do you use a seperate number for that or are you using your private number for the connection? I read something about that you should be aware that the bot can read any message you send and recieve via whatsapp than. is that correct or am I wrong here?

[brisk sand]

Deploying on Render: Where do I specify the port?

==> Deploying...
==> Setting WEB_CONCURRENCY=1 by default, based on available CPUs in the instance
==> No open ports detected, continuing to scan...
==> Docs on specifying a port: https://render.com/docs/web-services#port-binding
2026-02-13T09:56:09.704Z [canvas] host mounted at http://127.0.0.1:18789/__openclaw__/canvas/ (root /data/.openclaw/canvas)
2026-02-13T09:56:09.795Z [heartbeat] started
2026-02-13T09:56:09.800Z [gateway] agent model: anthropic/claude-opus-4-6
2026-02-13T09:56:09.801Z [gateway] listening on ws://127.0.0.1:18789 (PID 41)
2026-02-13T09:56:09.802Z [gateway] listening on ws://[::1]:18789
2026-02-13T09:56:09.803Z [gateway] log file: /tmp/openclaw/openclaw-2026-02-13.log
2026-02-13T09:56:09.808Z [browser/service] Browser control service ready (profiles=2)
==> Continuing to scan for open port 8080 (from PORT environment variable)...
Deploy cancelled
==> Continuing to scan for open port 8080 (from PORT environment variable)...

[livid vigil]

ya i use personal number , but have told bot that not to reply to any messaages except when its tagged like @@ openclaw

[urban harness]

Hello

can someone suggest an AI model provider which has a lot of models that I can use for my api requests ?

[obsidian jolt]

hey, how do i make my openclaw browse they tell me to use brave but it doesn't have a free api

[burnt plover]

macOS OpenClaw app (v2026.2.12) UI issue:

• Menu bar icon and app window never appear (only runs in background).
• Canvas snapshot always returns “A2UI assets not found”.
• Node connects, permissions (Screen Recording/Accessibility) granted.
• Reinstall tried: DMG, ZIP, Homebrew Cask, downgrade to v2026.2.9, reset app support/caches, reboot/logout.
• App bundle seems missing A2UI assets (find /Applications/OpenClaw.app/Contents -name "a2ui" => no output).
• macOS 26.2, Apple Silicon (Mac16,10).
  Please advise/fix packaging.

[unborn remnant]

let us know how it went...

[vague willow]

Any workaround for "Error: Session file path must be within sessions directory" until they patch it?

[unborn remnant]

I read about venice.ai on the Openclaw onboarding to be there #1 pick. Did you try that?

[past eagle]

Where do I get skills from? I only see the basic ones

[dim viper]

Hi everyone, I’d appreciate some guidance on setting up a workflow with OpenClaw.

Goal:
We want to use OpenClaw as an internal AI assistant connected with Slack (CFO channel), Google Drive, Trello, and possibly internal documents. The idea is:

• Store data from Slack channels, Google Drive folders, and other connected tools

• Then ask the bot questions like:

  • “What’s happening in the CFO Slack channel?”
  • “Create a file in Google Drive for this task”
  • “Summarize updates from Trello”

• Also allow the bot to perform actions (create files, update cards, etc.)

Questions:

1. What is the best architecture for this? (Cloud vs self-hosted)
2. How should we structure data ingestion from Slack/Drive/Trello so the bot can answer based on stored data?
3. Do we need a separate database/vector store, or does OpenClaw handle this internally?
4. Any recommended setup steps or best practices for production use?

Would really appreciate guidance from anyone who has implemented a similar multi-integration assistant setup. Thanks in advance!

[unborn remnant]

I payed 5$ to get 1000 searches..but I don't know if it uses that or if it get's the connection via my connected models..

[dim viper]

I am curios, Is it achieveable with openclaw ? As I am new to the openclaw, Do you suggest me to buy plan from the openclaw or buy VPS and configure it by myself ?

[past eagle]

Huh I’m so confused lol I’m on claw hub but what do I get here

Any you recommend?

[dim viper]

@north veldt Can u please guide if this is achievable, the scenario i have explained to you? Does claw allow us to connect all the platforms we possibly think of?

and can be controlled by the openclaw bot ?

[unborn remnant]

https://clawhub.ai is that trustworthy? There are already so many scammers outthere? I have a hart time finding out who is right and who is not...

[dim viper]

@north veldt I am glad to hear that, Soon I will start on this project, Its for one of my client.

@north veldt What do you mean by skill here ?

Ohh, Got itt 🙂

[obsidian jolt]

did you use the brave api?

[unborn remnant]

yes and its running

[past eagle]

which browser skill do I get?

oh and desktop skill

so lets say I want my bot to communicate with me via discord and I just want it to use chrome and go to a website and itneract there. what do I need?

[dim viper]

@north veldt I have a quick question regarding how ClawBot works inside Slack.

Where exactly does the ClawBot live, and from where can we message it?
Is there a way to create a Slack app/bot and connect it with OpenClaw, so that the bot appears inside our Slack workspace or channels?

If yes, can we then send commands directly in Slack like:
“Please create a file for me inside Google Drive,”
and the connected OpenClaw bot will perform that action through the integrations?

Just trying to understand the correct architecture and how the Slack bot interacts with OpenClaw.

[obsidian jolt]

do you recommend me to use it too?
what if we go above the $5 free trial?

that's wild
who knows what that devil openclaw might be doing all night

[tawny thistle]

Anyone facing this error randomly when talking to an agent: 'LLM request rejected: messages.3.content.0.thinking.signature: Field required'?

[past eagle]

I’m so confused on which skills to pick

[tawny thistle]

I rolled back to the latest 2026.2.13 version (without the pr) and no longer had the earlier issue where 4.6 opus wont work at all, but I got this error again, though it happens sporadically

[regal matrix]

Idk where to post this but... Looking for setup advice from power users

Running OpenClaw on a standalone VM with Opus 4.6, Telegram + Discord. Got about 30 skills, structured workspace (memory files, daily logs, task board), heartbeat crons, morning briefings, email automation. The basics work.

But honestly I'm struggling with a bunch of stuff and need help:

Coding projects feel janky. I can spawn Claude Code or Codex as sub-processes but it doesn't feel like a real integration. For anything beyond small fixes it's clunky. I mainly just use Claude Code on its own. What's the best way to have OpenClaw actually orchestrate dev work properly?

Memory is hit or miss. Got MEMORY.md, daily logs, memory_search, the whole thing. Agent still forgets stuff or fails to connect dots. Context compaction loses important things. What memory strategies actually work for you guys?

Sub-agents are unreliable. They go silent, don't report back, or produce half-baked output. How do you handle complex multi-step tasks without babysitting?

Zero visibility. No dashboard, no way to see task status or what the agent is doing at a glance. Anyone built something for this?

Skills - which ones are actually worth it? Anything on ClawHub that's a must-have?

I feel like I'm at maybe 30% of what this thing could do. Would really appreciate hearing how you guys have your setups dialed in or what I'm probably doing wrong.🙏

[past eagle]

I want autonomy with Claude I tell it to do something on my desktop or browser and it just does it

But which ones do I get I see 10 different ones that all look the same

[dim viper]

@north veldt can u please guide me on one of my messages above 🙂 so that i can breateeee

[tawny thistle]

well, I did not use that PR anymore since it causes my opus 4.6 to not work. But that PR is a fix to this thinking sig issue, so is there any other way to fix this thinking sig issue?

[regal matrix]

Any advice or good guides I can check to make my setup much better?

[hardy sable]

2 questions

1. Anyone have any advice on how to keep the cost down for heavy usage? I just started with a “stock” build using opus and sonnet. I’ve switched everything to sonnet which seems to help but would love a subscription like minimax does for $10-20 a month
2. Maybe this won’t be a thing depending on the answer to my first question but how do you keep your input tokens down? I keep running into the 30,000 token per minute limit for Claude and my logs are showing my baseline input is about 18k tokens.

[past eagle]

I cant get my bot to talk back to me in discord and its stopping itsself so quickly

[bright sky]

can some guide me how to Config Ollama API with OpenClaw in PowerShell, what are the commands required

?

how can do that? can you please guide me.

yeah i have done that

[past eagle]

yea I got that fixed but why does my bot like stop the action? why dosent it like keep going??

[regal matrix]

Is there any useful like poweuser guides or channels or YouTubers or any content I can learn and consume ?

[bright sky]

have to command this on PowerShell?

its says “missing path”.

i command “ openclaw config set”.

[nocturne zephyr]

So...multi-agent is broken right now. 🙁 At least communicating with more than one agent is. It doesn't use the right agent's path, and gets really confused. 🙁

[past eagle]

am I using my bot wrong

Im trying to tell it to find me leads but it just wont

[simple birch]

why openclaw not responding? ollama respond very fast

⠴ streaming • 49m 36s | connected agent main | session main
(openclaw-tui) | ollama/glm-4.7-flash:q4_K_M | reasoning | tokens 93k/200k
(47%)

[carmine vine]

What is the best general model to use on Ollama right now?

[hasty viper]

Hello,

Is there a way that I can get openclaw to check if the model is having issues to automatically change back?

It keeps getting stuck in openrouter/auto even after deleting it keeps coming back

[remote tartan]

Is it advisable to use Docker and LanceDB?

[remote ginkgo]

Hi everyone 🙂 I am using the free glm4.7 from openrouter for my openclaw, which models are you using? I feel glm4.7 is not following the heartbeat instructions as I requested maybe there is a better model for agents.

I tried gptoss120*b from ollama cloud as well

[high parrot]

gemini 3 flash

[remote ginkgo]

any free alternative? the I/O for G3 is $0.50/$3

[remote tartan]

someone use docker and lancedb in openclaw? or is better dont use?

[empty gazelle]

you gonna hit the free tier limits way quicker than you expect

i tend to use glm 4.7 flash, gemini 3 flash

minimax the new model is quite good for coding and apparently good at agentic tasks try taht aswell ie the new one

[upper python]

what do I do when I hatch and the bot just says (no output)?

[remote ginkgo]

kill tui and open again, the reason may appear.

[upper python]

how do I do that? I have no idea of linux

[remote ginkgo]

press control C multiple times, then run openclaw tui again

ah.. are you using the browser ?

[upper python]

no I am in powershell

[remote ginkgo]

try it.. idk powershell, maybe it works

[upper python]

I got an llm error

I tried using gemini 2.5 flash

just use a different version?

[remote ginkgo]

using openrouter/z-ai/glm-4.5-air:free

[past eagle]

Bro my bot is stupid or something

He acts so dumb and I don’t know why 😭

[thorn dome]

Hi! Dumb question. Please How can I install the new update? Just say it in Telegram? And does this also update my portal on Hostinger?

[young nebula]

anyone have a good way to get your bot to be able to use google drive? i tried with API stuff and it appears google blocks it from accessing the drive

[remote ginkgo]

keep tweaking it, configure the md files, change models to a better one and try again.

[remote ginkgo]

just ask it to update itself

[remote ginkgo]

gogcli?

[upper python]

LLM error: <!DOCTYPE html>

<html lang=en>
<meta charset=utf-8>
<meta name=viewport content="initial-scale=1, minimum-scale=1, width=device-width">
<title>Error 403 (Forbidden)!!1</title>
<style>
{margin:0;padding:0}html,code{font:15px/22px
arial,sans-serif}html{background:#fff;color:#222;padding:15px}body{margin:7% auto
0;max-width:390px;min-height:180px;padding:30px 0 15px} >
body{background:url(//www.google.com/images/errors/robot.png) 100% 5px no-repeat;padding-right:205px}p{margin:11px 0
22px;overflow:hidden}ins{color:#777;text-decoration:none}a img{border:0}@media screen and
(max-width:772px){body{background:none;margin-top:0;max-width:none;padding-right:0}}#logo{background:url(//www.google.
com/images/branding/googlelogo/1x/googlelogo_color_150x54dp.png) no-repeat;margin-left:-5px}@media only screen and
(min-resolution:192dpi){#logo{background:url(//www.google.com/images/branding/googlelogo/2x/googlelogo_color_150x54dp.
png) no-repeat 0% 0%/100%
100%;-moz-border-image:url(//www.google.com/images/branding/googlelogo/2x/googlelogo_color_150x54dp.png) 0}}@media
only screen and
(-webkit-min-device-pixel-ratio:2){#logo{background:url(//www.google.com/images/branding/googlelogo/2x/googlelogo_colo
r_150x54dp.png) no-repeat;-webkit-background-size:100% 100%}}#logo{display:inline-block;height:54px;width:150px}
</style>
<a href=//www.google.com/><span id=logo aria-label=Google></span></a>
<p><b>403.</b> <ins>That’s an error.</ins>
<p>Your client does not have permission to get URL
<code>/v1beta/models/gemini-2.5-flash-lite-preview-09-2025:streamGenerateContent</code> from this server. <ins>That’s
all we know.</ins>

what does that mean?

[high parrot]

did you put an api key even?

[upper python]

yeah I entered my gemini key

[young nebula]

Ya but when I run the browser it says like client deleted or something

[remote ginkgo]

it is not* so ease to make it run.. I am using a lot of env vars to make it work.

GOG_ACCOUNT/GOG_KEYRING_PASSWORD/GOG_KEYRING_BACKEND=file ...

then it disconnects every day

[upper python]

minimax also shows the llm error what is going on?

[queen flame]

can somebody here please tell me how does a service like simpleclaw.com , provisions and installs openclaw programmatically. I'm very close on building something like that but installation is the challenge.

[empty gazelle]

my guess is mostly likely docker containers,or using a vps reseller api

[queen flame]

I'm using hetzner api for provisioning the vps. The issue with docker here is that every instance then needs a docker installation which takes time. I want to somehow make it quicker

[remote tartan]

if i am using a vps, i need docker?

[tepid ferry]

you can aalso ask minimax on thee dev programm for free voucher, just nneeed to sescribe for what and why you want to use them. can get up to 100 Dollars. i also got 50 yesterday for OC! anyonee need a link?

[remote tartan]

So you don't recommend putting Docker on the VPS server?

[queen flame]

the flow is dead simple

• user pays
• create an instance for the user
• install openclaw on the vps

the installation is the trickiest part.

I did add the cloud init config , but for some reason that is not being run during the creation

syntax was correct

okay , let me check for it

[young nebula]

i get this when i try to connect to set it up
Sign in with Google
Access blocked: Authorization Error

[e-mail]
The OAuth client was deleted.
If you are a developer of this app, see error details.
Error 401: deleted_client

[empty gazelle]

Instalation is simple
Automating configuration needs more knowhow s bit more than most projects

[young nebula]

ya i keep doing that, its almost like google is blocking it somehow. when I go back it looks like it is accurate

[queen flame]

the installation done with the wizard is simple and sets up things properly , but since this supposed to be carried out programmatically we need to install it in a non-interactive way , which then does not enable many of the things by default.

[ornate bison]

I'm struggling with exec approvals with a node on a remote system. I want my agent to execute some ssh commands on the node. It requests the approval, but the approval fails immediately and I never see a prompt. I'm not sure what the settings should be in openclaw.json or exec-approvals.json. It seems that the node isn't passing approvals to the Control UI (which is open and connected) and I don't know how to do that. I don't want to run the macOS app.

[empty gazelle]

you can just install it from source ? rest is relatively simple to automate once you know how how the config.json is and where teh auth details should be

the main tricky part is actually automating the auth to oauth providers

gl with it,its a good market but is tricky

[queen flame]

what do you mean by source ?

[flint onyx]

can someone tell me why venice api is not working on openclaw?

[queen flame]

also If I just use a template for setting up the fs for openclaw like openclaw.json, identity , there can be a possibility of the cli getting updates and the template that I use might break.

[empty gazelle]

well yes openclaw is generally backwards compatable but thats the maintainence part of such a proj which is ever evolving

[flint onyx]

i really need help regarding venice api

[queen flame]

wouldn't installing it from the source and running it would open the wizard mode of installation ?

i'm just assuming this possibility , have you tried installing it like that ? @north veldt

[flint onyx]

i just dont get any response when i set the default model to venice. everything works proprly when i changed the default model to openai or anthropic

[queen flame]

yeah , but installing it through cli overall installs gateway , plugins and some other deps as well , which I would miss if I try to install everything manually

[lusty knot]

is it normal for gemini 3 pro to sometimes go berserk

alright thanks

[frigid ravine]

Good day - is it normal for all the conversations in telegram to start with "the user is asking ... and proceeds to give an answer. infact it is like it is showing me its thinking which i do not need to see. what could be causing this and how to stop it?

[lusty knot]

i just told it to stop showing me its thinking and it did lol

[frigid ravine]

haha didnt think of that - i thought it was due to using a free model. Tried out kimi and it sucked up all my tokens - so had to wipe the slate clean and try again - it is odd behaviour that i had not seen before thought there is a setting somewhere

[young nebula]

so it says connected and then this is what my bot says when he tries to use it. any thoughts?

Same unauthorized_client error — even with a fresh project, Google won't let the OAuth client use the Drive API from a headless environment.

This is a Google API design limitation, not a setup issue. The OAuth client can authorize in a browser but can't exchange tokens from a server.

yeah thats my thought, i guess i just dont know how others got around it. not being able to have my bot save stuff where I can also access it really limits the usefulness

[flint onyx]

@north veldt will you be able to help?

[sharp marlin]

Hey, people
Is heartbeat work with main session? if i set target: none, will it work in diffirent? tonight agent eat 33M tokens and it blaiming heartbeats being in main session, but we not sure

Im not set it to none yet
thx

[past eagle]

im trying to self host llama 3 but I keep getting this error Agent failed before reply: Model context window too small (4096 tokens). Minimum is 16000.
Logs: openclaw logs --follow

so how do I make it handle more?

I just said hi 😭

[stiff tusk]

14:22:23 error [tools] message failed: Local media path is not under an allowed directory:
any ideas how to deal with this?

[unborn remnant]

I used 15 of my 1000 Search requests.. 5$ is plenty so fare. I am more struggeling to get the bot running and get rid of the context overflow issues

[green sundial]

How do I configure the base URL during the onboarding process so I can get OpenClaw to use a local ollama AI model?

[livid vigil]

guys how to set google project tag ?
i tried :

  --tag-key=environment \
  --tag-value=Development

when i do : gcloud config set project openclaw-test-487311
i get error

Project 'openclaw-test-487311' lacks an 'environment' tag. Please create or add a tag with key 'environment' and a value like 'Production', 'Development', 'Test', or 'Staging'. Add an 'environment' tag using `gcloud resource-manager tags bindings create`. See https://cloud.google.com/resource-manager/docs/creating-managing-projects#designate_project_environments_with_tags for details.

[green sundial]

can somebody please help me? I've been struggling with this for hours and it is driving me insane. I've tried copy pasting the information into the openclaw.json file but it always ends with syntax errors

[past eagle]

how do I fix this?

⚠️ Agent failed before reply: No API key found for provider "custom-127-0-0-1-11434". Auth store: C:\Users\voids.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\voids.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir.
Logs: openclaw logs --follow

[past eagle]

just set a custom provider?

then leave the api blank

base url the same and put your model name

[green sundial]

I have no clue how to do that, I'm not much of a computer person but want to learn

[past eagle]

did you go throught begining config?

[green sundial]

yes, I set the model to ollama/qwen2.5:14b and verified it is up and running at http://localhost:11434

but when I try to get OpenClaw to use the model I get this error message:
[diagnostic] lane task error: lane=main durationMs=5 error="Error: Unknown model: ollama/qwen2.5:14b"
14:54:04 [diagnostic] lane task error: lane=session:agent:main:main durationMs=13 error="Error: Unknown model: ollama/qwen2.5:14b"
09:54:04 Embedded agent failed before reply: Unknown model: ollama/qwen2.5:14b

[bold meteor]

any idea why i am getting no or no reply as reponse , while losing token , i am using openclaw tui

any idea

https://freeimage.host/i/qH1QgrQ screen shot

is it something to do it with approvals

[green sundial]

I've been told I need to set the base url to "http://127.0.0.1:11434/v1", but I have no idea how to do that

[sinful field]

ive been having a issue in my private discord where my bot wont remember message to massge at all// anyone know how to fix?

[ornate bison]

If I have a headless node and that node triggers an exec approval - shouldn't the prompt show up in the Control UI? It just keeps failing immediately.

[past eagle]

5:07:37 [diagnostic] lane task error: lane=main durationMs=16 error="Error: No API key found for provider "custom-127-0-0-1-11434". Auth store: C:\Users\voids.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\voids.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir."
15:07:37 [diagnostic] lane task error: lane=session:agent:main:main durationMs=19 error="Error: No API key found for provider "custom-127-0-0-1-11434". Auth store: C:\Users\voids.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\voids.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir."
09:07:37 Embedded agent failed before reply: No API key found for provider "custom-127-0-0-1-11434". Auth store: C:\Users\voids.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\voids.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir.

but if I add an api key it dosent send a response at all

[bold meteor]

https://discord.com/channels/1456350064065904867/1471881871490023465

[random mauve]

sometimes in scenarios like that its easier to just run codex or claude in your openclaw dir, and tell it to go update the config to do what you want it to do.

[bold meteor]

hi , got any idea , kepp getting no or no reply trying openclaw tui

[green sundial]

I'm terrible with computers. I'm trying to learn how to use OpenClaw. I don't know how to use codex or claude

[random mauve]

no idea, sorry. I think in the past that's sometimes been a symptom of some hidden error. One thing you can do that works well for me to debug is again use codex or claude running in your openclaw dir, and just tell it to debug the issue. it should be able to dive through your gateway and session logs to figure something out

[random mauve]

how did you install openclaw?

[green sundial]

using npm

[severe nexus]

Hello guys has anyone figured out a way to have a browser with Openclaw that can login and solve captcha ?

i have heard of https://github.com/browserbase/skills?tab=readme-ov-file

but i am unable to figure out how to install it, seems to not be working

[heavy obsidian]

Hello!
Very new to everything ai related basically.
My plan is to run my own local model (deepseek 8B) though ollama. Running on windows 11.
I managed to get deepseek working though ollama, using gemini pro Ive made working openclaw.json that turn on, and though the browser UI I can see that deepseek is the default model.
However, when I use chat option, I get blank responses.
Its been 6 hours trying to fix it with AI, no effect.
Maybe someone running local model can send me an example of json file so I would understand what am I doing wrong?
Thanks

[past eagle]

what does this mean and hgow do I fix it?

400 registry.ollama.ai/library/llama3-16k:latest does not support tools

[past eagle]

Im having the same issue

[west kite]

I'd like to setup openclaw as a "chatbot who remember" and totally local via ollama, no access to stuff (unless I explicitly pass to the agent), no multiagents for the time being, no cloud and no access to other channels.
As secure as possible.
On Windows 10

I know it defeats most of the points of Openclaw but reading into it I'm not sure how to do it but basically I'm fascinated of the AI self-modifiable soul/identity/user and want to see how it evolves in local

[prime spindle]

Good afternoon, guys. Just a quick question. Has it ever happened to you that some messages from the lobster app on Telegram don't reach you, but you can see them when you log into the dashboard?

Yes. Everything seems to be fine.

[lusty knot]

agent used 1.2m/1m tokens
stopped working (gateway closed)
now im restarting gateway and it says it started
but when i check its stopped

Error: gateway closed (1006 abnormal closure (no close frame)): no close reason

whatever this is

[orchid wagon]

Hi, I am trying to use "antfarm" in order to have a multi-agent development setup. It is not really working well and the first agent is "stuck" while antfarm thinks it is running. I asked openclaw, whether he can talk to that agent, but apparently it is not possible for main to talk to another local agent, is this true?

[tall pier]

https://tenor.com/view/am-iright-anchorman-ronburgandy-willferrell-gif-5060412

agents and subagents

[past blaze]

Is it possible to the the A2UI canvas in the mac app working with a remote agent? I've got the rest of it working but the agent is having trouble getting anything to show up on the canvas.

[tall pier]

Whats the trick with getting sub agents -> run locally, ollama random models like 8-12b param models to correctly use tools like telegram, setting cron jobs but they just hose the cron job, it completes successfully but they dont send telegram messages. can switch model to the sub agent, and speak directly to the subagent ollama model via telegram. Something with the isolated is killing their ability to use the tool message

[orchid wagon]

I simpy ran the feature-dev "workflow" and it stuck already quite already. Antflow says it is "running" saying "feature-dev/planner pending"
In the Activity Logs it says:
"
Activity
09:53 AM Run started
10:00 AM planner Claimed step
10:15 AM Step timed out — Reset to pending (retry 1)
10:20 AM planner Claimed step
10:35 AM Step timed out — Retries exhausted — step failed
10:35 AM Step failed: Agent abandoned step without completing
10:35 AM Run failed ❌
"
No more infos on why it stopped

[tall pier]

yea its enabled, and setup and can /model ollama then talk via telegram, but using cron jobs its just not working. i can switch to a stronger model like deepseek for the cron job and it will work

[orchid wagon]

I have no clue why it is not working. And apparently main cannot "talk" to the feature-dev/planner agent to make him move or so.

[tall pier]

Using llama3.1:8b, and mistral-nemo 12b on ollama

[orchid wagon]

I saw yesterday a guy that used openclaw with multi-agents (without antfarm) and the agents could communicate among each other somehow

[green sundial]

here is my configuration file: {
"messages": {
"ackReactionScope": "group-mentions"
},
"agents": {
"defaults": {
"maxConcurrent": 4,
"subagents": {
"maxConcurrent": 8
},
"compaction": {
"mode": "safeguard"
},
"workspace": "C:\Users\tomclancy.openclaw\workspace",
"model": {
"primary": "ollama/qwen2.5:14b"
},
"models": {
"ollama/qwen2.5:14b": {
"streaming": false
}
}
}
},
"gateway": {
"mode": "local",
"auth": {
"mode": "token",
"token": "b399a2446e5fa726ddbcbae90172bba333f07277e8a91254"
},
"port": 18789,
"bind": "loopback",
"tailscale": {
"mode": "off",
"resetOnExit": false
},
"nodes": {
"denyCommands": [
"camera.snap",
"camera.clip",
"screen.record",
"calendar.add",
"contacts.add",
"reminders.add"
]
}
},
"wizard": {
"lastRunAt": "2026-02-13T14:45:29.095Z",
"lastRunVersion": "2026.2.12",
"lastRunCommand": "onboard",
"lastRunMode": "local"
},
"meta": {
"lastTouchedVersion": "2026.2.12",
"lastTouchedAt": "2026-02-13T14:45:29.111Z"
},
"models": {
"providers": {
"ollama": {
"baseUrl": "http://127.0.0.1:11434/v1",
"apiKey": "ollama-local",
"api": "openai-completions",
"models": [
{
"id": "qwen2.5:14b",
"name": "Qwen2.5 14B (Ollama)",
"reasoning": false,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 64000,
"maxTokens": 81920
}
]
}
}
}
}

there are no syntax errors but the agent doesn't respond when I try to chat with it

is there something wrong in it?

[young nebula]

Any suggestion how to stop my bot from sending me like 10 messages when I ask it to do something? It just started happening today like explaining every little thing

[green sundial]

it still doesn't seem to be responding

[orchid wagon]

Thank you, for your answer! I will have a look at these orchestrator agents. Maybe this is the solution to nudge the other agents in case that they are stuck

[green sundial]

can somebody please help me? I'm at my wit's end

[tall pier]

the first few days of oc are rough

[astral loom]

I'm suddenly getting 421 Misdirected Requests across multiple providers whenver I write to my bot

Yes I even reverted to a last known good openclaw.json

[ornate bison]

How do exec approval prompts work for a headless node? Should they forward the approval to chat (discord in my case) or the Control UI? Everytime I trigger a prompt it just gets auto-denied without forwarding a prompt.

[livid vigil]

any body can help with this issue of gmail integration ?

[ornate bison]

Thanks - I don't have approveByDefault in my configs. I haven't made a lot of changes just mostly defaults. If I run a CLI command to trigger an approval - something like openclaw nodes run --node my-node --ask always /usr/bin/true - I successfully get a prompt. However, if chat tries to run something that needs a commmand, it's insta-denied with ⇄ res ✗ node.invoke 41ms errorCode=UNAVAILABLE errorMessage=SYSTEM_RUN_DENIED: approval required conn=9aba077c…8e13 id=9a5c02f1…3c62

[rancid fjord]

woke up this morning to everything broken.
"disconnected (1008): unauthorized: gateway token mismatch (open the dashboard URL and paste the token in Control UI settings)"

i have updated to the latest version.
😭

its already in that dashboard, that's wehre i see the error.

[livid vigil]

ya but these commands ? proper command
gcloud resource-manager tags bindings create
give me error

(gcloud.resource-manager.tags.bindings.create) argument --parent --tag-value: Must be specified.
Usage: gcloud resource-manager tags bindings create --parent=PARENT --tag-value=TAG_VALUE [optional flags]
  optional flags may be  --async | --help | --location

what is this parent
I tried : gcloud resource-manager tags bindings create --parent=PARENT --tag-value=environment=Production it also not worked

[green sundial]

I gave up trying to use a local model and set it to ChatGPT 5.2. Now when I try to use it I get this message: 16:36:54 [agent/embedded] unsupported thinking level for openai/gpt-5.2-chat-latest; retrying with medium

[livid vigil]

open you openclaw.config file look for gateway section then from there copy token . it may ask for proval too of new device then do

openclaw devices list
openclaw devices approve <request id>

i tried with : gcloud resource-manager tags bindings create --parent=environment --tag-value=Production
but no success
can u pls help

[rancid fjord]

ok i was able to open this http://localhost:xxxx/?token=SOME_LONG_TOKEN and it now says connected. thanks!.

ugh ... now my telegram bot is giving me an error. i switched from anthropic to gemini to test something last night, and this morning i switched it back to claude and i get this error "HTTP 404: not_found_error: model: claude-3-5-haiku-latest (request_id: req_011CY6QC6VjLXp95DMMCH9Fr)"

[ornate bison]

A sanity check - I do not need the macOS app running for approvals, right? If I have approvals forwarding to discord, have a node running, and chat triggers an approval on the node, it should forward to discord, correct?

[rancid fjord]

hmmm found this in logs. "The model 'claude-3-5-haiku-latest' is deprecated and will reach end-of-life on February 19th, 2026" ...

[thin loom]

[OK] Node.js v24.13.1 found
[*] Installing OpenClaw (openclaw@latest)...
node.exe : npm error code ENOENT
At line:1 char:1
+ & "C:\Program Files\nodejs/node.exe" "C:\Program Files\nodejs/node_mo ...
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    + CategoryInfo          : NotSpecified: (npm error code ENOENT:String) [], RemoteException
    + FullyQualifiedErrorId : NativeCommandError```

I keep getting that error anyone know how to fix it?

[green sundial]

I have to say I really hate how it is basically impossible to get OpenClaw to work and learn how to use it as a non computer programmer. All the instructions and YouTube videos seem to assume I'm already a computer scientist. I want to learn but it is basically impossible at this point. I've used both ChatGPT and Gemini to try and help me to no avail

it seems like a cool program but it is basically impossible for people who aren't computer experts to use

[fierce gazelle]

The file path doesn’t exist

[fierce gazelle]

In a fresh powershell Run node -v then npm -v to see if npm actually exist and it’s on path

[thin loom]

v24.13.1
PS C:\Users\Sethl> npm -v
11.8.0```

It does

[green sundial]

I spent hours trying to get it to work with my local ollama model with no luck. When I threw in the towel and tried to set it up to run an OpenAI model with an API key that didn't work despite following every direction and Chatbot I could find

[sonic sandal]

Thanks a lot

[fierce gazelle]

Nice, node and npm are fine. That ENOENT usually means npm can’t find a path it expects on Windows, not that OpenClaw is broken. Quick check, what does this print?

npm config get prefix

[rose thunder]

Do u guys know if soul.md is still a thing to configure in newest ver ?

[thin loom]

C:\Users\Sethl\AppData\Roaming\npm

[fierce gazelle]

Do npm i -g openclaw

[soft pike]

I just set up on my macbook, and the agent never does what I want. For example, "Give me a digest of my unread emails." (I have himalaya setup) It just returns a raw list of email headers. I had to get specific with each step, then got to a point where it STILL hasn't given me a digest.

[zinc mica]

Has anyone else encountered your lead agent just not responding in telegram? Checked my mission control dashboard but everything seems systematically in order. Telegram is set up properly and everything.

[rose thunder]

Odd, these instructions in the docs are vague...

[soft pike]

Are you saying that the skill isn't robust?

Also I was using openrouter, and maybe the models it's picking aren't intellegent enough?

[fierce gazelle]

add a job rule that says “when user says digest, you must read the email bodies (or snippets) then summarize into Action Required / FYI with next steps.” Or create a dedicated mail_digest skill that does: list unread → read bodies → summarize.

[royal fog]

Is there an easier way to debug openclaw.json config issues than runningopenclaw then trying to parse errors like:

Invalid config at ~/.openclaw/openclaw.json:\n- agents.defaults: Unrecognized key: "list"
- agents: Unrecognized keys: "tools", "messages", "commands", "session", "channels", "discovery", "skills", "allow", "entries"

?
Are there docs specifically about configs cause openclaw agents claim https://docs.openclaw.ai/ is nearly unparsable

[rose thunder]

in my case i created 5 agents in orchestrated mode. Confused at this point how do i configure their roles.

[lofty pendant]

Having issues pairing telegram on a new instance.

🦞 OpenClaw 2026.2.12 (f9e444d) — Your task has been queued; your dignity has been deprecated.
[openclaw] Failed to start CLI: Error: No pending pairing request found for code: XXXXXXXX
    at Command.<anonymous> (file:///usr/lib/node_modules/openclaw/dist/pairing-cli-REnkGH30.js:104:24)
    at async Command.parseAsync (/usr/lib/node_modules/openclaw/node_modules/commander/lib/command.js:1122:5)
    at async runCli (file:///usr/lib/node_modules/openclaw/dist/run-main-w6FjSYWf.js:149:2)

[bronze stirrup]

I constantly get 429 rate limited, I switched to open ai chat gpt but what prompt can I use to switch the default model? Opus is just rate limiting me when I ask it for help

[fierce gazelle]

This is how I’ve fixed this exact “Telegram goes silent” issue over and over:

First I restart the gateway: openclaw gateway restart

Then I generate a fresh tokenized dashboard link: openclaw dashboard --no-open
Old links/tokens often die after a restart, so this rules that out fast.

Before I test Telegram again, I run logs in another window: openclaw logs --follow
Then I send a message and watch what the gateway actually receives. You’ll usually see the real error immediately instead of getting silence in chat.

[rotund hinge]

hello guys, im completely new and now trying to instrall the OpenClaw. This error pops up, node installed at lastest version. Mac mini running on Mac OS monterrey 12.7.6. anyone can please me please?

npm -v
which node
v24.13.1
11.8.0
/usr/local/bin/node
atlas@Mac-mini-uzivatela-Atlas ~ % curl -fsSL https://openclaw.ai/install.sh | bash
╭─────────────────────────────────────────────────────────────────────────╮
│ │
│ 🦞 OpenClaw Installer │
│ Less clicking, more shipping, fewer "where did that file go" moments. │
│ modern installer mode │
│ │
╰─────────────────────────────────────────────────────────────────────────╯

✓ gum bootstrapped (temp, verified, v0.17.0)
✓ Detected: macos

Install plan

OS macos
Install method npm
Requested version latest

[1/3] Preparing environment

INFO Homebrew not found, installing
ERROR Installing Homebrew failed — re-run with --verbose for details
Warning: Running in non-interactive mode because stdin is not a TTY.
==> Checking for sudo access (which may request your password)...
Need sudo access on macOS (e.g. the user atlas needs to be an Administrator)!
atlas@Mac-mini-uzivatela-Atlas ~ %

[warm raven]

Having this issue: OpenClaw connects to Chrome via Relay extension in dev mode and shows “ON”, but as soon as I open a page or click anything on a page, the extension switches off and the openclaw connection with browser is lost.
This used to work before. Now it happens even after a fresh reinstall and reconnect.
Windows + latest Chrome.
Anyone else experiencing this?

[fierce gazelle]

That error means your openclaw.json is using the wrong schema shape, not that debugging is hard, it’s telling you those keys are in the wrong place.

“Unrecognized key: list / tools / messages / commands / skills …” usually happens when you paste an agent block from an old example or from a skill file directly into the root config.

To continue making progress: generate a fresh baseline config and diff it.

Run openclaw init --force (or move your current openclaw.json aside and re-init), then compare the new file with yours and copy your values back into the correct sections only.

openclaw.json is runtime config, not agent/skill definition. Agents, tools, skills, messages belong in their own files, not top-level config.

For debugging, also use openclaw config validate and openclaw logs --follow while starting up. It will point to the exact path that’s wrong instead of guessing from the big error blob.

[royal fog]

Thanks so much for the informative reply! I guess part of my issue (other than not being particularly smart) is that I'm running via vps via docker which seems to add an extra-extra layer of complexity, especially when dealing with gateway stuff (like the openclaw logs --follow stuff from what I understand).
I'll look into all this though. Thanks again ❤️

[warped kettle]

I can't get rid of this error

Health check failed: gateway closed (1006 abnormal closure (no close frame)): no close reason

[rose thunder]

can someone point me to a good doc describing on how to configure your agents?

[warped kettle]

Ok, let me try

[tough lagoon]

Just get it running and talk to it and it can do it itself

[tough lagoon]

Yes but only tried that manner of browser control this week and assumed that’s how it always worked

Very annoying

[rose thunder]

so enable telegram and that will fill my soul.md, identity.md templates?

[tough lagoon]

Yes basically it will ask you to set it up

[warped kettle]

Didn't work

[rose thunder]

ha, so i have 5 of them and i want them to do specific tasks. Was thinking this was done via soul.md and identity.md

[tough lagoon]

It can edit its soul and other files on the fly anytime so you can start it up fresh and tell it to be whoever you want and anytime in the future tell it to be something else. Or be fun and ask it who it wants to be

Just make one first and then ask it to make some siblings

It will make new workspaces and new soul files for each

[left arch]

how do i fix npm error code 3221225477

[rose thunder]

ha interesting. I will pair telegram than and go this route see what happens. TY!

[scarlet echo]

I'm no longer gettings Whatsapp

[golden marsh]

Hey all when i update, do i need to update the client and the mac app?

theres no way to do it both from the cli is there?

[tranquil marsh]

yo folks, just starting out with openclaw. Any specific way to make the bot truley autonomous. It still needs to be prompted at every step of the feature. It says it will do X and it just stops responding until I prompt it again

[vast badge]

Hi everyone, I'm encountering a 'Gateway Error: Not Initialized' while trying to connect Openclow with the APK on my Android device. Does anyone know how to resolve this or what might be causing it

[runic breach]

guys i dont know why my bot asks for Brave API key

he already has web search functionality right?

[fierce gazelle]

If Docker is adding complexity, skip it for now and install OpenClaw directly on the VPS so you can see what is actually failing. It is much easier to debug when there is no extra container layer in the way.

On a fresh Ubuntu VPS install Node LTS using nvm or NodeSource, then verify with node -v and npm -v. Install OpenClaw globally with npm i -g openclaw. Start the gateway with openclaw gateway start and generate a fresh dashboard link with openclaw dashboard –no-open. While you test, keep openclaw logs –follow running so you see the real error instead of guessing from the UI.

Also keep the gateway running inside tmux or as a systemd service so it does not die when your SSH session closes. Get it stable like this first, then move back to Docker later if you still want it.

[tranquil marsh]

Nope. It does not. We need to give it the functionality either through brave API key or perplexcitiy

[fierce gazelle]

That braveAPI key enables your bot to use the web search functionality

[runic breach]

ah cool fair enough!

[fierce gazelle]

It’s the gateway websocket getting dropped. Quick sanity check is restart the gateway and generate a fresh token link with openclaw gateway restart then openclaw dashboard --no-open, then keep openclaw logs --follow running while you connect so you can see if it’s auth, a crash, or a timeout. If you’re on VPS + Docker, confirm the container is still up and the port is actually exposed because a restart or healthcheck fail often shows up as a 1006. Also if you’re going through a proxy, tunnel, or Cloudflare, try one direct connect first because those can kill websockets and make it look like a gateway issue.

[wet adder]

Has anyone else run into problems with the Claude CLI backend creating its own shadow memory directory and starting to use that as active MEMORY rather than the ~/.openclaw/workspace MEMORY files?

claude-cli spawns the claude binary as a subprocess and that ends up creating a subdirectory that has all the same memory files as openclaw, and we start to get divergence and memory failures.

[green sundial]

how do I use the Nano Banana Pro skill in OpenClaw? I put in my API key but it says Brew is missing

[runic breach]

guys my bot does not check emails as scheduled and do things automatically

It always needs to be prompted

i had email checks set up every 3 minutes

so he can reply to people emailing him

Any idea why? I am a mechanical engineer so i am a noob dev please help

I created his own Gmail, gave him IMAP and Oauth and everything he asked me for

[haughty hedge]

My bot was slow as hell. Kept trying to debug. Ended up stripping away everything 😭😭. Turns out a 27MB sessions file. Made it ssslooow!

[candid imp]

~80% reduction in token usage for agents just by adding this header
-H "Accept: text/markdown"

And for accessing websites not on Cloudflare, you can always use the Browser Rendering /markdown REST API to render a dynamic page or application in a real browser before converting

[halcyon veldt]

Does anyone know why I can not setup my bot with the Claude Setup token created from claude on Terminal? I have a claude pro subscription, but I keep getting HTTP 401: authentication_error: Invalid bearer token (request_id: req_011CY6YS9uBnew3dKD2Hv6Gp) error all the time, my tokens are not expired when setupping.

[candid imp]

Did you pay for cluad pro subscription or pay for API

[halcyon veldt]

Subscription, I checked from the settings and I have pro plan enabled

[candid imp]

Claude Pro ($20/mo): This pays for the Chat Interface (claude.ai) and the Claude CLI tool for personal use.

Anthropic API (Pay-as-you-go): This is what (like OpenClaw) use. It requires a different account, different credits, and a different key.....

the error Invalid bearer token happens because OpenClaw is trying to send your "Browser/CLI Session Token" to the commercial API endpoint, which rejects it immediately because it's the wrong type of key.

[fossil sand]

it used to work. i've used the setup token before but i believe anthropic does not like you doing that so they might be shutting down access

[halcyon veldt]

Oh so I need to create an API key and put some credits in it? Does the Claude Max plan support OAuth token creation for 3rd party services? Because OpenClaw has token and API key both as an option

[fossil sand]

just because openclaw has support doesnt mean anthropic likes it. same with google. if you try to use oauth instead of api key for google, theres a chance they remove your access

[halcyon veldt]

This might also be the case then, I wish they would talk about this publicly if they are limiting the access or something, which one do you use atm? I think using Oauth is more worth than direct API key

[fossil sand]

oauth is definitely way cheaper which is why they only want you to use api and the max plan is to be used either with the claude app or in the browser.

I've set up a bunch of agents. If you wanna go the api based route, i'd recommend openrouter with the auto config

it auto routes your message to appropriate models based on complexity

else if you really wanna use claude or codex, api would be way more stable (but way more expensive) than oauth

keep in mind that even through openrouter, if your queries hit opus, its expensive lol out of the 10$ i loaded, here's the breakdown

Spend
$10.1

Claude Opus 4.6 -9.31
R1 - 0.42
Gemini 3 Pro Preview- 0.192
Others- 0.224

Tokens 8.84M

GPT-5 Nano 5.48M
Claude Opus 4.6 1.84M
Gemini 2.5 Flash Lite 1.03M
Others 487K

[halcyon veldt]

How long did it took for you to spend the whole $10 approximately?

[fossil sand]

2 days lmao

[halcyon veldt]

Did you do some complex tasks etc?

[fossil sand]

i set a limit for the openrouter api key to stop working after a limit of 5$ per day

[fossil sand]

blew it all trying to set up a flow to create trello tasks from whatsapp messages (it still doesnt work)

[green sundial]

is it possible to use the nano banana skill on Windows? Everytime I try it says I am missing brew

[fossil sand]

you have to keep in mind that for most chatting usage and regular usage, opus is like taking a bazooka to a paintball game. so even if youre using opus for non intesnive queries, its going to be very expensive. which is why you should keep opus as a last resort or only switch to it manually unless you want your api bill to be higher than your car payment

im also trying to get my openclaw to build a prompt router for itself where it will automatically switch to opus only for tasks it thinks are too complex for other models

if you want a cheap model for general conversation and basic tasks, use: gemini flash, minimax 2.5 or kimi 2.5 or haiku

use sonnet for creative writing tasks. haiku for basic conversations, gemini for image gen, kimi for coding. opus for deep research or complex tasks

[green sundial]

how do I get the nano banana pro skill to enable? I put in the API key but it keeps telling me that bin:uv is missing

[late grail]

does anyone know how to get claude premium models api's without paying too much for them ?

[brisk tide]

im trying to setup openclaw via Windows powershell, ive installed homebrew using unbuntu and when i get to install skills in onboarding, i receive a message saying:
"◇ Install failed: 1password — brew not installed
Tip: run openclaw doctor to review skills + requirements.
Docs: https://docs.openclaw.ai/skills"

[copper gulch]

19:18:53 [ws] unauthorized conn=b34a383b-b409-451d-bee7-4cba9a31e19e remote=127.0.0.1 client=openclaw-control-ui webchat vdev reason=token_mismatch
19:18:53 [ws] closed before connect conn=b34a383b-b409-451d-bee7-4cba9a31e19e remote=127.0.0.1 fwd=n/a origin=http://127.0.0.1:18789 host=127.0.0.1:18789 ua=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/144.0.0.0 Safari/537.36 Edg/144.0.0.0 code=1008 reason=unauthorized: gateway token mismatch (open the dashboard URL and paste the token in Control UI settings)
19:18:55 [ws] unauthorized conn=11358b3c-e9ca-4679-bd67-4acc79e2e046 remote=127.0.0.1 client=openclaw-control-ui webchat vdev reason=token_mismatch
19:18:55 [ws] closed before connect conn=11358b3c-e9ca-4679-bd67-4acc79e2e046 remote=127.0.0.1 fwd=n/a origin=http://127.0.0.1:18789 host=127.0.0.1:18789 ua=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/144.0.0.0 Safari/537.36 Edg/144.0.0.0 code=1008 reason=unauthorized: gateway token mismatch (open the dashboard URL and paste the token in Control UI settings)
19:18:57 [ws] unauthorized conn=3de2a39e-b807-4643-85cb-46311add6321 remote=127.0.0.1 client=openclaw-control-ui webchat vdev reason=token_mismatch
19:18:57 [ws] closed before connect conn=3de2a39e-b807-4643-85cb-46311add6321 remote=127.0.0.1 fwd=n/a origin=http://127.0.0.1:18789 host=127.0.0.1:18789 ua=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/144.0.0.0 Safari/537.36 Edg/144.0.0.0 code=1008 reason=unauthorized: gateway token mismatch (open the dashboard URL and paste the token in Control UI settings)
19:19:01 [ws] unauthorized conn=d0fecc62-5bc5-4a92-b

i need fix error

disconnected (1008): unauthorized: gateway token mismatch (open the dashboard URL and paste the token in Control UI settings)

[split tusk]

Have this same issue. I found a thread in the help channel where a user pointed out that when pasting the token into the OpenClaw CLI that it can truncate it. Here is the text:

wait i got it, there's a openclaw bug, when the openclaw cli prompts you to paste the claude key, it truncates the last dash bit

[lusty knot]

why am i getting
LLM error: {
"error": {
"code": 429,
"message": "Resource has been exhausted (e.g. check quota).",
"status": "RESOURCE_EXHAUSTED"
}
}
when my usage stats aren't showing that im rate limited

[sturdy mural]

Hey everyone. Is there a CLI command I can run that will fully reset the workspace so I can start with a new chat but also keep my other settings like LLM and whatnot?

[proven axle]

how do i install

[green sundial]

so it turns out a lot of issues with OpenClaw can be fixed by telling the agent to fix them itself. assuming you can get the agent to work

[rigid herald]

anyone willing to share their config for discord channel? I have a bot set up and responding in a server , was able to pair via DM, but i cannot use commands like /new in the server, it says i am not authorized

[pearl solstice]

Anyone have experience setting up bluebubbles on a personal MacBook to go to an openclaw server

I feel like I’m hitting a wall

[azure dagger]

anyone using the lancedb memory plugin?

i cant seem to install it since it doesnt exist in the npmjs registry 🤔

[rotund hinge]

hey i had set up the bot and for some reason its disconnected and i cannot figure out how to get back....

[subtle yacht]

You have hit your ChatGPT usage limit (free plan). Try again in ~8486 min., and i have a open ai apikey, but idk how to change to the api

[candid imp]

first, check if the bot is actually running. open your terminal and type:

openclaw gateway status

[wide musk]

What can I do to my bot do trades in polymarket. It always says it can not do trades.

[whole notch]

Hey, can someone help me? My Claw keeps getting stuck on captchas and refuses to find solutions.🤔

[rotund hinge]

Config invalid
File: ~/.openclaw/openclaw.json
Problem:

• plugins.entries.email: plugin not found: email
• channels.email: unknown channel id: email

Run: openclaw doctor --fix
│
◇
Service: LaunchAgent (loaded)
File logs: /tmp/openclaw/openclaw-2026-02-13.log
Command: /usr/local/bin/node /usr/local/lib/node_modules/openclaw/dist/index.js gateway --port 18789
Service file: ~/Library/LaunchAgents/ai.openclaw.gateway.plist
Service env: OPENCLAW_GATEWAY_PORT=18789

Config (cli): ~/.openclaw/openclaw.json (invalid)
Config issue: plugins.entries.email: plugin not found: email
Config issue: channels.email: unknown channel id: email
Config (service): ~/.openclaw/openclaw.json (invalid)
Service config issue: plugins.entries.email: plugin not found: email
Service config issue: channels.email: unknown channel id: email

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

Runtime: stopped (state spawn scheduled)
RPC probe: failed
RPC target: ws://127.0.0.1:18789
gateway closed (1006 abnormal closure (no close frame)): no close reason
Gateway target: ws://127.0.0.1:18789
Source: cli --url
Config: /Users/atlas/.openclaw/openclaw.json

Service is loaded but not running (likely exited immediately).
File logs: /tmp/openclaw/openclaw-2026-02-13.log
Launchd stdout (if installed): /Users/atlas/.openclaw/logs/gateway.log
Launchd stderr (if installed): /Users/atlas/.openclaw/logs/gateway.err.log

i tried doctor repaid etc but still does not whats going on lol.

[candid imp]

actually have two options depending on whether you want to archive the old chat or nuke the current one...

neither of these commands will touch your openclaw.json, API keys, or model settings.

first one just reset -> this creates a new Session ID. the old chat is saved in your history (in case you need to check it later), but the bot starts with a blank memory slate.

openclaw session new

This will resetcontext window (chat history).

Another option

If you want to keep the same Session ID but force the bot to forget everything that just happened (useful if the bot is in a loop or hallucinating badly), use this:

openclaw session clear

And the last is

If by "workspace" you meant the actual files the agent created (code, text files, etc) in your folder, the CLI command is:

openclaw workspace clean

Note: It will ask for confirmation before deleting files. It does not delete your config

[slow geyser]

Trying to setup sandbox browser in docker container on VPS - Model defaulting to Anthropic even though my openclaw.json only has OpenRouter... anyone know how to fix this?

[candid imp]

your openclaw.json has a section enabling email, but the actual email plugin is not installed on your system.

[wild jacinth]

What is the best way to configure this setting and what prompts will be the best?
"compaction": {
"mode": "safeguard",
"reserveTokensFloor": 20000,
"memoryFlush": {
"enabled": true,
"softThresholdTokens": 30000,
"prompt": "Write any lasting notes to memory/YYYY-MM-DD.md; reply with NO_REPLY if nothing to store.",
"systemPrompt": "Session nearing compaction. Store durable memories now."
}
},

[rotund hinge]

yeah that was my last message to the bot. since i cannot acces to the dashboard, im not really good in the terminal to know whats exactly happening and what exactly i need to do...

[cobalt sphinx]

My agent plans a coding task, but in several cases doesn't start or at least not complete it.
When I ask it remember exacty what to do.

Trying for a few days to nudge or ask to report progress back every 10 minutes, it says yes, and then sleeps...

Many similarities with a teenager 🙂

Any advice for troubleshooting?

I've tried also multiple agent and letting one agent manage the other but with no difference.

Other similar multistep tasks gets completed very quickly (read git issue, triage, resolve, push PR, close issue). Complexity of the tasks are very similar.

Gateway and communication channel always up and stable.

[subtle yacht]

hello, guys how i change from free plan gpt to the apikey token for openai gpt 5.3 codex?

[candid imp]

running in docker, openclaw (and the underlying sandbox images) often prioritizes environment variables (set in docker-compose.yml) over your openclaw.json file.....sandbox browser runs as a separate micro-agent..

[flint onyx]

someone tell me please why venice doesnt work on openclaw

[boreal galleon]

Hey, who knows how I make subagents more assertive about reporting back to the parent? I’m trying to decide whether there’s any way to make subagent message announcements actually fast and predictable, or whether I should just rip the whole mechanism out and replace it with my own scripts.

Here’s the test scenario I’m using, it’s just a simple models health check :

I tell the main agent:

For each model available to you:
1. Spawn a subagent using that model and ask it to list files in its home directory.
2. If it succeeds, send me in Discord: "model_name" --- OK.
   If it fails, send: "model_name" --- "error_reason".
3. Wait 5 seconds.
4. Move on to the next model.

In practice, this is what happens:

• The main agent correctly spawns the subagents
• The subagents execute properly on-time (I can see successful responses in LLM proxy)
• However, the main agent reports: "model_name" did not return a result

In the gateway WebUI, I can see that subagent messages only occasionally make it back into the main agent’s queue, with a delay of 1-2 minutes. By the time they arrive, the main agent has already reported that the model failed
Then, after the main agent finishes reporting that all models were silent, a large batch of subagent responses appears in the session, arriving one after another with 1-2 second gaps. It looks like they were queued somewhere and then flushed in bulk. I’ve tried tuning the queue behavior, but it doesn’t seem to change the outcome

Also, most subagents in this type of task don’t return anything at all

A subagent task "save-home-list-mistral-medium" just completed successfully.

Findings:
(no output)

Stats: runtime 13s • tokens 9.9k (in 27.2k / out 759) • est $0.01 • sessionKey 

However, if I ask them to write their results to a file instead, they work perfectly and much faster than 13s
I asked in the help channel, and the guidance I got from @sharp parcel was so off the mark it almost blew up my infrastructure.

[fervent nest]

my gateway is up for like 10 seconds and then disconnects itself and the disconnection says no reason. that keep happening in a loop. i just installed it havent even been able to send my first promt. can someone help

[boreal galleon]

gpt-5.3-codex works only with OAuth

[slow geyser]

This looked interesting - https://x.com/jumperz/status/2020305891430428767

[crisp wing]

@jumperz via Twitter

JUMPERZ (@jumperz)

📰 I Built an AI Agent Swarm in Discord. It Works Better Than Anything I’ve Tried (Full Guide)
So I built this agent coordination system that lives in my Discord server.

they talk to each other, split work, and deliver results, I was surprised how easily this actually worked out. It's honestly kinda wild.

let me walk you through everything..

# The Entry Point ( YOU )

You type plain English in #orders channel.. no special commands. its a discord channel where everything starts.

• "Research the top 10 AI coding tools and write a comparison thread"
• "Build a landing page for my new SaaS idea"
• "Track engagement on my last 5 tweets and suggest improvements"
  That's it, just natural language..just drop your task like you're texting a friend.

=========

# The Brain (Coordinator)…

FxTwitter · Feb 8, 2026, 1:17 AM

[modest frigate]

Hello World

[sacred vale]

sorry to ask, maybe this was asked 100 times before. but while configuring openclaw for the first time I was overwhelmed with configuring local llm stuff and I answerd the hatch agent via TIU or web with SKIP for now. How can I hatch it later?? Just tell it to read and follow the BOOTSTRAP.md ???

[modest frigate]

If you have access to Claude Code, just point it to your folder and then tell claude code you need it to deal with the questions you skipped interactively.

[boreal galleon]

I already have agent swarm in Discord but it's slow and doesn't work well with marked text like MD or JSON, but looks cool, agreed

[indigo nebula]

on the command line, run
openclaw dashboard. That will either open your browser with the token (if you're on the same machine), or show you the URL with the token for you to copy.

[sturdy mural]

Thanks!!

[slow sluice]

I’m looking for someone to help me set this openclaw up, I work in real estate and want to get a ai to run through these list for me. I would be more than happy to pay someone to set this up or walk me through how to do this.

[lusty knot]

why's my gemini api key not working on openclaw anymore
giving (no output) 429 error even though i haven't hit any rate limit

[candid imp]

checkout if its not gemini-3-flash-preview

[spiral thorn]

Hey my number got banned from Telegram. Has this happened to anyone else? Is there any recourse?

[candid imp]

If you connected Openclaw (or any userbot) to your personal phone number instead of using a bot api token, you likely triggered their anti-spam systems.

[spiral thorn]

Yeah I went through the whole botfather process and did it properly

using an api token from the botfather

[candid imp]

maybe your bot probably had spammed the server..

You canend the appeal email immediately.....many devs get unbanned within 24-48 hours if they admit it was a coding loop error.

Personally did that but i was uding temporary phone number....donot use your personal contact

[solid kiln]

Anyone find a fix for "Cannot read properties of undefined (reading '0')" as the response?

[solid kiln]

I can assist you

[spiral thorn]

How are most people setting up chatting with their bots from their phones? I wish imessages via imsg or bluebubbles worked better. Its pretty unreliable for me. I like using imessages because i can walk and talk and reply with siri reading my messages

[olive oasis]

telegram

[candid imp]

are you using localAI, Ollama, Azure, or OpenRouter

[solid kiln]

As the model? Possible ollama but through Venice.ai

[candid imp]

1st: Telegram
2nd : Discord

[candid imp]

Venice does not acceptt ollama as a model name. If you send model.. ollama.. venice returns a 400 Error JSON.... OpenClaw tries to read that error as a success message, finds no choices and throws the undefined error.

[candid imp]

these ( venice.ai and ollama) are way two different.....does venice.ai have ollama in there model list

[solid kiln]

im not sure, but I see I was confused. I am using llama 3.3 (from venice) and gpt 5.1 codex. somehow it is just fixed now after two repeated occurences.

[kindred gate]

I am trying to setup clawbot with ollama but whenever i press space to choose a model it types in the filter secction. I have been tryiong to stop it from filtering buyt nothing has worked so far

[lusty knot]

doesn't gemini 3 flash preview come with google search?

[blissful ore]

hey guys, some one here can help or explain how to fix this message ?

openclaw Install GitHub CLI (brew)brew not installed

[kindred gate]

I think you dont have github CLi installed

[blissful ore]

second question, how to do that XD

[kindred gate]

search guthub CLi

and download it

it brings Github to your command prompt

its needed to run openclaw

[dusk geode]

Are openclaw bots allowed to post in the help session and talk to krill?

[kindred gate]

i think it can do it but im not sure if its allowed to do it

[dusk geode]

I checked rules didn’t see anything against it. Was just curious if anyone had tried it.

[kindred gate]

i dont think there are any rules per se

[blissful ore]

i have, i download it but its says to repair or remove

[kindred gate]

uhhhh

try the install openclaw command again

and see what it says

[blissful ore]

i have openclaw running

[kindred gate]

??

restart it i guess

[dusk geode]

Might give it a shot and see what happens. Im just interested in how that conversation would go.

[kindred gate]

i think its more that just keep it within basic human decency

and dont scam people

and other basic humaity

[stark hawk]

I got everything set up yesterday using a self-hosted vLLM instance running a Qwen3 Coder model. Things were working well but now it's stopped. I give the agent a task, it says it's going to do whatever the first step would be, and then nothing else happens. I don't see any errors in any logs or anything. Any guidance?

[livid gull]

Hi, can someone assist me on this error:

after running -> pnpm install

Failed in 2s at C:\Users\Admin\programming\OpenClawAI\openclaw\node_modules.pnpm\node llama-cpp@3.15.1_typescript@5.9.3\node_modules\node-llama-cpp ELIFECYCLE  Command failed with exit code 3221225477.

[haughty hedge]

Any way to get agents to only reply to certain telegram group topics?

[rough pond]

are you guys just telling Clawdbot to setup subagents to use local models or how do you integrate local models?
i am currently setting up and i would like to use some local models for sub tasks

[stark hawk]

I just created a couple of agents using the guide here. You can tell the main agent what model the subagent should use: https://github.com/digitalknk/openclaw-runbook/blob/main/examples/agent-prompts.md

[rough pond]

Thank you

[blissful ore]

i stil get this message ?

23:40:04 [ws] ⇄ res ✗ skills.install 78ms errorCode=UNAVAILABLE errorMessage=brew not installed conn=db72b103…cdc3 id=c83ce204…c866

[trail cedar]

Is iMessage just not working?

[blissful ore]

how do u mean ?

[autumn junco]

Hey everyone! I need some advice regarding AI APIs. Currently, I’m using Gemini Flash 2.5, but I frequently run into token per minute limits, which block my work and I have to wait a whole day for the reset. I’d like to avoid this. I’m looking for an API/solution with a monthly subscription plan, no token per minute limits (or very high limits), and good reasoning capabilities. Ideally, it would also have vision (image recognition), but it’s not a must. I need something that allows stable performance without constant throttling, especially for building a chatbot. Has anyone used something like this? Which providers have such plans? OpenAI, Anthropic, xAI, Google, or others? Thanks in advance for your help!

[pseudo oyster]

I am having great success using ollama $20/mo and asked my main agent to select and assign specific cloud versions to tasks.
I'm using openclaw for about 8-12 hours a day and have yet to hit limits running my lab 24/7.
And probably overkills but I also use Claude and Codex $20/mo subscriptions for bigger coding tasks, fixes, and larger architecture needs.

[autumn junco]

i am running open claw on aws can i still use ollama?

[blissful ore]

Can u send an link of what u use I’m doing the same almost and u make me curious now 😄😃

[winter depot]

I'm currently using kimi k2.5 coding api. You can get the lowest subscription for a dollar. You have to negotiate with the AI though.

[young nebula]

how do you get the crome extension to work when the vps doesnt have a browser i just constantly have the exclaimation point

[south cloud]

Hi all, a bit about my setup. Running LM Studio & Openclaw on a Ubuntu VM. I have GPU passthrough set up and the vm has 16GB of vram. Currently in th elm studio developer tab, I have a 9.12GB model loaded with a context length of 30000 tokens. According to watch -n 1 nvidia-smi, the model and context length take up a total of 15.01GB - so maxed out my GPU. In the openclaw.json, I have the contextWindow set to 30000. Why am I telling ya'll this? Well I looking for someone that is also experimenting with a similar setup. Someone else that is attempting to optimze openclaw

Reach out if you're trying something close, we can compare notes.

I am looking to optimize pruning and memory to understand how the openclaw builts and stores memory without losing context.

[strong sonnet]

30k is too small

[south cloud]

Yes it is. I am purposefully doing this to try and understand how the compaction feature is triggered, and what gets written to session memory. I am trying to understand pruning and wrting to the /memory/YYYY-MM-DD.md

Suggestions and knowledge are welcomed and asked for.

@strong sonnet are you suggesting I switch to a smaller model and increase context length? Or to not bother because my hardware won't run a larger model and context length?

[cedar dagger]

Hey guys, i started to do voice chats with my openclaw, it's really slow to do the transcription (i'm on oracle's free vps). I think it's using something called whisper to do the transcription locally. Is there a faster option?

[rotund sun]

you'll need to pay Whisper API with openai whisper to get quick transcripts, but I with an rtx 3060 using ollama 7b ran fine up to 2 minutes of audio

@south cloud The biggest problem is managing large contexts, in the current form We need to have a weekly purge, but without losing memory, using applications like SuperMemory, but in any case, for reading this co-text, it would consume many tokens, so I see that the ideal in this scenario is to use Multi-Agent, that is

[south cloud]

@rotund sun A cursory read through SuperMemory AI-agents leads me to believe that is what openclaw is kinda attempting to solve with the ./openclaw/agents/main/seesion/ , the /.openclaw/workspace/memory/YYYY--MM-DD.md and the /.openclaw/MEMORY.md. That is why I'm trying to understand Companction, the Pi npm module and how to trigger it. ATM, from the developer logs in LM STudio, I am noticing that openclaw gateway send a lot of (seemingly) useless into to LM Studio. If my guess is correct, a lot of useless stuff is taking up precious context length. I need to dig futher into both what the gateway is sending and exactly what LM Studio model is receiving.

[rotund sun]

and you even set up the context limits and tokens per agent? as well as memory management with the MEMORY.md files? I was also trying to understand this consumption management tokens x context because I noticed that channels like slack spend more tokens, as well as things that we can program outside of LLM to save tokens, I wanted to understand how to persist memory like you trying

[drifting crescent]

Guys how do I edit the json on Mac to lm studio

[stiff grotto]

What model(s) are you loading in LM Studio?

[south cloud]

This is the first couple of lines sent to LM Studio from the openclaw gateway. The full json object is 1515 lines long. deleted and edited wall of text.

I can delete the last couple of posts to get rid of the wall of text, but wanted to share it here so peeps can take a quick look. openclaw is sending a lot of useless info to the model

[south cloud]

There are a couple of edits that you'll need to make to get openclaw to use a model in ML Studio.

[rotund sun]

It even seems purposeful to spend more tokens.

[copper ruin]

Hello, I want to set up openclaw with an anthropic API. For moderate use how long do you think $30 in credit will last?

[rotund sun]

@south cloud https://docs.openclaw.ai/concepts/memory#automatic-memory-flush-pre-compaction-ping

[south cloud]

@rude From docs.openclaw.ai/reference/token-use "Openclaw tracks tokens, not characters. Tokens are model-specific, but most OpenAI-style models avaerage ~4 characters per token for English text."

[rotund sun]

you already tried this?

[south cloud]

Yes I have already made the edit to the .json file and I'm testing now. I'm running into glitches but not sure if it's openclaw side or lm studio side

[rotund sun]

isolate then and use minimax with oauth now for tshoot

its for free

just need to create a account https://www.minimax.io/

[south cloud]

I'll look into it later, maybe tomorrow. In the middle of testing locally right now.

[rotund sun]

no problem thanks for sharing ur experiences i'm trying to explore autonomous agents to create like a "nurse" remotely to my grandma

[south cloud]

That's a cool and useful usecase for agents and openclaw

[tiny flame]

Anyone have an idea why my agents default back to my brains model? I’m using sonnet for my brain but I’m trying to use googles model for research. I have the api key setup but whenever it spawns the agent called researcher google fails and falls back to sonnet

[south cloud]

I've only been working with openclaw for a couple of days now, so I'm not completely certain. AND I have only used openclaw in the browser; no telegram, signal, whatsapp etc. But to my limited understanding, you need to start a 2nd agent. 1 for you main with it's assigned model and API key, and a 2nd with it's own model and API key. Switch use between the 2 as you need.

[tiny flame]

Hmm ok. When I set up I created agents with names and designated the models it needed to use but it just won’t work.

[south cloud]

Stupid question, but have you tried asking your agent how to setup and configure a 2nd agent, then switch between the 2?

[oblique yarrow]

I am getting this issue Error - Sessions file path must be within sessions directory

[south cloud]

@Lrod When I asked my agent, this is how it responded: To create a second agent with a different model, you'll need to define it in your workspace configuration. Here's how you can proceed:

Define the Second Agent:

Create an AGENTS.md file if it doesn't already exist.
Add a section for the second agent, specifying its name, model, and any other relevant details (e.g., personality or preferences).
Configure the Agent:

Use the gateway tool to apply your configuration changes. This will ensure that OpenClaw recognizes the new agent.
Switch Between Agents:

Once the second agent is defined, you can switch between it and the main agent using commands like sessions_spawn.
Let me guide you through this step-by-step. First, I'll check if an AGENTS.md file already exists in your workspace. If not, I'll create one for you.

[oblique yarrow]

Openclaw doctor failed : error : sessions file path must be within sessions directory

[frigid ravine]

how does one have the gateway service always running. I use cursor to edit the config now and then. when I open cursor the gateway works. When I close cursor i see disconnected (1006): no reason and i just cannot seem to keep the dashboard connected/active

[south cloud]

I've just been using openclaw gateway restart everytime I need to edit openclaw.json

[frigid ravine]

no expert but have you tried cursor or any IDE such as VS code to edit the json file

I pretty much now know all the commands there is thanks to all the troubleshooting i have had to do past couple of days. Will see if there is another way instead of needing terminal or even cursor to open the dashboard and have it remain connected as telegram is

[rotund sun]

The main question is how the service is running, did you follow the official installation guide?

[placid saffron]

im having trouble switching models in chat on the web dashboard the only model that shows up is the primary model but in my json i have a different model set as primary for agents and 2 fallbacks, none of which are showing up. I'm also not able to switch models, I keep getting the allowlist error but I don't have an allowlist set up.

[frigid ravine]

yes i did - several times even. perhaps the service is running as user service and not system service

[rotund sun]

I only make the statement like this more or less in mine to work the fallback: Primary: xxxx, Fallbacks: [yyy,zzz]

[frigid ravine]

in telegram or even on the web dashboard - have you tried putting in /models and seeing what comes up

[placid saffron]

yes so everything comes up but it doesnt let me switch

[cinder marten]

CDP port mapping issue with sandbox mode "non-main".

Basically cannot get browser going in docker container.
Setup: openclaw v2026.2.12 installed on bare metal VPS with sandboxing mode 'non-main'.
Per Gemini:
It is now clear that for v2026.2.12 on bare metal, modifying the config.json sandbox settings isn't enough to solve the "CDP port mapping" error for isolated sessions. The Gateway's internal resolution logic is timing out (around 110ms) before the Docker bridge—or even the host network—can register the new container's metadata.

The recommendation to switch to sessionTarget: "main" isn't just a "workaround"; in this version, it is the architecturally stable way to run fast, recurring tasks like weather checks on VPS hardware.

Problem with cron job using sessionTarget: "main" only allows for notifications, not agentic workflow, or at least that's what openclaw bot tells me.

Am I missing something? Seems like people are doing a million things with openclaw, and I can't even get it to automate fetching the weather, which I can easily do with autohotkey. FML! Can anyone point me in the right direction?

[placid saffron]

so when i do /models and select one it just gives the same error

[rotund sun]

https://docs.openclaw.ai/providers

[frigid ravine]

whats the error as that tells alot as to what the issue could be

[rotund sun]

Openclaw models show for you some useful logs

[placid saffron]

model set failed: Error: model not allowed: anthropic/claude-haiku-4-5

but when i look at the faq

and i check the config for the default

it says its unset

[rotund sun]

Now you can append" list" in command to visualize which models is configured or miss configured

Openclaw models list

[placid saffron]

Model Input Ctx Local Auth Tags
anthropic/claude-opus-4-6 text+image 195k no yes default,configured,alias:opus
anthropic/claude-sonnet-4-5 text+image 195k no yes fallback#1,configured,alias:sonnet
anthropic/claude-haiku-4-5 text+image 195k no yes fallback#2,configured,alias:haiku

[reef blade]

It seems OC is reverting back my sandbox settings from off to all. Am I missing some setting or chron? Is this happening to anyone else?

[placid saffron]

is there a way i can disable the allowlist so i can just use whatever model i set

[obsidian jolt]

Man why did my hostinger vps telegram GitHub copilot gpt 4.1 tell me it’s gonna work all night and update my website with info
But today it’s telling me you’re absolutely right and I will do it now type shit?

How do we make it autonomous for real?

[rotund sun]

In my case, my bot forgot to run the cron 3x in a row, even though I was talking to run automatic, very strange.

[frigid ravine]

based on limited few days experience - i have notice openclaw doctor mess things up in the name of fixing them. one slight issue and your whole working setup i snot working anymore because the good OC doctor decided it is not structured properly

[tiny flame]

Yeah I have and I’ve done all those things but I’m trying to set it up to where the brain can call the agent do the work

[placid saffron]

anyone have any ideas

[south cloud]

Ask the main agent how to use a sub agent

[rotund sun]

I'm sorry, but do you have API credits on these LLM tasters?

[tiny flame]

It has no problem summoning the agent to make it work but it fails to sue the model despite it explicitly having the model set to something different so it falls back to sonnet

[rotund sun]

Even if you set up but don't have credits, it doesn't work.

[placid saffron]

i have an api key with a few dollars

and opus works but sonnet doesnt

nor haiku

[rotund sun]

But are you using 3 models from the same provider??? why?

The idea is a little different I believe, it is the same to have 2 energy sources from the same source.

You got it?

[rotund sun]

Try to use something that accepts only text input and another fallback that accepts image and text input, then you will understand a little better I think.

[placid saffron]

well no my goal is to use cheaper models for less complex tasks

but i only have an anthropic key 😜

[rotund sun]

You can configure antigravity too

And minimax

Its for free

[placid saffron]

i'm not entirely sure what those are

but my first issue is i cant switch to any model

even openrouter free ones

[rotund sun]

Antigravity is from Google.

[placid saffron]

isn't it an IDE?

[reef galleon]

Okay, strange question - I was using ChatGPT 5.3 Oauth for free. I've hit my rate limit and my gateway agent is useless. Is there a free model out there that I can quickly connect to in order to get me through spinning up my own local?

[rotund sun]

Yes, but take a look, they and Minimax are offering LLMs for a while.

[placid saffron]

i see

my issue still

is that i cant get my openclaw to switch models

like right now i dont want to use opus i want to use haiku

but when i try running /model anthropic/claude-haiku-4-5 or selecting it from models list

it throws the error

[rotund sun]

It depends on your computer.

[rotund sun]

Did you try to use only 1 and see the result?

[placid saffron]

i cant switch at all

[reef galleon]

I'm not really looking to find out the best open source local model, I'm just trying to find a good free oauth model that I can connect to just to get my agent back up for a bit.

[placid saffron]

im a little lost with the config settings

right now the only one working is opus-4-5 which isn't even in my openclaw models list output

[rotund sun]

I believe that if you don't set up multiagent no, because they are from the same provider.

[rotund sun]

Antigravity

[reef galleon]

I thought antigravity was only an AI powered IDE, not a full model that you can use for general purposes. Thanks, I'l check it out.

[placid saffron]

i just switched my primary model in my config to claude opus 4-6 but on the web version it still shows it as opus 4-5 under agents

[pearl salmon]

Does anyone know why discord bots only respond in DM? Is that a known issue?

[rotund sun]

I think there's OAuth for antigravity, yes, take a look.

[south cloud]

@rotund sun Ok, so I can run LM Studio headless and with it's SDK, there's a lot I can do with context length and resetting it in LM Studio. Now I just have to understand how OpenClaw triggers and handle Compaction, Pruning and writing to the various memory.md files. If I can also get an agent to call JavaScript scripts, I'm golden.

[spring crow]

does anyone understand the code, specifically https://github.com/openclaw/openclaw/blob/84ed9ab5541d4d9486b6155362103af11507a446/src/config/types.models.ts#L10

[rotund sun]

Yes, I believe that all magic happens when we end the right contexts at the right times and whether or not we know whether or not we will reuse it, combined with the power of memory.md and the YYYY-MMDD/Memory files work.

[pearl salmon]

What models are you local guys running? I’ve only managed to get qwen3 4B instruct 2507 to work properly with openClaw.

[spring crow]

I am trying to figure out where this stuff can be configured. I have a bug that I believe would be fixed by setting supportsStore to false. I found this in the code but can't figure out how to set it and it's not in the docs

[winged bridge]

Hey I got an issue with first installation I cant open the getting started docs website and the quick start one liner doesnt work either in the terminal saying my OpenSSL version number is wrong but its up to date and the Ubuntu is freshly installed

[south cloud]

There are abstract objects. Classes in Java or C# but using TypeScript/JavaScript

[spring crow]

I'm confused as to how those objets get set in the code

ModelDefinitionConfig looks like what you control via the openclaw.json

but i was curious if we could also set ModelCompatConfig objects somewhere

[rotund sun]

qwen2.5-coder:7b-instruct-q4_K_M

[south cloud]

I haven't gotten that far into Openclaw yet, but I doubt we could access the instantiation of those objects - at least not without it breaking eveytime OpenClaw updates.

[spring crow]

unfortunate

so only the ModelDefinitionConfig interface is indirectly exposed through the json?

it's just weird because i recognize a lot of these objects as being fields that can be easily configured

[south cloud]

It seems some Gateway properties are as well, but again I've only been at it for a couple of days now and still limited in my understanding

[spring crow]

https://zeabur.com/forum/posts/69815910ecc019f67a6d12c4 there is also a screenshot of the same issue I am having being fixed in a similar way

[pseudo oyster]

unsure. maybe? ask your favorite LLM. that's what i do.

[south cloud]

I see what you are saying, but that image cuts off some context. It looks like it can be set in the model definition of the .json

[pseudo oyster]

if you go onto Ollama you can sort by cloud and most current and most popular. dig in and read what makes sense for what you want to do.
For big decisions i cant figure out; I ask Gemini, then share that conversation with Claude (copy/paste), then to Gpt; back and forth until they help me make a decision i feel confident with. and let them know you're doing it and that they're your team. it woks awesomely

[spring crow]

yes but i have tried that and supportsStore seems not to be recognized

using the latest version of openclawd

[pearl salmon]

Thanks for replying. Have you had any luck with any qwen3 models? I’m trying to get on a reasoning model but I only have 8GB vram

[spring crow]

maybe not the latest but, 5 hours ago 😄

[south cloud]

Will I've found even the install scripts to be buggy, so maybe there's alreay an open issue that will get fied in the future.

[rotund sun]

no my friend 🙁 it's seems to be like you need a NASA PC to run locally models hahahaha

[pearl salmon]

Truth. Does that model do reasoning ?

[remote crow]

C:\Windows\System32>npm i -g openclaw
npm error code ENOENT
npm error syscall spawn git
npm error path git
npm error errno -4058
npm error enoent An unknown git error occurred
npm error enoent This is related to npm not being able to find a file.
npm error enoent
npm error A complete log of this run can be found in: C:\Users\rylan\AppData\Local\npm-cache_logs\2026-02-14T04_08_18_932Z-debug-0.log

[rotund sun]

i'm not sure but you can find some details here: https://ollama.com/library/qwen2.5-coder:7b-instruct-q4_K_M

[remote crow]

can anyone help

[weak otter]

that new update broke my discord session and all sessions.....

even TUI wont work

[south cloud]

Make sure you're using nvm and have node v22.22.0 installed. Clean and update all npm packages? I haven't used windows in 10 yrs so.....I'm just guessing

EDIT: also set default node version to 22.22.0 it's what openclaw wants

[placid saffron]

root@openclaw:~/.openclaw# openclaw update
Updating OpenClaw...

│
◇ ✗ Updating via package manager (309.82s)

Update Result: ERROR
Root: /usr/lib/node_modules/openclaw
Reason: global update
Before: 2026.2.12

Total time: 310.13s

Anyone know what this means

[weak otter]

Not funny @south cloud but what is funny is that happened on 2026.2.13 not on the previous one 😩

[placid saffron]

the new update just broke everything 😂

[south cloud]

Thats what happens when bootcamp developers start thinking they're software architects.

[split dew]

Hi! I'm trying to set up Claude Max API Proxy (https://docs.openclaw.ai/providers/claude-max-api-proxy) for the first time. I've been running into various errors setting up the configuration. I did hear that the latest update overhauled the structure of openclaw.json. Can someone please share the specific file config that works for you?

[weak otter]

Did just just called claw devs bootcamp developers?

[south cloud]

Well, not ALL of them

[rotund sun]

have you exposed the endpoint?

http://localhost:3456/v1/chat/completions

[pearl salmon]

Did the recent update break discord bots? Mine will only reply in DMs

[rotund sun]

bad news: This is a community tool, not officially supported by Anthropic or OpenClaw

https://docs.openclaw.ai/providers/claude-max-api-proxy

[celest vigil]

Yes it did. I need to downgrade. Let me find the version that i downgraded to

[pearl salmon]

Ahhh well that explains the 4 hrs of confusion I went through following tutorials and unable to get my bot to talk in the channels lol. FML. 🤦

[celest vigil]

Downgrade it to v2026.2.9 https://github.com/openclaw/openclaw/issues/15238

https://github.com/openclaw/openclaw/issues/15187

[pearl salmon]

Thank you kind sir

[wispy spruce]

new setup, onboard doesnt ask for AI model? I had to manually type openclaw configure --section model ?

[candid imp]

Security checklist: gateway not public, pairing required, filesystem scoped (no /), and access via Tailscale/SSH tunnel.
If a provider can’t show you “nmap clean” + no root mounts in 5 min, don’t hand them your API keys.

[fierce gazelle]

Do you have node and nvm installed ?

[reef galleon]

Does anyone know what free version of Antigravity I can use with OpenClaw? I've tried a couple of models and my OpenClaw just hangs up and never gets a response.

[rotund sun]

yes my models are configured some like this:

openclaw models list

🦞 OpenClaw 2026.2.12 (f9e444d) — The only crab in your contacts you actually want to hear from. 🦞

Model Input Ctx Local Auth Tags
minimax-portal/MiniMax-M2.5 text 195k no yes default,configured
google-antigravity/gemini-3-flash text+image 1024k no yes fallback#1,configured
openai-codex/gpt-5.1-codex-mini text+image 266k no yes fallback#2,configured

[reef galleon]

are those all free models?

[rotund sun]

no codex are paid but not expensive because use oauth too

[reef galleon]

omg i figured it out. I setup anthropic gemini flash. not google antigravity

[rotund sun]

lol

[raven sand]

How do i use my openai sub to use the model in openclaw ?

[reef galleon]

I used oauth to utilize my Pro subscription with chatgpt

[celest vigil]

The API subscription?

[graceful oxide]

Why would/wouldn't i get a moltbot of my own? What are the pros and cons?

[eternal finch]

How can I complete the Gemini OAuth flow on a headless Ubuntu Server 24.04 (Raspberry Pi) via SSH, since my previous API key attempts failed and I have no browser?

[meager cypress]

pros: everything, cons: you won't sleep for a week

[graceful oxide]

sounds too good to be true, youre either missing something or hiding something

[wanton cedar]

how can i get the free ones?

[meager cypress]

well if you want to explore what agents will be in the near future then its cool, if you want to automate a lot of things, its good.
cons: its too rough around the edges and buggy

but the bugs will be ironed out

[graceful oxide]

i see.

[graceful oxide]

as a person skilled in jailbreaking ai, no not all of them

if i sound a tad hostile/confrontational i dont intent to be, thank you very much for answering my questions

[meager cypress]

well i wouldn't call prompt injection a bug

all AI systems are vulnerable, basically computer version of social engineering

[graceful oxide]

inability to handle them well is

[meager cypress]

design flaw =/= bug

[graceful oxide]

from what ive heard they occasionally jailbreak themselves by accident through hallucination

and then uhh leak peoples personal info

[graceful oxide]

fair point

but then is the design flaw not a con

[meager cypress]

but yeah i mean if you're paranoid about that then you can just have one completely locked down but the fun and utility is limited, the whole point is letting the thing go do its thing

[graceful oxide]

fair

but that doesnt seem like paranoia that just seems like common sense

[weak otter]

2026.2.12 broke my whole setup, 2026.2.13 didn't fix anything. Now what? Grab a chancla squash the crab 👀

[graceful oxide]

like ive built a setup on antigrav that if i wanted to i could have all the functions of a clawdbot in it

but i still implement things to make sure it doesnt like brick my pc

[meager cypress]

well just run it on a VM/Docker/VPS

[weak otter]

be aware google is banning accounts for using openclaw with antigravity OAuth

[tiny cedar]

Hi there everyone.
Can someone please help with maybe stupid question? I decided to use Gemini 3pro and connect trough googles antigravity auth via url. And all went well but then after using it a little I’ve got this massage: Cloud Code Assist API error (403): This service has been disabled in this account for violation of Terms of
Service. If you believe this is an error, contact [e-mail].
and I have a student subscription. Did someone got anything like it in a similar situation? And figured out how to avoid it?! Please help if so.

[weak otter]

LOL:

banned

[meager cypress]

on projects like this I hold off for a while before updating, half the time its new bugs galore rather than fixes

[weak otter]

Yeah bro i just said that 👀

[meager cypress]

you can get pretty decent models for free with openrouter if you don't need image support

[tiny cedar]

🤣 I didn’t saw lol. So nothing can be done?)

[meager cypress]

and get some more stuff from zen

[tiny cedar]

Like?

[weak otter]

Ah you can try contact em and lie, but they know everything you do the moment you use their service... Its Google 😄

[tiny cedar]

And what about open ai auth through subscription a codex?

Do they ban also?

[weak otter]

Man you should contact em or read their ToS before doing anything, i mean chances are they don't want you using it either

or just ask gpt if it is allowed

[tiny cedar]

I just don’t get it why there is even an option for using it if it’ll get banned anyway 🤷🏼‍♂️🤦🏻

[meager cypress]

https://opencode.ai/zen you can get glm 4.7, minimax 2.5 for free here (with rate limits)

[tiny cedar]

And do u guys may be know how to get the bot serf the web and do web search as well without getting it a brave API?

[wanton cedar]

do you know which if you dont mind me asking? or how

[weak otter]

Has anyone encounter this: error Delivery failed (discord to 1467238027557601313): Error: Ambiguous Discord recipient "1467238027557601313". Use "user:1467238027557601313" Looks like some variables changed but can't find anything related to it

[left arch]

im trying to startt a new session and it just deleetes my message

[meager cypress]

you can use tavily

[meager cypress]

in openrouter the free models keep changing but they have an smart free router now so you can just point it at that and it gives you the best available, https://openrouter.ai/openrouter/free

[weak otter]

i made a tool to use x.ai/grok the web_search tool, cuz $2 bucks here, $5 there and on and on can get crazy expensive and openclaw doesn't come with many free options pre-installed

[wanton cedar]

thank you

[weak otter]

Openrouter is broken after this update as well. It doesn't understand the api endpoint

[eternal finch]

Please!

[amber path]

Hey! Can someone help me:

1. Can I make separate Telegram bots for different sub agents?
2. Can agents send progress/plans on long-term tasks? If so, how can I set-up it?

[weak otter]

bro when you click the link on the console its gonna give you and error cuz the ip of your machine doesnt match the raspi

[tiny cedar]

Tanks guys both of you.

[weak otter]

but on the browser you will see on the address bar the refresh token hidden just copy from there paste it on the console

[graceful oxide]

not what i mean

[eternal finch]

Got it. Can I fix this by using SSH port forwarding (tunneling) to map the Pi's local port to my machine, and if so, what's the best way to set that up for the Gemini flow?

[weak otter]

you only need to do that once, for what i seen the refresh token never change. But your account could be disabled. Good luck good sir

[meager cypress]

must be your setup, mine is working fine

[weak otter]

Doubt it, was working before update

[eternal finch]

How would you recommend doing that? Should I run the OAuth flow on my local machine first and then SCP the token.json to the Pi, or is there a better way to handle this headlessly?

[wanton cedar]

if free models dont end up working for me whats the cheapest ai i should use?

[weak otter]

kimi k2.5

[wanton cedar]

where can i get?

[weak otter]

openrouter but i think there is cheaper providers don't quote me on that

[terse briar]

API provider returned a billing error — your API key has run out of credits or has an insufficient balance. Check your provider's billing dashboard and top up or switch to a different API key.

openclaw keeps returning this error for openrouter when I have credits in my account. Does anyone have a fix

[weak otter]

yep pay your bills homie

[terse briar]

i have credits loaded

[weak otter]

That isn't openclaw issue, you should contact em. They may be blocking you from using it or something else

[patent spire]

can i use codex's search tool in openclaw?

im running it with chatgpt subscription

[terse briar]

will do, is there a way to change an existing api key?

[weak otter]

from whoever your provider is usually under account > API keys> revoke existing one create a new one

[terse briar]

yeah im struggling to find that within openclaw though when i already have one set

[weak otter]

Yeah i know openclaw has the worse ui ever but its not that hard to do directly in the openclaw.conf

[terse briar]

Thank you, sorry if some questions are dumb as rocks

[weak otter]

in the panel go to configuration then within it you will see a toggler on the botton that says RAW click on that to open the config file, you can change the api from there then save, restart gateway

[ivory gazelle]

i'm also doing a first time setup, and i get to this step: "Invite the bot to your server with message permissions." but i don't know how to invite the bot. the bot isn't a discord user that i can click on and hit 'invite to server'

[robust zenith]

Hey folks

Not looking to recruit - just willing to pay for a couple of hours for someone to fully setup openclaw on my VPS and get headless browser working
(I got as far as telegram channel, gog workspace integration (calendar updates, email, sheets/etc))
If you have bandwidth and are a college kid looking for extra bucks, please DM me.

[ivory gazelle]

nevermind, i realized that was discord-specific steps, not openclaw steps.

[terse briar]

got it working again, there was no env, had to add it manually

[meager cypress]

discord bot config is unnecessarily confusing (it's discord's fault) but the openclaw docs are not amazing either

[ivory gazelle]

well they're good enough to get the gateway up and the webpage showing 😅

i'm running a 3080LHR. i'm still kind of peeping this thing to see if i want to use it so i don't mind if it's a bit slow - what's a good llmstudio model (huggingface) to connect it to?

or is the answer "none 🤣 "

[meager cypress]

easier to run something off openrouter than local tbh otherwise you'll be waiting forever if its slow

too many tool calls, background stuff

[vale zealot]

Anyone got this message before: The AI service is temporarily overloaded. Please try again in a moment. If so what fixed it? Thank you 🙏🏻

[ivory gazelle]

ok, thanks. i will try that. should i just put the "config snippet" in the openclaw doc into openclaw.json?

the cli setup doesn't seem to work the way i expect it to.

[ivory rapids]

Hey - which memory option do you use for clawdbot, the official one with an embedding provider model via api?

thinking about if i need thedotmack/claude-mem or if i use the internal memory function

[azure dagger]

boring answer, but waiting fixed it 😄

[edgy zenith]

I am finding it fascinating how bad openclaw with 5.2-chat-latest is at knowing its own internals. The only way I can configure and fix it reliably is with Codex 5.3 on the side, otherwise it hallucinates itself into broken state, every single time. https://x.com/bitfalls/status/2022589647252316587?s=20

[crisp wing]

@bitfalls via Twitter

Bruno Skvorc (@bitfalls)

A short story.
︀︀
︀︀Lesson: you cannot rely on @openclaw setting itself up AT ALL in any regard. It is completely wrong on all counts all the time no matter how much information you give it.
︀︀
︀︀This is codex 5.3 correcting it, for the 103rd time.
︀︀
︀︀Open to suggestions on how to make it less retarded. 5.3 not available via API so nothing I can do on the model side.

**👁️ 118 **

FxTwitter · Feb 14, 2026, 8:32 AM

[hard relic]

is mac mini m4 base good enough to run it and futureproof at least some decent amount of time?

[alpine jay]

Hey guys how can I get access to the ios testflight? Or source code?

[rain relic]

Hi

I thought I could run this on a raspberry pi 8gb ram model I had. So I threw on the latest Raspberry PI OS and updated the packages.

After ruining the one liner it skips over the model setup. Which I understand immediately breaks "hatching". Does anyone here have ideas what to do?

I've tried uninstalling via "openclaw uninstall" and removing openclaw directories

Do I need to be running this as root?

[dapper wren]

guys how u working with subagents, i tried to spawn 5 agent to do separated tasks in project and they all died within 1-5 minutes with tons of errors and "connection error"

[pastel night]

I guess I'm not the only one with broken Discord after the update. Now I've got reason to hook telegram for the first time.

[void cargo]

Do you mean that you didn't get the choices for provider/model? Sorry, I haven't had my coffee yet

[rain relic]

Yep, it skips over it entirely. Tried first over ssh, then again by using the terminal app included with Raspberry PI OS

[void cargo]

I can't tell you exactly why it's happening but a hack you can do is changing the config manually before starting the hatching process

Unless things have changed, I used this solution before and it worked

[rain relic]

How would I go about getting the right auth keys? I have access to the consumer tiers of OpenAI and Anthropic

[rain relic]

I could give it a go

[void cargo]

For that let's see if we can get anyone else support as I only use local models

[rain relic]

Okay

[half dagger]

I'm stuck on how to even get started.

o, as a non-technical founder, in "getting started,"
Control UI
The Control UI is a small Vite + Lit single-page app served by the Gateway:
default: http://<host>:18789/
optional prefix: set gateway.controlUi.basePath (e.g. /openclaw)
It speaks directly to the Gateway WebSocket on the same port.

<host> = ? what do I do with this?

Is there a page or .md somewher with setup instructions for non-engineers?

Thanks!

[west flax]

Anyone wanna learn together via video call / shared screen?

[lapis tapir]

hey guys i got a good openclaw setup going with claude cli as my only backend. my only concern is that i'm not using claude in an interactive mode so questions about my tasks don't reach me, sometimes i get a message back saying "while i am waiting for your feedback, i did the following“ but the task then never finished because it never got me answers.

claude command args use ‘—print‘.

is there any solution to this or is this the drawback i have to live with?

[devout flame]

I want to scrape public fb group posts, I'm running it on m1 max with ollama gpt-oss:20b. Couldn't find related skill. Does anyone here know how to do it? It can access browser and is not good with scraping post content, I'm providing post urls, it produces py script, but does not know how to run that itself on CLI.

[barren gate]

I built https://clawtick.com — a cloud scheduler for OpenClaw agents.

OpenClaw's built-in scheduler silently drops jobs. No logs, no errors, just missed messages.

ClawTick fixes that:
→ Reliable cron in the cloud
→ Dashboard + CLI
→ Delivers to WhatsApp, Telegram, Slack, Discord

[dusky jackal]

Hi all. Im pretty new here and was looking to get a bit of help or even just some answers. Im running Clawbot on an unraid server through the docker app. But i've noticed a few issues with the UI for it. Is anyone else using it and having similar issues? My agent is working fine. but I cant actually see what model is running in the "Agents" tab.

[hybrid veldt]

I’m running OpenClaw v2026.2.13 on a VPS. I get ‘Unsupported channel: telegram’ and ‘Unsupported channel: whatsapp’ errors. How can I enable chat/messaging channels? Is there a public plugin or new release?

[lone verge]

was the doctor command renamed or something?

error: unknown command 'doctor'
(Did you mean docs?)

(latest git commit 3bbd29be)

[jade sundial]

how do i actually make openclaw working continuously? right now it only respond to my morning briefings.

[lean ether]

I would appreciate a little advice. How are you coordinating multiple specialized AI agents to recogonize, communicate and work together on projects? See thread for details....

[pulsar urchin]

how do I add skills to my clawbot?

[steel hinge]

I installed and now how do I talk to it?

this is what I get in terminal ⠸ dillydallying… • 5m 10s | connected
agent main | session main | unknown | think low | tokens ?/200k

what now?

[verbal laurel]

the easy way is to run "openclaw gateway" and then use the webuy at 127.0.0.1:default_port

[steel hinge]

I already installed and connected it to discors

but now no longer answering in discord

are there commands?

[mental silo]

That open tui and "unknown" could be an issue

Unknown and ,tokens ? Might indicate no model connected?

Running a local model or connected via API?

[brisk ledge]

Got the same issue. Have reached out to support who have forwarded it to their technical team. Is pretty fundamental to getting the agent to be able to do anything!

[golden mural]

was the session file path error supposed to be fixed in 2026.2.13?

still getting it after updating

[rain relic]

It seems to be a known bug, just spotted this https://github.com/openclaw/openclaw/issues/16134

[unreal bear]

All of the guys with multiple agents, are they persistent or just spawned ad hoc by main???

[golden mural]

I have multiple persistent agents with different primary models / workspaces

I set up multiple discord bots, each bot gets treated as its own agent in openclaw. But I let them treat each other like sub agents too

[rare roost]

Hi Folks. Any news on the set-up issue where Model/Auth provider gets skipped (jumping straight to Channel select)? I'm sat here wanting to hatch but I have a lobotomized assistant!!

[vivid thicket]

is it possible to let openclaw control google chrome? as in open a tab go to a page and do a specific action? like click a button or write a post?

[storm copper]

i have 2 persistent agents. it's pretty comfy

[verbal laurel]

what is the best way to create 1 orchestrator (main agent) + different agents (different models) that will act as subagents but with own identity/soul ?

[void cargo]

Good find

[unreal bear]

What local model are you using? And how did you set them up? My main created two and then set them up on qwen3 but said it ended up putting them on Claude which just burns tokens

[storm copper]

i'm just using opus I can't find a decent local model.

[unreal bear]

Ah ok

[rain relic]

npm install -g openclaw@2026.2.12

Seems to be be working, it's letting me go through full setup. Something broke with the 2026.2.13 release

[prime spindle]

Just a quick advice. If someone is having this issue, just modify your DNS to:
nameserver 8.8.8.8
nameserver 8.8.4.4
nameserver 1.1.1.1

[fossil lichen]

W for coming back in with the solve

[prime spindle]

Thanks... Love you nick btw.

[indigo nebula]

Neat but why not just the built in system cron? I’ve been using system cron for basic tasks like ‘morning summary’, ‘email check’, and task type jobs. It’s free too

[fossil lichen]

system cron had issues up until 2.09 ish i think

[viscid plover]

pretty sure my ollama instance is running properly fine, I think it's caused by time outs by openclaw because the model is running on 1 tokens per second.```
◇ What would you like to change?
│ Change base URL and model
│
◇ API Base URL
│ http://127.0.0.1:11434/v1
│
◇ API Key (leave blank if not required)
│
│
◇ Model ID
│ llama3.3:70b
│
◇ Verification failed: This operation was aborted

[unreal bear]

Anyone know how to install openclaw chrome extension

[mild garden]

Hey! Anyone else running OpenClaw on Windows natively? I'm hitting a weird issue where Telegram inbound messages come in as plain text — no envelope metadata at all (no sender ID, no timestamps, no [message_id: XXXX]). The funny thing is reaction notifications DO include the message_id, so the Telegram plugin clearly has access to it — it's just not making it into the message envelope.

Running 2026.2.12, default config, embedded backend, no envelope overrides. Same setup on macOS works perfectly with the full envelope.

Also getting 200k context window instead of 1M for Opus 4.6 with no explicit contextTokens set, which feels related somehow.

Anyone seen anything like this? Starting to think it might be a Windows-specific bug.

[unreal bear]

Do you get normal responses in the dashboard?

[mild garden]

The responses are normal but in telegram every message comes with a message_id , but in my Windows openclaw it’s just plain text i don’t know why

[north lava]

hi why my agent struggles to click on a video in youtube, he successfully opens youtube searches 'lionel messi' but fails at opening one of the videos, i'm using qwen3:8b btw

help am i doing something wrong or is qwen3:8b too dumb for browser automation

[bitter orchid]

What's the reason for wanting to open a video? There are libraries to invest the content in a bit more streamlined way

[north lava]

im trying to get it to automate browser, i'm giving simple tasks as test, but it's failing the simple tasks, am i doing something wrong or is qwen3:8b too weak for browser automation

[bitter orchid]

8b in general isn't going to be a great model.

[north lava]

is there a better free alternative

[bitter orchid]

What's your hardware set up chief?

[north lava]

rtx 3050 8gb vram, 16gb ram, i3-12gen

[bitter orchid]

Honestly you probably won't be able to do YT automation with that.
However, you might can tweak things to get it there.

Try telling it to look at the inspect window

[unreal bear]

Guys how do I see ./openclaw in finder

[bitter orchid]

Clicking play there

[north lava]

alright will try some workarounds, thx for the help

[austere fjord]

Am i ment to have the HEARTBEAT.md, BOOTSTRAP.md and MEMORY.md files or do they get automatically generated?

nothing is broken im just wondering

[frank ginkgo]

I opened OpenClaw onboard and the option to choose the model is not appearing; it goes straight to Telegram, Discord, etc. Any solution?

[sinful shore]

Hi, I need help, I have been stuck on this error for too long, checked docs and with Claude, no luck

from logs:
"[ws] closed before connect conn=fafc6694-6f85-457b-b556-533475216c53 remote=172.18.0.1 fwd=n/a origin=http://localhost:18789 host=localhost:18789 ua=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/144.0.0.0 Safari/537.36 Edg/144.0.0.0 code=1008 reason=pairing required"

[bitter orchid]

They get made when u start.

[bitter orchid]

Looks like your remote gateway shut off

[bitter orchid]

Try to run openclaw setup. What's it say?

Or openclaw doctor

[frank ginkgo]

🦞 OpenClaw 2026.2.13 (203b5bd) — If it's repetitive, I'll automate it; if it's hard, I'll bring jokes and a rollback plan.

Config OK: ~/.openclaw/openclaw.json
Workspace OK: ~/.openclaw/workspace
Sessions OK: ~/.openclaw/agents/main/sessions

[viscid plover]

pls help, openclaw is not responding because my device is generating at least 1 tokens per second

[bitter orchid]

Ok. Now try doctor

Also if you are able to talk to your bot still. You can ask it to fix itself. Or to put in the API key

Or ask it to tell you how

[frank ginkgo]

nah i cant talk to my bot, since the first install i couldnt select any model to send my api key

😩

Updated ~/.openclaw/openclaw.json
Backup: ~/.openclaw/openclaw.json.bak
│
└ Doctor complete.

[bitter orchid]

Just purge it and start over

I've had to do it several times

[frank ginkgo]

ok thx thx

[bitter orchid]

Part of learning 😂

[flint turtle]

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

To prevent this, increase your compaction buffer by setting agents.defaults.compaction.reserveTokensFloor to 4000 or higher in your config.

[bitter orchid]

Annoying I know

[frank ginkgo]

appreciate you @bitter orchid

[bitter orchid]

You need to optimize your bot
Give this a watch. Very helpful
https://youtu.be/ZrBvPoFBVUo?si=wIeeBuaqG9uwd___

[sinful shore]

how do I fix it? I running openclaw in docker so the gateway config is confusing to me, can't seem to have any progress. Tried local and LAN with no luck...

[bitter orchid]

Your docker container isn't port forwarding properly.

Ask grok or something. I'm really not great at that

[frank ginkgo]

◇ I understand this is powerful and inherently risky. Continue?
│ Yes
│
◇ Onboarding mode
│ QuickStart
│
◇ QuickStart ─────────────────────────╮
│ │
│ Gateway port: 18789 │
│ Gateway bind: Loopback (127.0.0.1) │
│ Gateway auth: Token (default) │
│ Tailscale exposure: Off │
│ Direct to chat channels. │
│ │
├──────────────────────────────────────╯
│
◇ Model check ───────────────────────────────────────────────────────────────────────────╮
│ │
│ No auth configured for provider "anthropic". The agent may fail until credentials are │
│ added. │
│ │
├─────────────────────────────────────────────────────────────────────────────────────────╯
│
◇ Channel status ────────────────────────────╮
│ │
│ Telegram: not configured │
│ WhatsApp: not configured │
│ Discord: not configured │
│ IRC: not configured │
│ Google Chat: not configured │
│ Slack: not configured │
│ Signal: not configured │
│ iMessage: not configured │
│ Feishu: install plugin to enable │
│ Google Chat: install plugin to enable │
│ Nostr: install plugin to enable │
│ Microsoft Teams: install plugin to enable │
│ Mattermost: install plugin to enable │
│ Nextcloud Talk: install plugin to enable │
│ Matrix: install plugin to enable

again @bitter orchid

[bitter orchid]

Ah. Do manual

Not quickstart

[sinful shore]

I'm using Claude to help but still, came here as last resort 🙁

[frank ginkgo]

manual - local gateway?

or remote one

[bitter orchid]

Local

Manual is the whole process

[bitter orchid]

I have never seen it on docker.

[frank ginkgo]

Model check ───────────────────────────────────────────────────────────────────────────╮
│ │
│ No auth configured for provider "anthropic". The agent may fail until credentials are │
│ added.

[bitter orchid]

Does the dashboard pop up?

[frank ginkgo]

nope

insta sent no model check

no auth configured for provider anthropic

[bitter orchid]

openclaw models auth paste-token --provider anthropic

[frank ginkgo]

paste-token = kimi token right?

[bitter orchid]

Your running local AI?

[frank ginkgo]

error: too many arguments for 'auth'. Expected 0 arguments but got 1.

[frank ginkgo]

im trying to run an AWS server w/ kimi 2.5

[bitter orchid]

Oh jesus

[frank ginkgo]

In my other instance, the setup was normal, but now it is disappearing with the model section lol

same installer, everything the same

[bitter orchid]

Firstly, your aws bill is going to be astronomical

And I'm unfamiliar with cloud deployment

[frank ginkgo]

So what I find weird is that before, the model option appeared in quickstart, but now it no longer appears

in a different instance in the case

[storm harness]

I'm having trouble logging into my account to access my agent. Everytime I try it says agent unavailable and when I try refreshing or restarting the agent it doesn't change anything. Has anyone experienced this before?

[wise rock]

Hey guys, I run openClaw in local with Qwen 3 8B, but it fails at calling tools reliably. WHich model do you guys run?

[frank ginkgo]

a dedicated server is the best way?

[rain relic]

@frank ginkgo I literally ran into this earlier. Seems to be an issue with 2026.2.13 release. See the link in reply. Starting a older version seems to work.

Uninstall openclaw
openclaw uninstall
npm uninstall -g openclaw

Then try the last release, you'll get some warnings about using old releases. That's fine right now.
npm install -g openclaw@2026.2.9

Then this should start the first setup prompts.
openclaw onboard

If docker, pull the last release

[broken mortar]

if anyone can help in #1472254740099235912 that would be very appreciated, thank you

[eager flume]

Hi everyone,

is it only just me or does Openclaw not work on tasks anymore reliable?
It says it will notify me if it is done but 5 hours later nothing happens.
After i ask what the status is it tells me again it contacs me with a summary what it has done after it is finished....

I updated today to the 2026.2.12 version.
Is this a bug?

[frank ginkgo]

lemme try ❤️ appareciate you shape

[frank ginkgo]

worked gg

[rain relic]

Awesome, glad to help

[proud crescent]

Last night after a complete uninstall, I freshly installed OpenClaw version 2026.2.13 and during the installation, it didn't prompt me for Model configuration. However, few days back when I had installed it, Model configuration was part of the installation. I eventually ended up using openclaw configure --section model to configure the Model I want to use but curious to know if something changed recently?

[meager cypress]

have you actually restarted the gateway

many (most?) of the configs don't hot reload

[teal oar]

been trying to fix You are not authorized to use this command., i have commands.allowFrom.discord set to "user:534859465648898071" in ~/.openclaw/openclaw.json

[split lion]

I have a likely related problem to @little blade I have a two tier system and default is Ollama. It works fine. I had to make sure it applies it to all jobs.

1. However, some nights something triggers and openclaw thinks Ollama is not available the it switched to Claude and burn a lot of token. When I go back to my Mac mini Ollama looks good but openclaw shows it as missing. Could this be an auto update?
2. I use iMessage and I assume the single thread makes it use a lot of tokens. I am planning to experiment with asking it to auto new the thread and rely on memory files.
   Appreciate all suggestions and pointers

[austere fjord]

everything has just stopped working. -_-
Imma reinstall it

[austere fjord]

why is it that openclaw always tries to use anthropic stuff:
C:\Windows\System32>openclaw models list

🦞 OpenClaw 2026.2.13 (203b5bd) — Shell yeah—I'm here to pinch the toil and leave you the glory.

Model Input Ctx Local Auth Tags
openai/gpt-5.1-codex text+image 391k no yes default,configured,alias:GPT

It gives this error in the gateway:
16:52:58 Embedded agent failed before reply: No API key found for provider "anthropic". Auth store: C:\Users\user.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\user.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir.

[fluid locust]

Hello all, anyone has the issue browser.act not working ? My OpenClaw can't click or anything in a webpage w/ peekaboo

[austere fjord]

#1472275845115805736
Someone help plz 🙏
This ai isnt very helpful

[amber path]

Any chance I can make multi-agents setup for diffrent telegram bots?

So each bot get own agent?

[cursive pecan]

I have a question. I have ollama/qwen2.5-coder:32b-instruct-q5_K_M model locally. But openclaw is saying it's missing. What am I doing wrong?

[hexed bay]

Are you sure you don’t have multiple gateways running without you knowing, via LaunchDaemon + LaunchAgent, and maybe even one started manually in verbose mode?
It happened to me and caused my buddy to hang all the time until I noticed I had three gateways running

[nocturne hull]

Agent failed before reply: All models failed (11): anthropic/claude-opus-4-5: Provider anthropic is in cooldown (all profiles unavailable) (rate_limit) | nvidia/moonshotai/kimi-k2.5: Provider nvidia is in cooldown (all profiles unavailable) (rate_limit) | openai-codex/gpt-5.2: Provider openai-codex is in cooldown (all profiles unavailable) (rate_limit) | openai-codex/gpt-5.3-codex: Provider openai-codex is in cooldown (all profiles unavailable) (rate_limit) | google-gemini-cli/gemini-3-pro-preview: Cloud Code Assist API error (429): You have exhausted your capacity on this model. Your quota will reset after 14h24m54s. (rate_limit) | google-gemini-cli/gemini-3-flash-preview: No available auth profile for google-gemini-cli (all in cooldown or unavailable). (rate_limit) | moonshot/kimi-k2.5: Provider moonshot is in cooldown (all profiles unavailable) (rate_limit) | perplexity/sonar-pro: 401 <html>
<head><title>401 Authorization Required</title></head>

anyone got this before? i have been getting it the last two days. my openclaw worked for a little while today but then i just started getting this error again. i have subscriptions to gemini, claude and openai and i've only used opus 4.5 today just for light chatting and troubleshooting. switching models and doing a gateway restart does not fix it, openclaw doctor --fix doesnt help either, restarting my vps connection doesn't help either

[blazing epoch]

Hi all. MS Teams - anyone using the plugin? I’m getting an error about duplicate plugin ids all the time. I’d appreciate any help how to fix this.

[cursive pecan]

google-gemini-cli/gemini-3-pro-preview: Cloud Code Assist API error (429): You have exhausted your capacity on this model. Your quota will reset after 14h24m54s.

Your claude code api quota was hit

As well as your other models are on cooldown as well

[proud crescent]

Last week I used Google's Antigravity OAuth to use Google models with OpenClaw since I have Google AI Pro subscription. However Google TOSed my usage and I can no longer use Antigravity Code Assist and Gemini CLI. I have already emailed [e-mail] requesting service restoration as I have realized my mistake but I haven't heard back from them. Has anyone managed to get a reply to their appeal and managed to get the code assisst and Gemini CLI service restored?

[half lance]

how good of a computer is needed to run openclaw

[wheat urchin]

hey, is there a way to switch auth profiles on the fly within a single agent/session?

i have two anthropic max subs (work + personal) and want to toggle which one im using without spinning up separate agents or telegram chats

something like a /profile command or per-session override would be sick

[median forum]

I have mine on a 2019 macbook pro which is kind of old but it runs great

[nocturne hull]

i get that is how it reads, but that can't be possible i havent used my other models for a single thing today, no subagents, no cron jobs, nothing

[north falcon]

when i start the onboarding i don't get to select any providers. the videos I have seen shows a list of providers that users can select from. Is it possible to use openclaw with a normal anthropic claude subscription or does it have to be API. when i select models in config ui (web or mac) I only get API selection posibillities

[rain relic]

@north falcon you might be seeing this issue

[austere fjord]

I had this problem.

Do this in your powershell:
openclaw configure --section model

[north falcon]

@austere fjord thanks, now i see that selection, thanks 🤩

[austere fjord]

np 🙂

[worthy ledge]

Thank you, Thank you, Thank you.

I think we should update the docs since I wasn't able to find that.

[austere fjord]

@mental mica idk if ur the right person to ping (i saw u giving help to someone earlier) but alot of people are having this problem

[umbral charm]

@low dragon ?? im guessing this isnt intended

[sour slate]

@austere fjord, please don't ping Peter for issues, use #1459642797895319552 or #users-helping-users if you need help, use #clawtributors to discuss PRs, or use any of the many other channels in this server as they're intended.

If you have a problem with the Discord specifically, use #report or DM @mental mica
-# Your message was reposted above without the ping active for the sake of conversation.

[worthy ledge]

I am hitting a new issue. when i send a message in the chat, the AI assistant loads for couple of seconds them sends an empty message. How to resolve this and make it respond?

[hollow ermine]

man I'm getting frustrated trying to run local reasoning models, I can't get any response in the openclaw chat interface, I've tried multiple Qwen3 and Ministral reasoning ggufs but none work

the only model that seems to work for me is Qwen3 4B Instuct 2507 Q4 K M

[molten gulch]

what's your issue?

what are you using to run the ggufs?

what did you try? what's the error you're getting, if any?

[unreal bear]

To people who run multiple persistent: do you run them all on a paid API

Or

one on paid API and the others on local models?

[dire orbit]

known bug where openclaw onboard doesn't prompt configuring api?

I setup a bot a few weeks ago, no issues.. brand new vm... and it isn't prompting me like it used to

[hollow ermine]

I run the ggufs on a dedicated ubuntu rig with a GTX 1070ti 8GB, I use llama-swap with llama.cpp to load up the models

@molten gulch I am connecting to the LLM just fine from openclaw, I see the request from openClaw chat go to the LLM and I see the LLM respond, but the response never shows up in the chat window

[pastel sierra]

hey, im new to the Open Claw stuff and i want to get set up but also be safe. im only 17 years old and i want to take advatage of the oppurtinity that ai is giving us too create side hustle ( for now ) and later on buisnesses. one day i want to become a software engineer or developer. if someone thats an expert and is genuine in helping me out please reach out. im down to get on a call and talk about it.

[unborn remnant]

I have a discord question: Is there a posibility to get a summary of todays thread? Like Questions that have been answered?

Besides a copy/past job into a model

[pure hollow]

i have set up OpenClaw a couple times now.. each times it seems to go through different steps... but i can't seem to get past the initializaition... it is kerfuffeling for 10+ minutes each time then i start over... should i wait longer this time?

[south cloud]

I'm running local LLMs as well; though through LM Studio. What I've notice is that if the Context Window is too high, I don't have enough GPU VRAM. With you having 8GB of GPU, use a small model that is like 4GB, then start with a low Context Length/Context Window (1000 tokens), test for a response, then increase from there until you don't get a response. Also, your desktop and other open apps will use a small bit of your GPU ram. I should mention that 1000 token might only be enough to says 'hello' and possibly get a response. On my 16GB GPU, when I run a 9GB model, my context window is only 30000. This is small and openclaw is constantly running the Compaction feature.

[void cargo]

You mean the bootstrapping process?

[hexed bay]

Humm do you have only one main agent or also some sub agents ?

[coral zodiac]

Hi, I'm trying to configure Kimi (Moonshot AI) in OpenClaw but getting authentication errors.

Setup:

• Platform: platform.moonshot.ai
• Active API key (sk-...) with credits
• Account tier: Tier0 with active limits

What I've tried:

1. Using kimi-coding provider with https://api.kimi.com/coding endpoint → authentication_error
2. Using moonshot provider with https://api.moonshot.cn/v1 endpoint → invalid_authentication_error
3. Created fresh API keys multiple times after adding credits
4. Verified key format (starts with sk-), no spaces, active status

Current config:
{
"models": {
"providers": {
"kimi-coding": {
"apiKey": "sk-...",
"baseUrl": "https://api.kimi.com/coding",
"api": "anthropic-messages",
"models": [{"id": "k2p5", ...}]
}
}
}
}

Question: What's the correct provider configuration for Moonshot AI API keys from platform.moonshot.ai? Should I use a different provider name or endpoint?

Olli gave the answer : openclaw configure --section model

[north falcon]

what would be the recomended model to get best quality for the bucks with openclaw. i read that OpenAI GPT models are on par or better than Opus 4.6 but cheaper, Im currently using sonnet-4.5 because i noticed that the tokens fly

[burnt lava]

Is there an option to change my model using OpenAI Oauth to standard chat gpt 5.2? Not codex? I have some weird enterprise subscription which doesn’t allow api keys -,-

[hollow ermine]

thanks, yes I think this may be causing the issue

[eager ridge]

Which ollama model is best for openclaw if i am trying to use openclaw as an construction company assistant and a sales/marketing assistant?

Thanks in advance 🙂

[south cloud]

I have been reading through the openclaw docs. There are ways to optimize token use and how many default tokens are used per call. It's a bit complicated, but I'm starting to understand it. So small GPU ram kinda sux, but won't put me out of the game.

[mental silo]

Yeah the previous version was a cinch to setup local providers

I just put the new version on a different box and defaults to Anthrophic ignoring the openclaw.json which point correctly at ollama

That auth-profiles.json doesn't even exist so where is the Anthrophic tag coming from?

Ok run openclaw configure

I sorted mine

[prime shell]

Anyone see this error before: 410 status code (no body)

[mental silo]

Ollama as local provider

[wraith sigil]

how to disable openclaw startup?

[mental silo]

Verify ollama working with a loaded model and then run openclaw configure and enter the model manually

It will then appear in the next selections and you select it

[hollow ermine]

can anyone see why my models dont show up in the agent drop down menu, here's the section for models from my json
"models": { "mode": "merge", "providers": { "llama-swap": { "baseUrl": "http://192.168.68.156:8081/v1", "apiKey": "llama", "api": "openai-completions", "models": [ { "id": "Reasoning_Alice", "name": "Reasoning_Alice", "reasoning": false, "input": [ "text" ], "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }, "contextWindow": 16768, "maxTokens": 4096 }, { "id": "Alice", "name": "Alice", "reasoning": false, "input": [ "text" ], "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }, "contextWindow": 32768, "maxTokens": 8192 } ] } } }, "agents": { "defaults": { "model": { "primary": "llama-swap/Alice" }, "workspace": "/home/openclaw/.openclaw/workspace", "compaction": { "mode": "safeguard", "reserveTokensFloor": 0, "memoryFlush": { "enabled": false } }, "maxConcurrent": 4, "subagents": { "maxConcurrent": 8 } } },

[mental silo]

You need to run openclaw configure and manually add your local providers and models

Then they will show up in the model list

[night prism]

Whenever you first install your openclaw and it opens your dashboard link, is there supposed to be anything on that page or it it just supposed to show that you can't reach the page?

[weak crane]

I'm burning through Claude Tokens incredibly fast. I'm using Sonnet 4.5 and now disabled Prompt caching and reduced all my cronjobs where possible:

Any tips on tuning claude for the best price/token usage?

[indigo nebula]

Re-Enable prompt caching that saves you money, first call is expensive but rest are lower but make sure you change it to 60 minutes not the default 5. https://docs.openclaw.ai/providers/anthropic#prompt-caching-anthropic-api
Also look at implementing QMD https://docs.openclaw.ai/concepts/memory#qmd-backend-experimental

[south cloud]

There is a lot of default tokens in each call to an LLM. There are ways to optimize openclaw to include less tokens when making a request to an LLM. The openclaw docs cover ways to send fewer tokens to a model and I'm just digging in to understand how to optimize openclaw.

[indigo nebula]

Also look at all of your cron and heartbeats. Ask your openclaw to change those to run on a free model. Does it really need opus to check if there’s a new email? Any cron jobs that are truly automated (no thinking, just give me the report) migrate to the system cron and run from a shell script. Ask your claw to build the script and set the cron in the OS.

[weak crane]

Thanks @indigo nebula and @south cloud I'm currently working my through switching options on/off until I find the best setup.

Do either of you run some sort of memory consolidation (using OpenAI Text embedding) I did this early on (recommended by my bot) but not sure if this is part of the issue.

[weak crane]

Thanks a ton

Also, any recommended length on the SOUL.md. I think I went overboard with mine...LOL

[south cloud]

For your main agent & any multiagents that you run, all the .md files in the /workspace directory get included - they can be 'trimmed' a bit by a property in the openclaw.json. The smaller those files are the less (default) tokens get sent. Subagents don't include all those files, bare bones sort of thing. Also, you might want to look into the Compaction feature of openclaw. I'm just starting to learn there are a few different ways to optimise token/context window use. Gonna take some trial and error.

[indigo nebula]

I’m not doing any memory consolidation. I told it to install & use QMD and search local for memories. I also gave it the following question: “I’m burning through tokens, what can we do to optimize and be efficient with calls to the LLMs? “ It made several recommendations like the ones I’ve shared and had it implement. I then gave it directives and said “Be efficient when making calls, batch commands when possible, when writing code or calling APIs if you can use the OS commands to from bash do calculations or gets the status of something for free, build the script don’t ask the model for the answer”. This helps some too, anything I can have it offload doesn’t burn tokens

[hollow ermine]

i made an email account for openclaw, I gave it the email address and password and told openclaw to set it up and send me an email. I dont think it really did it, it just keeps saying it did

[turbid bloom]

hey guys, facing a weird issue right now. My claw worked fine until i asked it to change it so i didn't need to mention it each time to get a reply (on Discord). It changed guilds.<myguild>.requireMention: false in the config, and when i restarted it, it didn't pick up any messages anymore, with or without mention, removed the channel link and tried to set it up again, still the same thing happened. can anyone help? Thanks

[south cloud]

I've done something similar. My main agent has been helpful 😀

[indigo nebula]

Pass it back to your AI, tell it “I think I went overboard on your SOUL.md, review it and propose a list of changes to shorten it without compromising on your vibe and identity. Make sure you are running on a good LLM model like Opus when you do this one.

[weak crane]

And here's the answe... based on the reference template: ~1,800 characters (~450 tokens)

Your current SOUL.md: 7,670 characters (~1,917 tokens)
You're 4.3x larger than the template!

LOL, I guess I went a bit overboard.

[indigo nebula]

If oyu haven’t install gogcli and follow all the instructions in the REAMDE. https://github.com/steipete/gogcli

[hollow ermine]

thank you

[indigo nebula]

One other thing you should do to save tokens. run /compact periodically to compact all the session when you’re chatting. And if you’re switching topics from doing improvements on openclaw to say writing some code, run a /new otherwise all that previous stuff keeps getting sent back to the model. I typically tell mine “I’m ready to switch topics, save all your memories and work then start a new session” and she’ll do it for me.

[gaunt charm]

I find it best to ask it to do things like this. Always pleasantly supprised it manages to. Avoid passing keys though to the chat.

[gaunt charm]

I will say Opus from Anthropic

[mental silo]

Anyone know why local models the bootstrap always shows pending in openclaw status --all

[flint onyx]

anyone is facing issue to use venice api?

[turbid bloom]

anyone got discord channels working without needing to mention the bot?

[frozen sparrow]

guys canot figure working with lm studio pls help i edited openclaw json but still canot get success
ppl have access to open all models and select local model i dont see this options at all when i start oboarding

[mental silo]

Run openclaw configure

Select the custom model and then enter the local model details manually

[frozen sparrow]

ye ye this all is work - pls add it to manual official

[dawn pollen]

Hey team, stuck with the "Error: Session file path must be within sessions directory" error, running latest version 2026.2.13.
Tried restarting it, no luck.
I'm a bit confused with everything going on in the Git issues and PRs, 13 was supposed to fix it, but it's still not working for me.
Any recommendations?

[frozen sparrow]

i did it bro but still not work after daemon restart

[mental silo]

If running ollama check the model is loaded and working via ollama first....

[frozen sparrow]

21:13:33 [diagnostic] lane task error: lane=main durationMs=36 error="Error: Unknown model: openai/gpt-oss-20b"
21:13:34 [diagnostic] lane task error: lane=session:agent:main:main durationMs=41 error="Error: Unknown model: openai/gpt-oss-20b"
00:13:34 Embedded agent failed before reply: Unknown model: openai/gpt-oss-20b
21:13:51 [ws] webchat disconnected code=1001 reason=n/a conn=43d6b811-1088-4674-a193-7e7fc748ad4f
21:13:51 [ws] webchat connected conn=cec53a39-2c23-4046-9b53-668c6b3cef6e remote=127.0.0.1 client=openclaw-control-ui webchat vdev
21:14:42 [diagnostic] lane task error: lane=main durationMs=26 error="Error: Unknown model: openai/gpt-oss-20b"
21:14:42 [diagnostic] lane task error: lane=session:agent:main:main durationMs=29 error="Error: Unknown model: openai/gpt-oss-20b"

[south cloud]

"models": {
"mode": "merge",
"providers": {
"lmstudio": {
"baseUrl": "http://localhost:1234/v1",
"apiKey": "lm-studio",
"api": "openai-completions",
"models": [
{
"id": "ministral-3-14b-reasoning", //the model you're using in lm studio
"name": "ministral-3-14b-reasoning",
"contextWindow": 30000, //this is important and goes hand in hand with LLMs size and you GPU ram
"maxTokens": 30000, //trial and error setting based on your GPU ram
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"reasoning": true
}
]
}
}
}, "model": {
"primary": "lmstudio/ministral-3-14b-reasoning"
},
"models": {
"lmstudio/ministral-3-14b-reasoning": {}
}

[frozen sparrow]

sec testing

[mental silo]

Yeah it's ignoring the openclaw.json

I had the same issue

Had to run the openclaw configure and I managed to put in the right port and model name

[frozen sparrow]

and how to fix it ?

[mental silo]

Openclaw configure

[frozen sparrow]

iam ther bro yes

model _ providers ?

right

[mental silo]

Run that command and only edit the model settings

[south cloud]

Dude, see my above reply to your first post

[mental silo]

Model auth provider > custom provider

[frozen sparrow]

i used it still not working

[frozen sparrow]

21:16:18 [diagnostic] lane task error: lane=main durationMs=14 error="Error: Unknown model: openai/gpt-oss-20b"
21:16:18 [diagnostic] lane task error: lane=session:agent:main:main durationMs=18 error="Error: Unknown model: openai/gpt-oss-20b"
00:16:18 Embedded agent failed before reply: Unknown model: openai/gpt-oss-20b

[mental silo]

Yeah it's hunting for the wrong model

[frozen sparrow]

its runned good after i put u setting but this error

delete openai/ ?

[south cloud]

Dependends on your model. Different model send and receive differently. If your using an anthropic model, that need to be your provider

[frozen sparrow]

no no i use openai oss

why i canot load here screnshot ?

00:16:18 Embedded agent failed before reply: Unknown model: openai/gpt-oss-20b
21:19:36 [reload] config change detected; evaluating reload (models.providers.lmstudio.models)
21:19:36 [reload] config change applied (dynamic reads: models.providers.lmstudio.models)
21:20:00 [ws] webchat disconnected code=1001 reason=n/a conn=e437aed7-7765-4511-8877-dee5d6b04da7
21:20:00 [ws] webchat connected conn=8e0de2f1-02ec-4b38-9dd5-072d280b378e remote=127.0.0.1 client=openclaw-control-ui webchat vdev
21:20:05 [diagnostic] lane task error: lane=main durationMs=20 error="Error: Unknown model: openai/gpt-oss-20b"
21:20:05 [diagnostic] lane task error: lane=session:agent:main:main durationMs=25 error="Error: Unknown model: openai/gpt-oss-20b"
00:20:05 Embedded agent failed before reply: Unknown model: openai/gpt-oss-20b
21:21:22 [diagnostic] lane task error: lane=main durationMs=39 error="Error: No API key found for provider "anthropic". Auth store: C:\Users\User.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\User.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir."
21:21:22 [diagnostic] lane task error: lane=session:agent:main:main durationMs=44 error="Error: No API key found for provider "anthropic". Auth store: C:\Users\User.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\User.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir."
00:21:22 Embedded agent failed before reply: No API key found for provider "anthropic". Auth store: C:\Users\User.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\User.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir.

[south cloud]

try removing the openai from openai/gpt-oss-20b

gpt-oss-20b

How much GPU ram do you have?

[frozen sparrow]

16 64

21:22:54 [reload] config change detected; evaluating reload (models.providers.lmstudio.models)
21:22:54 [reload] config change applied (dynamic reads: models.providers.lmstudio.models)

[south cloud]

16GB card and a 64GB card?

[frozen sparrow]

64 ram

[south cloud]

Of GPU ram?

[frozen sparrow]

no lol

gpu 16

vram

64ram

[south cloud]

Ok, look at the setting in LM Studio. Don't offload to system ram, just use GPU. Use a model of no bigger the 9GB and set your context window to about 25000-30000.

[frozen sparrow]

20b i use model

contex 32k

i dream about rtx 6000 )))))) and run 120b models

[mental silo]

Have a look at models.json in agents>main>agent

[frozen sparrow]

kk sec

[mental silo]

Make sure your local model is correctly defined there

[frozen sparrow]

{
"providers": {
"lmstudio": {
"baseUrl": "http://localhost:1234/v1",
"apiKey": "lm-studio",
"api": "openai-completions",
"models": [
{
"id": "openai",
"name": "gpt-oss-20b",
"reasoning": true,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 30000,
"maxTokens": 30000
}
]
}
}
}

[mental silo]

You should then be able to select it from the list of you use openclaw configure

[frozen sparrow]

ye ye list not working bro

[mental silo]

You tested the model locally in lmstudio?

[frozen sparrow]

openclaw fix this file themself after u change it in openclaw json

model working good

[south cloud]

20b parameters isn't the model size that fits into your GPU ram. If the model size is too big and the context window too big, openclaw won't run it.

Go by the size of the model on disk EG: 9GB

That will leave you some room for a context window

[frozen sparrow]

which one is betst for now ?

its not seding any api request

to my local llm server

[mental silo]

Oh I thought he must have a 32gb card haha

[frozen sparrow]

i dont see any req in terminal

[mental silo]

Use a model that neatlyn fits in the VRAM

[frozen sparrow]

i use gguf model

its working fast

very fast

for me

best of all what i used

12gb

[south cloud]

On my machine with 16GB of GPU ram, "name": "ministral-3-14b-reasoning",
"contextWindow": 30000, works ok for me. I'm still trying to figure out how to optimize openclaw to use less of the context winodw of 30000. openclaw usies a lot of context widnow by default.

[frozen sparrow]

how much use openclaw ?

[frozen sparrow]

becose its big agent thats why 🙂

to keep context window better

[mental silo]

I running on a 4060 I have the IBM granite 8B

It's responsive but my issue is that no matter what model I use it has no memory

Nothing is ever written to the workspace folder by any model I use

[frozen sparrow]

i think its big fail of devs to skip so easy option integration with local models :/

like i spend 4h and still canot fix it