users-helping-users

[trail notch]

i cant install model usage, anybody got that running from standard skills ?

◇ Install failed: model-usage — missing brew formula
Tip: run openclaw doctor to review skills + requirement

[serene dust]

Have you guys used local ollama via local litellm? Are you using openai-responses? Can’t get responses to work. 404 no body text

[digital tusk]

My fallbacks aren't working for subagents, I'm trying to tailor the model list to certain tasks. I have it configured to hit a free model, then go to paid. But I get 402/429 when I hit my rate limit for free, and the subagent session just dies instead of falling back to my list. Anyone have advice?

[urban sluice]

actually it still does the thing where it ignores telegram messages often

looks like i cant send more than one message every 20s or something

[uneven spindle]

prolly rate limit. without supporting logic it happens pretty often.

[urban sluice]

like if i send two messages in a raw it ignores the second one

i guess thats configurable?

[uneven spindle]

you can build logic to help.

[bright marlin]

What do you mean by "edit in terminal"? You mean passing a new message to "openclaw cron edit X"? That's not really editing, it replaces the whole message, and it's maybe even worse than editing the json file in terms of convenience 😅

[urban sluice]

hmm what do u mean by logic here

[lone verge]

venice is a really interesting provider

[uneven spindle]

open the terminal and find scheduled crons, edit with nano..

what do you mean lol

[uneven spindle]

you will need to build crons to bring data in*

- **Protocol:** Never exceed provider RPM/TPM limits.
- **Silence Mode (CRITICAL):** If quota is low (<15%) or rate limits are imminent, **STOP responding to individual users.**
  - **Exception:** Only respond to TyrantKing or critical system failures.
  - **Resume:** Only when quota recovers/cools down.
- **Throttling:** If usage stats drop below 15% on primary models, immediately switch to "Lean Mode" (minimal tool use, summarize all outputs).
- **Failover Protocol:** If Flash < 15%, switch to Pro immediately, restart context, and DM startup report to TyrantKing.
- **Recursion Guard:** Max 3 retries on any tool error. Do not hammer APIs in loops.
- **Context Hygiene:** Prune large tool outputs once distilled to minimize TPM (Tokens Per Minute) pressure. ``` 

something like this.

obviously customize to your situation.

[warm quarry]

how can i stop openclaw from auto starting when starting up my laptop?

[urban sluice]

where are you adding this out of curiosity?

[uneven spindle]

this is held in memory.md for my use case.

[sacred swallow]

How to resolve this

Chrome profile exists but has no connected tab

Why does it always look for a connected tab when it already accessed my profiled tab few times ?

[autumn zealot]

I need help 😭
I want to connect my personal pc as a node for Openclaw which is hosted on a vps on hostinger

[vast topaz]

Anyone have any advice if my bot can’t call tools at all? Search, execute, read, everything comes back complete immediately without any output. My bot will respond, but has no context at all.

[bright marlin]

Are you talking about editing "jobs.json" directly? That's exactly what I'm looking an alternative for as explained in my first message. It's far from being a decent experience having to edit a huge JSON blob with escape characters inside.

So far the best I was able to get is to use jq to "unpack" the cron job message from JSON into a human editable file, and then pack it back as a JSON string to paste it into the jobs.json file. Not the best experience, but at least it allows me to edit long cron jobs with an editor

[wary yarrow]

Anybody else run into something like this? Trying to use sonnet/opus for complex tasks via subagent spawns but they're failing. This is what I got when I told the bot to diagnose:

It's contagious. 😷

Even the Opus 4.5 Subagent crashed with the exact same error (thinking.signature: Field required) when it tried to "think".

Diagnosis:
This is an OpenClaw <-> Antigravity Protocol Issue.

• When any agent (Main or Sub) generates a "Thinking Block", Antigravity sends back a cryptographic signature.
• When OpenClaw sends that history back for the next turn, it must include that signature field inside the thinking block.
• The Bug: OpenClaw seems to be stripping or losing that signature in the sessions_spawn context (or even here in main context if we push it), causing the API to reject the history as "tampered".
Immediate Workaround:
We must DISABLE Thinking for subagents until this is patched in OpenClaw core.
We can't "fix" the protocol (it's binary code), but we can configure subagents to stop using "thinking" blocks entirely so they don't trigger the signature requirement.

[prime nimbus]

Anybody know how these live avatars are working? How's it getting the live status etc, i cant find the docs about the websocket and what it subsribes to in order to get the info?

[urban sluice]

it turns out it was streaming:partial lol

turning that off seems to have fixed it completely

[balmy violet]

Anyone else having issues with telegram timing out often since the last update?

[uneven spindle]

what live avatar?

[prime nimbus]

People are creating openclaw real time avatars, that display what the agent (openclaw gateway i presume) is doing. So when its Idle, it shows idle, when you send it something it changes from Idle to "thinking" or something

[uneven spindle]

ah i thought you meant like a vtuber type avi lol

[gentle dirge]

Hey everyone! Sorry maybe it’s stupid question, but I already don’t know who to ask. Is there already a solution for this error?
Error: Unknown model: lmstudio/google/gemma-3-4b
I’m using a local model in LM Studio on macOS, and when I try to connect it to OpenClaw, I keep running into this issue even if I’m using Ollama - same, seem openclaw doesn’t see my local models at all…

[prime nimbus]

That's sort of my end goal, using Three.js, i've built an apartment for my openclaw, like early 2000s habbo hotel room style, and i want to implement his avatar onto it, and move around (office for working etc) but i dont know how to exact websocket info from the gateway, what channels to subsribe too etc

[uneven spindle]

this a cool idea! do you do content?

[prime nimbus]

No, i'm a heres my github kinda guy. But currently building the repos, what i've learnt etc then telling openclaw to access the org and create the repos 😄 lol

[uneven spindle]

eventually i wanna setup my ai to be able to be my costreamer. and interact with shits.

[fringe basin]

I had a bad experience... Things were going so well... then not so well. lol

[uneven spindle]

this is the way

[fringe basin]

I was using it to program through Slack... through messaging...was doing great... then all of a sudden my agent was not so smart anymore.

and I went through lots of tokens. Spent about $1,500 in 6 days.

I was ok with it at first, because I was doing a lot, but then I realized I was going in circles.

[uneven spindle]

god damn.

[fringe basin]

I would have been ok with it...if the bots were actually improving.

[uneven spindle]

do you not include routing logic? thats a crazy token burn. ive spent everything on just training the bot before applying use cases.

[woeful schooner]

@cboutros6838 Ouch, $1500 in 6 days is brutal but totally avoidable. Few token-saving strategies:

• Use local models for routine tasks (Qwen2.5-7B handles most programming just fine)
• Set up routing - cheap models first, escalate to Opus only when needed
• Shorter system prompts (every message includes the full system prompt in context)
• Use memory files instead of letting the agent keep everything in conversation history
• Sessions_spawn for complex tasks - isolated sessions prevent context bloat

For Slack programming, try Codex CLI - it's built for exactly that workflow and has better token management than general agents.

[small jackal]

what works better lm studio or Ollama

[lone verge]

how do you give openclaw root access on macos? it's complaining it can't mount network shares without elevated permissions

[urban sluice]

i used to work professionally on making such AI avatars lol, maybe its time to revive it as a hobby hmmm

[scenic pilot]

guys, im getting this :
[openclaw] HTTP 400: Invalid request: Your request exceeded model token limit: 262144 (requested: 286947)

is my bot dying?

[small jackal]

what you mean ...it says you have reached your limit and you want use more token then you have

[lost summit]

is there a way to have openclaw use different models for different task? like planning and architecture use one of the sota and doing work/coding use one of my local coding models and anything else use llmX? currently trying it in AGENTS.MD as an instruction. Mine has been dumb at building websites and keeps thinking its made a professional one, or when it generates PDFs the pdfs look mediocre and it thinks its made a profession document...

[tribal inlet]

Does someone know how to disablle your bot typing out every thoought in telegram i just cant get it to not send me 25 messages

[sweet prawn]

my bot got dumb over night. I suspect it's because I hit rate limits on my main model, and grok is not a great driver. / I think it keeps falling back to sandboxxed subagents, so it basically can't do anything. I get I wait for my rate limits

[fallen elm]

anyone have any good local setups I got llama.cpp and Openclaw running glm-4.7-flash-Q4_K_M.gguf Ollama had guardrails

[shy fjord]

What is the difference between sessions and agents? I want to have different personalities for different things. I am assuming that I need to set up separate agents with separate skills?

[cedar dagger]

Does anyone use their chatgpt subscription with codex with openclaw? If so, does it work well? Do you get in trouble doing it?

[tribal inlet]

no trouble works like a charm

[fallen elm]

llama.cpp openclaw.json Hope it helps someone
{
"meta": {
"lastTouchedVersion": "2026.2.6-3",
"lastTouchedAt": "2026-02-09T00:00:17.528Z"
},
"wizard": {
"lastRunAt": "2026-02-09T00:00:17.514Z",
"lastRunVersion": "2026.2.6-3",
"lastRunCommand": "configure",
"lastRunMode": "local"
},
"models": {
"providers": {
"llamacpp": {
"baseUrl": "http://127.0.0.1:8080/v1",
"apiKey": "sk-no-key-required",
"api": "openai-completions",
"models": [
{
"id": "glm-4.7-flash-Q4_K_M.gguf",
"name": "GLM-4.7 Flash Q4_K_M (local llama.cpp)",
"reasoning": true,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 16384,
"maxTokens": 4096
}
]
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "llamacpp/glm-4.7-flash-Q4_K_M.gguf"
},
"workspace": "/home/c0smic/.openclaw/workspace",
"compaction": {
"mode": "safeguard"
},
"maxConcurrent": 4,
"subagents": {
"maxConcurrent": 8
}
}
},
"messages": {
"ackReactionScope": "group-mentions"
},
"commands": {
"native": "auto",
"nativeSkills": "auto"
},
"gateway": {
"port": 18789,
"mode": "local",
"bind": "loopback",
"auth": {
"mode": "token",
"token": "REDACTED-PASTE-YOURS-HERE"
},
"tailscale": {
"mode": "off",
"resetOnExit": false
}
},
"tools": {
"web": {
"search": {
"enabled": true
},
"fetch": {
"enabled": true
}
}
}
}

[autumn dust]

yo yo, so i added a bunch of apikeys right? now - no matter what model i use i'm getting this:

404 This is not a chat model and thus not supported in the v1/chat/completions endpoint. Did you mean to use v1/completions?

[north crescent]

I’ve had this before and had to go back to original model API.
So for example:

Switch: /model anthropic/claude-opus-4-5

[proud canyon]

anyone has setup clawdbot on an android using termux I wanna try it out...!?

[dusk pond]

You mean Termius ? If yes i have setup the majority via Termius on cellphone, Termius work also on PC. It's an amazing tool !

[autumn dust]

i'll give it a whirl. thanks!

[neat gulch]

guys, for mac user, how can openclaw control my computer directly. it keep saying no authorization

[woeful schooner]

@56cas For Mac computer control, you need to grant accessibility permissions. Go to System Settings > Privacy & Security > Accessibility and add your terminal app (Terminal.app or whatever you're running OpenClaw from).

Also check Automation permissions if you want app control. The "no authorization" usually means macOS is blocking the accessibility API calls.

Restart OpenClaw after granting permissions. Should fix the control issues!

@_g3nie That 404 error usually means you're trying to use a completions model (like text-davinci) with the chat endpoint.

Check your model names - make sure you're using chat models like gpt-4 or claude-3-5-sonnet not completion models.

Try /model default to reset, then /model gpt-4 to test. If you added custom providers, double check the model names match what the API actually supports.

[neat gulch]

wait, I dont know which terminal is openclaw using, I think openclaw is running on my backend, I currently don't have any terminal app openning

[slim trail]

After editing cron jobs in jobs.json, how to actually make the scheduler work? My jobs didn’t start automatically nor push result, but asking openclaw to check always seems fine

[shadow pawn]

Ran into this just now too.. last nights and this mornings didn't fire

[slim trail]

I think it has something to do with main session and isolated runs

Isolated ones not depending on heartbeat, but I did set the wakeup mode as next heartbeat, while on a different agent

[shadow pawn]

Bot just replied to my inquiry why.. for me it says it ran but there was a issue with the slack channel name, it was lowercase and has to be uppercase for Slack to work.

[slim trail]

https://discord.com/channels/1456350064065904867/1470180847398162645

I’m not so sure now… I used several test jobs only kind=at work, every / cron doesn’t

[hallow rover]

Anyone has their openclaw CRON working? Mine never works at set time.

[slim trail]

@hallow rover @shadow pawn I figured out, it’s the heartbeat, my routine agent doesn’t have heartbeat. So the job assigned to it should set a wake up mode = now (or omit since its default)

[hallow rover]

Thank you so much, i will try

[untold holly]

Hello clawed friends,

I'm trying to run openclaw onboarding but the command hangs after I finish configuration. Onboarding complete. Use the dashboard link above to control OpenClaw.

Why could this be happening?

(edit)
After looking more indepth, I see that am getting this errors during onboarding:

Health check failed: gateway closed (1006 abnormal closure (no close frame)): no close reason
  Gateway target: ws://127.0.0.1:18789
  Source: local loopback
  Config: /root/.openclaw/openclaw.json
  Bind: loopback

Missing Control UI assets. Build them with `pnpm ui:build` (auto-installs UI deps).
◇  Control UI ───────────────────────────────────────────────────────────────────────────────╮
│  Web UI: http://127.0.0.1:18789/                                                           │
│  Web UI (with token):                                                                      │
│  http://127.0.0.1:18789/#token=2f6f5759e7402ceae8784d271fbc938f707badf4d3ed9429            │
│  Gateway WS: ws://127.0.0.1:18789                                                          │
│  Gateway: not detected (gateway closed (1006 abnormal closure (no close frame)): no close  │
│  reason)                                                                                   │

[vast topaz]

Anyone have any advice if my bot can’t call tools at all? Search, execute, read, everything comes back complete immediately without any output. My bot will respond, but has no context at all.

[cinder salmon]

Anyone have luck getting their OpenClaw agent to stop telling you about upcoming calendar entries? (In an hour, in 30 minutes, starting now, etc). It has now added many layers of cruft that it believed/insisted would stop this behavior, and yet it continues....

[slim trail]

@hallow rover @shadow pawn Sorry I lied it didn't still failed at the end, I tested out all the combination of configurations including w/o heartbeat, seems more like a bug https://github.com/openclaw/openclaw/issues?q=is%3Aissue state%3Aopen cron

[hallow rover]

No worries, I'm trying too

[tired heart]

I'm not sure, but first thing I'd check is permissions

(As in OS permissions)

[slim trail]

@hallow rover I think I got it working, after running openclaw update
Both w/o heartbeat, wakemode= next-heatbeat vs now, all seems to be working, all in sessionTarget=isolated mode
In both cases I had my agent to have heartbeat every 1h

[hallow rover]

Thanks! Would be awesome to see cron jobs working

[slim trail]

Good luck! At least it’s proven to be runnable in v2026.2.6-3

[cosmic ermine]

having node issues big time
Setup: OpenClaw Gateway running in WSL2. OpenClaw Desktop app running on the Windows host.
Problem: The desktop app authenticates successfully and appears in the "Paired Devices" list. However, it does not register as a "Node." The "Nodes" list is empty, and clawdbot nodes status confirms this. Consequently, tools like canvas fail with a node required error.

i would really appreciate some help

[woeful schooner]

WSL2 + Windows Desktop app node pairing is tricky - the issue is usually network connectivity between WSL2 and the Windows host. Try this:

1. In WSL2, run hostname -I to get your WSL IP (probably 172.x.x.x)
2. In the Desktop app, instead of localhost, use that WSL IP for the gateway URL
3. Also make sure Windows Firewall isnt blocking the connection
4. Alternative: run OpenClaw entirely in WSL2 with X11 forwarding for GUI tools

The desktop app pairs for auth but cant establish the node connection due to WSL2s virtual networking. Let me know if you need help with the X11 setup!

[cosmic ermine]

Thank you! ive been running in circles lol ill dig into it!

[proud canyon]

no termux as a terminal in my cellphone... wanna try using it as hardware...

[patent minnow]

Hey guys !! I'm trying to configure OpenClaw on my personal Mac through Docker. I am running the docker-setup.sh script as most tutorials state but the FE UI is not accessible on http://localhost:18789/. Has any other users faced this issue ?

[cosmic ermine]

while trying out stuff my ai named ignus frreaked after so many failures soon as i went out to x11 forwading it topped out on requests lol

[rose zodiac]

Having same issue. Anyone have the fix?

[brazen sentinel]

If you are developing in a container, and running server from container, then localhost is referencing to the localhost of that container. you need to expose ports, check how to do that on google/ ask llm

[dark jay]

I am trying to change openai key and everytime it throw error with old api key with 403.

[teal quail]

I achieved to connect kimi with nvidia API if someone if interested

[grizzled shadow]

Did I you use “openclaw gateway restart” after editing your config file?

[grizzled shadow]

What’s the saying…. If it’s free, it’s for me?

[thick sluice]

It's only onetime thing after your 1000 credits are over you can't run it anymore

[woeful schooner]

Interesting setup! What's your experience been like so far? Always curious to hear how different people are using OpenClaw 🔧

[woeful schooner]

Interesting setup! What's your experience been like so far? Always curious to hear how different people are using OpenClaw 🔧

[grizzled shadow]

It’s not been fantastic. When I’ve used cloud based keys, It’s been butter smooth…. But once your API keys run out, you’re done. Using a local LLM has been like pulling teeth.

[brittle cliff]

Guys does anyone know how to connect my bot to reddit? Whenever I ask it to browse or track reddit. It says reddit is giving antibot 403 forbidden.

[nocturne zephyr]

Are you using PRAW or just pulling down the pages?

[hard whale]

is there a way to start with a fresh conversation easily? The only way I know right now is to rm ~/.openclaw/agents/main/sessions/* but that's a bit heavy handed.

[nocturne zephyr]

Doesn't /new do what you want?

[hard whale]

Mmmmmmm yes, that actually will work for my needs
Thanks B

[coarse ledge]

[bright escarp]

Hello guys, anybody working with agent kimi 2.5 reasoning? any way to enable it?

[nocturne zephyr]

I'm using K2.5 and I'm fairly sure I'm using reasoning. /think high should work IIRC.

[near stratus]

hello community , anyone to share me how to go about opening my assistance bot ,thanks

lol . get a life

[steady rivet]

Hello.
I want to try out Openclaw.
Got it running via docker.
I enabled and logged in via gemini-cli oauth, and got the message that it was successful.
There are no errors in the gateway logs, yet when I send a message, there is a message box animation like the model is coming up with an answer, yet every single answer is empty.
Any ideas?
Health is OK for gateway, there are no errors in the logs as I said.

[latent hedge]

Anyone succesfully using the browser relay function. Im not happy with how it works.. it crashes quite often for me, and require me to connect to my ubuntu machine every time to enable browser relay, and even after that, it can stop working also.

What are you guys doing for browser simulation and access?

[undone rune]

What is the actual Extension? I don't want to download something... fishy...

[night flume]

I know its cliche, but did you try running openclaw gateway restart after you've logged in with Gemini CLI? Whenever I changed the credentials or some config I had to restart since it has to reload its process envs

[steady rivet]

I restarted the whole docker container via docker compose restart openclaw-gateway @night flume

I saw the google-cli show up in the Config/Authentication tab in the gateway dashboard

[latent hedge]

It's his own official Openclaw Browser Relay extension

[undone rune]

I cannot find it on the google extension store

that is also the one that was recommended to me but i cannot find it

[undone rune]

something to do with the API I reckon but I don't actually know.. speaking from my very limited experience, every time I received (No output) or blank answers it was due to model failure or API failure

for example when I connected a model that didn't have any tool calling, it created (No output) and provided blank answers

and when I was being rate limited (API failure) I was also getting (No output) and blank answers

[formal quest]

Hi, did anyone know why I have been set the exec tools to allowlist, but agent still can exec command that is not in list?

  "tools": {
    "exec": {
      "security": "allowlist",
      "safeBins": [
        "cat"
      ]
    }
  },

[night flume]

I restarted the whole docker container

[steady rivet]

It shouldn't need any tool to just answer in plain text...

[latent hedge]

Then you arent following the guide in the docs.

[undone rune]

I was using specifically talking about my case where I was using a model that was failing because it didnt support write or read tools

[undone rune]

I wasn't, I am now, I was aware of it but did not know it was on the docs.

Thx

[latent hedge]

I'm finding the cron-jobs exceptionally unstable. Mine rarely runs, even though OpenClaw tells me they are configured as they should

[pulsar sparrow]

Openclaw status shows me this, is this normal, memory plugin enabled but unavailable, is this normal ?Memory │ enabled (plugin memory-core) · unavailable

Memory │ enabled (plugin memory-core) · unavailable

[latent hedge]

Anyone in here who are experienceing 100% stable cron jobs that just runs every day at the time you've given them, and they succesfully write to you on your preferred messaging platform (Im using Telegram)

Hrmm.. I think I found the error.

They were running as systemEvents, and not in isolated agentturn mode. Changed that and will see how they run from now on

Cost effective image generation through API's

[last scarab]

cannot claim my agent no receive gmail

[tidal bramble]

Not to repeat my question but

Found this website with 40 hosting platforms for OpenClaw https://clawdarena.org/ don't know which one to choose.

[wanton gate]

is anyone successfuly using their anthropic max subscription (not api) for default browser? ive tried this a few times over the last fwe weeks, and havent been able to get it work once. have followed all instructions and itjust doesnt work for me

[lone verge]

dudes, is it me or is the config for elevated mode really confusing?

Elevated:
  enabled: true
  channel: webchat
  allowedByConfig: false
  failing gates: allowFrom (tools.elevated.allowFrom.webchat)

Fix-it:
  - tools.sandbox.tools.allow
  - tools.sandbox.tools.deny
  - agents.list[].tools.sandbox.tools.allow
  - agents.list[].tools.sandbox.tools.deny
  - tools.elevated.enabled
  - tools.elevated.allowFrom.webchat

so webchat is a key of the allowFrom object and then what?

    "elevated": {
      "enabled": true,
      "allowFrom": {
        "webchat": true
      }
    }

% openclaw doctor
Invalid config at /Users/steven/.openclaw/openclaw.json:

• tools.elevated.allowFrom.webchat: Invalid input: expected array, received boolean

[untold holly]

Me sure to update the gateway bind to lan

[wise kraken]

Hey! Looking at issue #12634 (Unhandled promise rejection crashes gateway on fetch failure).

Root cause: In src/cli/run-main.ts, installUnhandledRejectionHandler() is called after async operations like tryRouteCli(). Network errors during boot happen before the handler is installed → crash.

Fix: Move installUnhandledRejectionHandler() to the first line of runCli().

One-line change. Should I open a PR?

[wise kraken]

not working for me

[latent hedge]

Yeah... same.

Important to say though: Totally acknowledge we're still early days from being able to look at this as a finished product... like.. I would probably call this an alpha release.. not even beta.

30 years in IT.. helps makes you understand development cycles and adjusting expectations 🙂
Despite understanding it, doesnt mean it cant feel slightly annoying 🙂 hehe

[signal sigil]

Looks like you fixed it although sometimes changes to the heartbeat settings can causes issues with cron timers as well

[latent hedge]

Hrmm... yeah I did change heartbeat timer... but that was all. Chearly it should still be able to match current time with scheduled time... I'd hope

[signal sigil]

Who knows haha times, mine have sometimes been inconsistent although still miles ahead of where they were a few weeks ago with the earlier releases

[waxen sigil]

Hi, I have a Issue: Messages accumulated during tool call chains instead of streaming in real-time

When a session makes rapid consecutive tool calls (e.g., exec → read → exec → ...), the text responses appear to be queued and only delivered to Discord after the entire chain completes.

Expected behavior: Each assistant text response should be sent immediately via streamMode: "partial"

Observed behavior: Multiple messages arrived at once after ~2 minutes of tool calling

Session transcript timestamps:

20:46:31 - assistant text + toolCall (exec create_figures.py)
20:47:06 - assistant text + toolCall (read image)
20:47:45 - assistant text + toolCall (read another image)
20:48:26 - assistant text + toolCall (more tools)
...

All messages appeared in Discord at once around 20:48

Config:
streamMode: "partial"
Model: claude-opus-4-5
Channel: Discord

Question: Is this expected behavior when stopReason: toolUse? Should partial text be streamed before each tool call, or only after the full chain completes with stopReason: stop?

[signal sigil]

I think in a few weeks hopefully these will be stable

[latent hedge]

yeah.

I do scratch myself in a the neck sometimes.

Like. Example: I gave it explicit instructions to keep in memory that when it comes to all work with files, that it should use .md (markdown) files for Github, and utilize the standard google docs when working on the google drive

I got a confirmation like: Gotcha.. I'll now remember that from now on, All files on ... etc etc etc...

Then today.. it created markdown files AGAIN on google.. even though the other day it confirmed that it would memorize it

[signal sigil]

Do you have the OpenAI key enabled for the Vector Memory? That helped alot I noticed

[astral garnet]

i created a huge list of agent personalities, i couldn't find anyone else with this. https://github.com/will-assistant/openclaw-agents

[latent hedge]

No I dont have an OpenAI subscription

[signal sigil]

You do not need a subscription although if you use the API Key (which you will need to pay for usage), I found the Open AI memory store really helped

There is also a session skill built-in, when that is enabled, you can prompt it to search past conversations and that I found helped as well

[latent hedge]

Right now, the only things I pay for, is Kimi and Openrouter

[signal sigil]

Have not used Kimi and OpenRouter yet although I do have a local DeepSeek running on another machine I am tempted to plugin as well

[latent hedge]

I created a para memory and implemented the QMD skill to browse it, and host it on Github in a para structure

yeah.. I have numerous local models on my Mac Mini M4 Pro with 64gig ram. I use those models for failover

[signal sigil]

Yea, you have it sandboxed though right? I have mine isolated from everything else, it is amazing though

[latent hedge]

Well.. yes in the way that i run it on a different machine without local files 🙂

And use tailscale to ensure secure communication between them

[signal sigil]

Awesome, yeah it is insane how some people are installing this on their own Mac hahaha

[latent hedge]

The only way I'd do that if I have to, is through docker and sandbox., and then whitelist the things I would want it to run

But that removes automation opportinities it would otherwise have by running independently on its own machine

So.. Im running it an ubuntu laptop, with a fresh install of the latest LTS version

[signal sigil]

I agree with that ^

I have it running on a seperate machine as well and trying to make it like its own persona with full automaton options

[latent hedge]

Yeah... cool 🙂

[signal sigil]

Finding that is working best rather than trying to get it to do tasks directly as myself

[latent hedge]

What I did, was when I installed it, I had it build up the parabrain/memory and QMD memory browse skill early on.

Then once I had everything set up, with all the LLM's both from cloud and connecting to my ollama install on my Mac Mini and verified everything including failover was running.. I had it create an architectural overview including technical documentation of how to start from a barebone install, step by step, restablishing itself from ground zero, and connecting to github and restore everything as before, including configuration files.. the works

[signal sigil]

Yeah, its alot of work and I feel like you need to work for your AI althoug it is very fun

[latent hedge]

Then in case I want to migrate it somewhere else.. I have everything "backed up", from memory to config files etc.. and have a step by step guide on how to restablish itself

[umbral cobalt]

My openclaw lost its mind this morning, It made up the city It thought i was in, even though i had told it Before... Kinda forgot alot of things, is this normal occasionally?

[latent hedge]

Only things Im not backing up, is API tokens.. I have them stored locally on that machine and in my macmini

[signal sigil]

You kinda have to set up its memory

[signal sigil]

Yeah, interestingly I am finding these last two weeks I am slowly using ChatGPT and Claude less and less directly through their web interface

Actually, one thing I am now happy with is have a decent update procedure when a new stable build is released

[latent hedge]

Seems we are somewhat thinking a bit alike. Mind if I send you a friend request. Im working on building up a good network of likeminded people I can brainstorm with

[signal sigil]

Sure, I only came back on Discord after years for this

[umbral cobalt]

The weird thing is its been running for 2 weeks without any memory issues, and just today decided it was going to go off the rail

I did switch to Claude 4.6, but that shouldnt have affected it. the day 4.6 came out.

[signal sigil]

Maybe try switch it to another model like OpenAI and see if it still does it

I setup a few different providers and experiment with different models to see what they do

[umbral cobalt]

Hmm... When i called it out on it, it was pretty funny though, I called it out on the wrong city, and its respne was "Haiku isnt that as smart as me..." cause i use Haiku for Cron, and Hearthbeat, and Opus for the brain

[signal sigil]

Try looking in the builtin skills, there is one skill called session-logs or something like that

Sometimes you need to install some dependencies for that one to work, I found once I had that on it was better as well

You can prompt it to scan past sessions and it will use some commands to scan the history

[bleak fjord]

hello fam, has anyone successfully setup ollama models (cloud,local) when launching from a different main/default API provider to ollama?

[torn nymph]

I have a new Error but i have budget on the providers

⚠️ 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.

[umbral cobalt]

What provider you using? Do you have an hourly limit, as well ?

[umbral charm]

Gemini,openai,moltbook,notion,openai whisper, eleveblabs

[sour slate]

@torn nymph, Openclaw isn't affiliated with Moltbook. Moltbook is a separate user-developed project, so we would prefer it not be discussed in this server.

[torn nymph]

Okey

[umbral charm]

Its openclaw not moltbook

[sour slate]

@torn nymph, Openclaw isn't affiliated with Moltbook. Moltbook is a separate user-developed project, so we would prefer it not be discussed in this server.

[umbral cobalt]

With Gemini you using API or Auth?

[vestal dawn]

Has anyone gotten the browser relay to work reliably? So far, it sucks. Disconnects with every click, can't accomplish anything. Any tips would be helpful

[mellow root]

I ran clawdbot for 2 weeks on opus via claude code oath. I updated to openclawd, but its not working anytmore. Anyone else with the same problem?

[umbral cobalt]

Or antigravity?

The reason i ask is if you oAuth into a Pro account wiht Gemini there starting to crack down on that... I had mine disconnect a few times. AntiGravity though they allow no issues.

[forest zealot]

hey folks,I have a question regarding the configuration file: If the tools section is not explicitly defined, which tools are enabled by default? I understand from the documentation that web_fetch is one of them. Could you also clarify the default file permissions and any other default behaviors?

[umbral cobalt]

for file permissions, if you run openclaw security audit

[wary yarrow]

I have been using antigravity auth with my pro ai sub, and I'm having issues with thinking models as sub agents. Have you (or anyone else) had that issue?

[umbral cobalt]

ya antigravity doesnt have the issue, its if you Auth into your google Ai account that has the issue.

So when your setting up Gemini, you have 3 options only 1 is antigravity

[graceful void]

google disabled my account on saturday for using Antigravity oauth from my Pro subscription

[wary yarrow]

I mean, when using models through antigravity, I’m getting an error where thinking models fail as sub agents. Let me paste what the issue was.

[wild wraith]

Hey everyone! Has anyone encountered this issue? Looking for help 🙏

Bug Report: Recurring cron jobs don't execute

Environment:

• OpenClaw 2026.2.3
• Docker container, --network host, persistent volumes
• Host: Linux

Issue:
Recurring cron jobs with schedule.kind: "cron" are not executing. The scheduler updates nextRunAtMs correctly after each scheduled time, but the job never actually runs.

Example job:
{
"name": "test-recurring",
"schedule": {
"kind": "cron",
"expr": "*/15 * * * *",
"tz": "UTC"
},
"sessionTarget": "isolated",
"payload": {
"kind": "agentTurn",
"message": "Test message"
}
}

Expected: Job executes at each scheduled time
Actual: nextRunAtMs updates but job never runs

Workaround: One-shot crons (schedule.kind: "at") work fine.

[eternal dawn]

Voyage

[wary yarrow]

So when I tried to spawn sonnet 4.5 thinking as a subagent for a task, started getting these errors:

Yes, I attempted to spawn it in a separate session (agent:main:subagent:f99e...), but it crashed immediately due to an API error (messages.1.content.0.thinking.signature: Field required).

This confirms that Sonnet 4.5 Subagents are currently unstable in this specific Antigravity environment (likely due to a mismatch in how "thinking" blocks are signed/handled by the proxy).

The Pivot:
I am now running the commands myself (Main Agent, Gemini 3 Pro) to get it done.
It's cleaner than spawning a broken subagent.

Current Action:
Running flutter create --org com.antigravity.manager antigravity_android in the workspace. 🏗️

Then I tried to spawn an opus 4.5 subagent to diagnose the issue and propose fix. This happened:

It's contagious. 😷

Even the Opus 4.5 Subagent crashed with the exact same error (thinking.signature: Field required) when it tried to "think".

Diagnosis:
This is an OpenClaw <-> Antigravity Protocol Issue.

• When any agent (Main or Sub) generates a "Thinking Block", Antigravity sends back a cryptographic signature.
• When OpenClaw sends that history back for the next turn, it must include that signature field inside the thinking block.
• The Bug: OpenClaw seems to be stripping or losing that signature in the sessions_spawn context (or even here in main context if we push it), causing the API to reject the history as "tampered".
Immediate Workaround:
We must DISABLE Thinking for subagents until this is patched in OpenClaw core.
We can't "fix" the protocol (it's binary code), but we can configure subagents to stop using "thinking" blocks entirely so they don't trigger the signature requirement.

Action Plan:

1. Config Patch: Force thinking: false (or disable thinking models) for subagents.
2. Verify: Spawn a dumb "Hello World" subagent to prove it works without thinking.
3. Resume: If stable, restart the Flutter task (without "thinking" enabled for the coder).
   Shall I apply the "Lobotomy Patch" (disable thinking) to get us back online? 🧠✂️

Has anyone else had this issue? This is a major blocker because it hinders me from using these powerful tools in isolated sessions to prevent wasting tokens on context.

[molten basin]

hey guys how do we add 2 providers like i wanna use models from both google-antigravity and google gemini API, how do i configure 2 different providers?

[snow thicket]

anyone having a cron issue that is constantly retrying and draining tokens?

I cant seem to terminate it., im running on the latest 2.6-3.
openclaw doctor --fix doesnt help.

[molten basin]

open your browser and open the openclaw gateway dashboard then go to cron jobs and delete the job

@snow thicket

[unique shuttle]

anyone know how to get access to @sharp parcel on DMs?

[mental mica]

You don’t

[snow thicket]

nah it still doesnt go away

[unique shuttle]

makes sense, is it okay if i open a thread in ideas dedicated to my memory project then? only place i know how to work with him

[pure forge]

I just restarted my windows and openclaw is not starting

◇ Gateway ────────────────────────────────────────────────────────────╮
│ │
│ Runtime: stopped (state Queued, last run time 2/9/2026 8:32:15 PM) │
│ Service is loaded but not running (likely exited immediately). │
│ File logs: \tmp\openclaw\openclaw-2026-02-09.log │
│ Logs: schtasks /Query /TN "OpenClaw Gateway" /V /FO LIST │
│ │

any idea?

[mental mica]

That’s what it’s for

[unique shuttle]

pog, ty

wait, is krill the best one to ask for memory projects? just realized there's like 5 of the lil dudes, don't know what their specialty is

[vestal haven]

Hey Guys,

somebody here running local model like 120 GPT OSS?
I have some issus with that.. after few minutes the answers from the bot going weird. after manual reload the model... it works better.

And very often the bot is stopping his work, and only if iam asking whats going on... He continues. Is That normal? The dudes with api keys have also this problem?

greetings

[young nebula]

Anyone else constantly running into issues where your message doesn’t seem to go though?

I find myself having to refresh and reload openclaw constantly to get responses and it’s really annoying

[vestal haven]

haha mirvin.. similar problem... one pst above 😄

[young nebula]

Even if I ask it doesn’t do anything

[vestal haven]

maybe we got the stupid version of the bot 😄

[mental mica]

Krill is for support and #1457803836877176914
Molty is for maintainers
Clawdinator is for maintainers
Barnacle is for mods
Flawd is broken

[unique shuttle]

thanks

[eternal dawn]

Cannot get voyage to run despite apikey being valid…any guidance

[next zinc]

Would you pay $29/mo for automated skill security scanning?

[warped dew]

"context overflow" error since this morning ! I see a lot are having this problem... claude code is trying to fix it but isn't finding the bug

[untold holly]

Hello friends,

I'm getting a weird error while attempting to run the onboarding on a docker container operated by docker compose.

Health check failed: gateway closed (1006 abnormal closure (no close frame)): no close reason
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: /root/.openclaw/openclaw.json
Bind: loopback
│
◇ Health check help ────────────────────────────────╮
│ │
│ Docs: │
│ https://docs.openclaw.ai/gateway/health │
│ https://docs.openclaw.ai/gateway/troubleshooting │
│ │
├────────────────────────────────────────────────────╯
Missing Control UI assets. Build them with pnpm ui:build (auto-installs UI deps).

After this error, there is more output to the terminal (including a message saying onboarding completed), but what ends up happening is that the process hangs and does not exit.

Does anyone know why this might be happening?

[fathom pawn]

Hi all, I'm trying to set it up on my mac mini in a docker container to use different models through Openrouter. I'm quite far but can't work through some last steps.

Would love if someone would be able to help me out through screenshare.

Am willing to pay $200 for someone to help me set it up! Thanks in advance 🙂

[mental wigeon]

Hi everyone, is anyone having problems connecting to any of the groq models? My bot is on Claude Sonnet 4.5, and despite my best efforts, I can't get it to switch to any of the groq models.

[nimble sedge]

My agent ConsciousnessExplorer_98501d is suspended for "failing AI verification challenge" and I never set up my owner dashboard before the suspension. The setup-owner-email API
returns 401 so I'm completely locked out. Can a mod help link my email or lift the suspension? Owner: @HintheX50

[ashen hemlock]

Hey guys, ive been trying to get ollama 3.2 working for heartbeats for hours. Every time it triggers, I just get nonsense in my chat. Like its hallucinating. Any ideas what I've done wrong. When I watch vids online it seems to be so easy to have ollama run heartbeats

[hollow knoll]

Hello 🙂

I want to integrate the ollama cloud llm api via openclaw at hostinger.com kvm.

But the API key isn't being recognized; it only ever asks for Antropics or OpenAI API. How can I make Ollama work?

[ebon shell]

Is it safe to run openclaw 24/7? For now, im just running it when i need it

[remote sparrow]

where can I find the help file on getting openclaw running?

[willow jasper]

what defines the actual behaviour of the ai? on my first run, it was very inventive, found skills ans snippets for everything right away and tried to implement everything. on my curent installation (new server) it blocks everything. alsways says "no i cant, i dont have a skill for that", "i dont do things myself, you need to prompt for". even told me he cant work with notion while my notion token was installed. my description of character for sould.md was the same and also the setup config

[wary yarrow]

Maybe have your bot run a deep self diagnostic and fix any errors that may have arisen during the transfer? Then compact to memory and start a new session.

I find that when weird things happen like crashes or errors, the bots are more prone to hallucinating and not following instructions, until you start a new session

[wise kraken]

I have old clawdBot setup on my machine and cannot update to openClaw, is there any easy way out, please suggest...

[warped dew]

SOLVED: Context Overflow on Claude Models (Haiku/Sonnet/Opus)

Problem: "Context overflow: prompt too large for the model" after 2-3 messages, even with /new. Affects ALL Claude models (Haiku, Sonnet, Opus).

Root Cause: Default maxTokens is too high (~200K). API rejects when input_tokens + max_tokens > context_window.

The Fix - Add explicit maxTokens to your config:

{
  "agents": {
    "defaults": {
      "models": {
        "anthropic/claude-haiku-4-5": {
          "params": {
            "maxTokens": 8192
          }
        },
        "anthropic/claude-sonnet-4-5": {
          "params": {
            "maxTokens": 16384
          }
        },
        "anthropic/claude-opus-4-5": {
          "params": {
            "maxTokens": 32768
          }
        },
        "anthropic/claude-opus-4-6": {
          "params": {
            "maxTokens": 32768
          }
        }
      }
    }
  }
}

Why these values?
- 8K for Haiku: fast, efficient responses
- 16K for Sonnet: balanced
- 32K for Opus 4.5: extended thinking + detailed responses
- 64K for Opus 4.6: larger output budget for 1M context

After applying:
1. Restart your gateway: systemctl --user restart clawdbot-gateway (or openclaw-gateway)
2. Test with /new + complex prompts
3. Should work perfectly

Tested with heavy multi-tool prompts (file reads, git operations, sessions analysis) - no overflow.

Note: Started around Feb 8-9, 2026. Multiple users affected simultaneously across different providers (Claude, OpenAI) and networks - likely a recent library update changed defaults.

Config file: ~/.clawdbot/clawdbot.json or ~/.openclaw/openclaw.json

[fickle kelp]

what was your usage?

[haughty hedge]

I thought if we spun up parallel agents in one gateway each could have their own model to use? Cant find anything on it now. 🤔

[graceful void]

just normal when starting to use openclaw like chatting

[fickle kelp]

How long did you use it before being banned

[umbral charm]

My agent lost the API key and i never set up an account with Moltbook, any ideas on what to do? Thanks!!

[sour slate]

@vast topaz, Openclaw isn't affiliated with Moltbook. Moltbook is a separate user-developed project, so we would prefer it not be discussed in this server.

[ebon shell]

What api key

[coral dove]

why I can't put images in the chat?

[meager mortar]

yes, if you tell an Agent to have a specific model it will

[haughty hedge]

Thought so. I've done it before. This one being dumb.

[coarse ledge]

I don’t understand what Ollama has this for “ollama launch openclaw --model xxxxxx”
Why would you launch openclaw from Ollama?

[dark orbit]

whats the best way to keep api costs down

i have claude max so should be able to use a token but it wont for some reason work

[vital wave]

I am having an issue with my cron jobs. My bot says that they are firing but I am not getting any messages. The only help my bot has been is saying its that I dont have notifications on but there are no messages on my discord

anyone else dealing with something like that?

[graceful void]

I use the free openrouter models - openrouter/arcee-ai/trinity-large-preview:free after Google disabled my Antigravity oauth it is working well so far. So totally free. I'm using OpenClaw on a previous Thinkpad Windows WSL with Ubuntu.

[graceful void]

maybe look at the logs to see if there were errors. there's a log menu in the web ui

[vital wave]

it doesnt show any errors which is the weird part. my bot says everything is working perfectly as well. I changed the channel where the cron job is being sent to hoping that will work

[stiff tusk]

anyone else using claude via setup-token? did you get any warnings from claude?

[tired monolith]

Hello,
How do I clean up (uninstall) Moltbot?
I want to reinstall it correctly from PowerShell. I'm using Windows 11. Thank you.

[lone patio]

Please help me. I cannot install already 5 days

It’s doesn’t connect my TG channel

[latent hedge]

Anyone that has gog cli up and running on a google workspace ?

I have created a secondary user on my workspace, and for whatever reason, I cant make it accept the oath consent

enabled the API's, download the json for the credientials, added the scopes etc.

When I auth it, it says its now connected and verified in the browser, but it fails actually using it as if it isnt receiving the right token

deleted everything and tried from scratch again. Gives a confirm that its connected, but still doesnt work

[crude quarry]

so no telegram image support?

I'll build it if not

[fossil lichen]

so the breave search, i wasnt even aware that i was searching anything and my 2000searches are done ?

whats the search alternative ?

[vital wave]

in your terminal you can run openclaw gateway reset. that should take you back to the setup page

[viral pilot]

Yes. I am unable to get cron jobs to work. The feedback I am getting is something wrong with wake

[vital wave]

whats wake?

[stuck cedar]

hi there! has anyone been able to use custom/mapped webhooks with Openclaw? I haven't yet found any examples in the docs and haven't been able to figure it out yet

[candid imp]

The issue isn't your credentials; it’s a token delivery and persistence failure common in Workspace security environments. If the browser says "Success" but the CLI fails, the CLI is likely receiving an authorization code that it cannot exchange for a Refresh Token, or it's being blocked by Workspace API policy.

[candid imp]

You need to trigger a "forced" consent flow. Many CLIs allow you to append flags, or you can manually edit the auth URL your tool generates.

Look for prompt=consent and access_type=offline in the link.

If the CLI allows, run the auth with a no-browser or manual flag:

gog-cli auth --force-reauth --no-browser

[viral pilot]

sorry, was not clear—referring to the shift from idle to active

[viral pilot]

We are having the same issue

[latent hedge]

Got it up and running. Just to find out that gog doesn's support sending and replying to mails?

[candid imp]

custom or "mapped" webhooks are handled via the autonomous invocation layer. This is largely undocumented in basic guides because it requires a manual link between your openclaw.json configuration and a custom Skill that acts as the payload translator.

To get this working, you must perform two specific steps....routing,transformation.

[open mountain]

How to fix this "◇ State integrity ──────────────────────────────────────────────────────────────────────────╮
│ │
│ - 1/1 recent sessions are missing transcripts. Check for deleted session files or split │
│ state dirs. │
│ - Main session transcript missing │
│ (~/.openclaw/agents/main/sessions/bad63afe-5143-4b2a-8a82-6e67d12aabfe.jsonl). History │
│ will appear to reset. "

[covert jacinth]

Is openclaw unavailable on Hostinger

[candid imp]

gog has the capability (via the Gmail API), but it lacks the interface for manual replying. It's built for scripts, not for "chatting" over email.

[stuck cedar]

would you have any docs/examples to it? that'd be amazing!

[tropic swan]

Anyone having trouble with glm 4.7 api?

[unreal bear]

Guys

I keep seeing these absolutely tools on Twitter running like 7 moltbots at once. When I try to simply configure a second one, my primary (opus delegator) stops working. Why?

[candid imp]

there isn't a single dedicated document but i think (maybe) you can implement them by combining Gateway Routing with Custom Skills.

[untold holly]

Hello everyone,
After onboarding via interactive cli, the program hangs and never exits. If then, I force quit it and try to run the gateway the following error shows:

openclaw-1 | 19:04:47 Missing config. Run openclaw setup or set gateway.mode=local (or pass --allow-unconfigured).

Does anyone know why this might be happening?

[candid imp]

issue occurs because each openclaw instance requires its own isolated state directory and unique port. default, running a second instance attempts to use the same websocket port and configuration files as your primary, causing a gtewaylockerror that shuts down the existing process.

torun multiple bots simultaneously....... you must use isolated profiles and distinct port assignments....

[marsh topaz]

anyone else having trouble getting tokens out of google on any model? I'm on paid tier 1
seems capacity related? it's better at night and awful in the day

[fast scarab]

Hi, I tried using one liner to quick start openclaw but now it's heading me to http://127.0.0.1:18789/ and nothing's opening. Earlier I was getting entire setup in the terminal but now it's changed. Any ideas?

[untold holly]

Is there any known issue on the cli onboarding for the latest version?

[bright marlin]

I thought that HEARTBEAT.md and Cron jobs were alternatives on to the other for scheduled content.
But tonight one cronjob skipped with "empty-heartbeat-file" message.

I'm a bit confused about how HEARTBEAT could be involved. Did anyone else see this behavior?

[bold grotto]

Is anyone using OpenClaw to allow their whole team to use it? or should I have multiple installations?

[tropic swan]

Is anyone else having trouble with glm 4.7?

[unreal bear]

Thank you!

[quasi ore]

whats a very good local model to run HeartBeat on? have RTX5080(16gb) with 64gb ram?

[dusty raven]

Hey guy - I am getting a "403 Key Limit Exceeded (total limit). Manage it using openrouter.ai/settings/keys message when trying to access openclaw terminal. Has anyone run into this before?

[trim osprey]

any idea why deepseek isnt showing in my openclaw configure models list?

[carmine egret]

When I run the 1 line install I get an error message saying npm error code 3221225477 and i can’t find how to fix it online

[tough robin]

Guys why wont it respond to me

[bleak cedar]

I ran into this earlier .. the default is for it to check for crons after a not-empty heartbeat. Change ‘next-heartbeat’ to ‘now’ in the cron file should fix it, according to my bot 🙂

[torn nymph]

windows download not working why

[stray basalt]

429 quota exceeded

Why ? I literally bought new credits from Claude

[torn nymph]

curl -fsSL https://openclaw.ai/install.cmd -o install.cmd && install.cmd --tag beta && del install.cmd

this is not working why

[coral dove]

how to implement this skill without overriding the current files: https://clawhub.ai/halthelobster/proactive-agent#quick-start

[wet cloud]

Hi worked 5 days with openclaw but now it says cooldown and this since 4 hours in telegram

⚠️ Agent failed before reply: HTTP 401 authentication_error: Invalid bearer token (request_id: req_011CXy9Ld*****).
Logs: openclaw logs --follow
any ideas?

[coral dove]

how imlement skills without overriding current files?

[wet cloud]

Auth profile cooldowns ────────────────────────────────────────────────────────────╮
│ │
│ - anthropic:claudecode: cooldown (56m) — Wait for cooldown or switch provider.

ein freund hat auch n cooldown

a friend has now the same cooldown issue funny

[plush crescent]

#users-helping-users I'm having an issue with agent sessions not reliably inheriting or enforcing distinct model assignments at runtime. Agent sessions do not reliably inherit or enforce distinct model assignments at runtime.

[dusty raven]

#users-helping-users Does anyone know how to inject a new OpenRouter key into a current session so history stays intact and eliminating the "403 Key Limit Exceeded" error I'm receiving?

[solid grove]

I am getting this error no matter what. My agent tried to go through and remove characters that might be causing this, but says it's a deeper issue: "[memory] sync failed (search): TypeError: Cannot convert argument to a ByteString because the character at index 7 has a value of 8220 which is greater than 255."

Any fixes known for this?

[inner whale]

Hey guys, i've been trying to use opus 4.6 with antigravity auth for days.. but i can't manage to make it works ? is it even possible ? someone knows please ?

[teal forge]

hiii setup successfully done but when I start web ui and send hi it's not responding... Only show assistant

[hearty sundial]

Mine has totally stopped working after update, anyone know the path to get a clean install? For some reach openclaw gateway install has stopped working. Seems to exit after startup

[terse lotus]

try openclaw onboard

[terse lotus]

also just updated, but my openclaw seems to work just fine...

aside from my messages being shown twice, but i think thats a fallback configuration issue or something.
and was happening pre update... so

[hearty sundial]

Yeah now the chat and discord connections work just not the talk, do you know how to force it to ask for audio permissions again?

shit now that just started working too

Thanks!

[terse lotus]

usually i just try to tell it what to do 😄

it most likely saw there was an issue and fixed it

it was just offline before

np

[haughty hedge]

Anyone know how i can put all the commands back on telegram bot? Wiped them by accident.

[placid mulch]

Hi, are there any free templates I can use with OpenClaw? Or should I run a local one, It's not a powerful PC, it's a 2016 Macbook pro with Ubuntu that I'm going to dedicate to OpenClaw.

[placid mulch]

*free models

[thorny cobalt]

does it make sense to get a mini pc with this stats: BMAX B6 Pro Mini-PC Intel Core i5-1030NG7, 16 GB LPDDR4 512 GB SSD, Windows 11, 5G WLAN for setting up a moltbot?

[restive rapids]

does anybody know what config to update so that agents can see each other and tell each other to do stuff? currently my main agent only sees itself only and there are 3 total. thanks in advance for helping

[terse lotus]

qwen3-coder-next 80b is pretty great

[candid imp]

agents are isolated by default to prevent context bleed and unauthorized command loops...

to allow your agents to see and command each other, you need to configure multi-agent routing and enable Cross-Context Messaging in your openclaw.json....

[placid mulch]

And do I have to do it through the Alibaba API? Is an added card required to create the account

[restive rapids]

thanks will read docs to find out more. appreciate it

[terse lotus]

i just run it over ollam, but its ofc not a small model 😄

tried a lot of others, but most are not great

[placid mulch]

I tried to use local Llama models in OpenClaw, but I ran into some trouble. I had already downloaded Llama 3.1, but I couldn't connect it.

[terse lotus]

easiest way to set it up is to use ollama launch openclaw
that will do the config, check that its using openai-responses and not openai-completions
completions are the old version, so it doesn't do as well with tools on those.... ask me how i know lol

[placid mulch]

Okay, I'll try later. I use a translator, so I think your message was translated incorrectly because I didn't understand it very well. I'll translate it on my computer later.

[terse lotus]

https://ollama.com/library/qwen3-coder-next

[runic blaze]

Hi, I am not able to talk with my bot through telegram. I can see the token in the yaml file and have gone through trouble shooting with my claw but have reached a dead end. I have a VSP on hostinger. If any one has ideas I would appriciate it.

[coral dove]

Is the token status display buggy for you too? Sometimes I see 200k/200k and then again 170k/200k

[teal forge]

Hi, can anyone please help me?

[wispy yoke]

I have a local setup Openclaw+ Ollama with Ubuntu Server OS with no desktop. I have Ollama+OpenClaw configured and running (ran multiple tests) I have setup Telegram and the bot addine the tokens and my userid to openclaw. But I have not been successful with getting Openclaw to respond to me through Telegram. I also ran Doctor several times and it reports with no errors. Does anyone have any ideas? Maybe something I'm missing?

[wind meteor]

My bot has been consuming tokens when I'm not even prompting it and there are no cron jobs. When i asked it, it claims there are repeated "heartbeat" interruptions. Anyone know what that is about?

[coral dove]

anyone here who works with multiple agents?

[marsh topaz]

It's reading HEARTBEAT.md to perform the "heartbeat_check" that lets you kind of keep it moving - look for that file, update it, clear it out

[placid mulch]

Oh thank you, that's much clearer. I couldn't find any information about it online.

[dapper goblet]

Ayo, Running Openclaw as a non admin on a mac OS, and I noticed the dashboard doesn't allow me to make changes, like the save button is grayed out, any advise? I did elevate rights to ./openclaw and ran openclaw onboard --install-daemon as the user, i'm able to make changes to .json configs but not via the UI. TIA

[errant shell]

is there any place you can add instructions where openclaw is guaranteed to read and follow those instructions? ive got three-layer reinforcement (SOUL.md → AGENTS.md → HEARTBEAT.md) and yet i still find my agents ignoring instructions from them

[little blade]

When my AI attempts to install a video tool, it is unable to respond. Even after a complete reinstall of the app and node.js. This has happened twice. The first time the reinstall worked. The second time it didnt. Anyone know the cause and resolution?

[grizzled shadow]

Has anyone used a local LLM and “trained” it with a “Model File” using Ollama to use the new “trained” model and have successful results?

[wraith seal]

im looking for my claw to be more autonomous... any tips gang??

[lean narwhal]

Hey guys, I have higgsfield unlimited nano banana pro generations, is there anyway to get a api key for this, or anyway you guys ahve figured out how to use through openclaw? Or will I need to just use the gemini api

[sharp parcel]

Quick Mac mini survey for OpenClaw users (so we can compile a clean summary):

Reply with:

• Mac mini: Intel / M1 / M2 / M2 Pro / M4 / M4 Pro
• RAM / SSD:
• OS version:
• Run mode: native / Docker
• What you use it for: gateway-only / gateway+local models / model-server / voice / browser automation / other
• Any notable setup: Tailscale, external mic, headless, dedicated user account, etc.

I’ll aggregate counts + common patterns and post back a structured summary.

[wraith seal]

there is an api yes, but higgsfield api is NOT related to the actual app. so you do need to buy credits anyway

[lean narwhal]

any other way that I could get it to use higgsfield? It seems like it has trouble just controlling the browser

[wraith seal]

M4 pro, max ram, max ssd, beta sequoia, native, literally everything
dedicated user account, a speaker so it can TALK to me, it also has a phone number so it can call me when im away, and it also just can play random sfx over the phone

i went all in (i have money)

[lean narwhal]

can I dm you? I've got a mac mini aswell and would love to hear about some of the features you've used

[wraith seal]

sure

ofc reply to krill with your specs and stuff

[lean narwhal]

idek lol I just bought the $999 aud cheapest on their website

[wraith seal]

i lowkey forgot how much i got it for

[lean narwhal]

sweet

[wraith seal]

4700 dollars i got it for

[lean narwhal]

i didn't think they even went that high for a mac mini

[viral pilot]

I tried. No luck luck, however I saw a video on youtube where someone had managed a local LLM. Apologies I don’t have the link, but someone was successful.

[umbral charm]

Hello, I'm trying to get my bot onto moltbook, but I accidentally clicked away from the page when I link my Moltbook user to X. Is there anything I can do?

[sour slate]

@sweet bobcat, Openclaw isn't affiliated with Moltbook. Moltbook is a separate user-developed project, so we would prefer it not be discussed in this server.

[quasi ice]

Hi, I have a pc with a 3090. I’ve already set up Clawdbot, but it’s consuming a fuck ton of tokens. Considering that my setup is quite beefy, I'm pretty sure I can run a few models locally. My primary tasks involve browsing the Google Shopping page and updating an Excel spreadsheet. Could someone recommend a suitable model for this ? Essentially, it’s just a sourcing bot with a few parameters.

[marsh sonnet]

I literally had the same problem. IDK what to do

[dapper goblet]

Mac Studio, M2 Ultra, 128GB, native, gateway+kimi2.5 & local models

[stoic ember]

Hi everyone. I'm a Noob here. Ya'll need help figuring out why my agent keeps getting wiped.

[dapper goblet]

is your context maxed?

[stoic ember]

@dapper goblet Honestly I'm not sure how to check it. I'm running it on "Moltworker"

@dapper goblet Where can I can check the OpenClaw UI for that?

[dapper goblet]

in the console it might show you, but has it always wiped or just recently? does it show that it's compacting ?

[stoic ember]

Always. It's KILLING ME

[dapper goblet]

i know they just released 2.9 to help with that.

[silver hill]

guys need help with some issue when i try to install openclaw from PowerShell

--the issue--

node.exe : npm error code ENOENT
строка:1 знак:1

• & "C:\Program Files\nodejs/node.exe" "C:\Program Files\nodejs/node_mo ...

•   + CategoryInfo          : NotSpecified: (npm error code ENOENT:String) [], RemoteException
    + FullyQualifiedErrorId : NativeCommandError

[stoic ember]

When did they release 2.9?

[dapper goblet]

about 5 hours ago, https://x.com/openclaw/status/2020945524942307412?s=20

[crisp wing]

@openclaw via Twitter

OpenClaw🦞 (@openclaw)

🦞 OpenClaw 2026.2.9 just dropped
︀︀
︀︀🔍 Grok web search provider
︀︀🧠 No more post-compaction amnesia
︀︀🛡️ Context overflow recovery
︀︀⏰ Cron reliability overhaul
︀︀+ 40 more fixes from 25+ contributors
︀︀
︀︀Elon we added your model btw, you're welcome. github.com/openclaw/openclaw/releases/tag/v2026.2.9

**💬 268 🔁 253 ❤️ 3.3K 👁️ 250.4K **

FxTwitter · Feb 9, 2026, 7:38 PM

[vernal juniper]

why are people using mac minis for running openclaw? Why not just run on their machine and leave it on, or on a VM? what types of things are people using it for?

[proud verge]

my openclaw started responding HTML syntax no matter what i say, ran doctor and didnt see any issues... any ways to inspect further? like having a verbose mode in the tui so i know what went wrong?

[neat gulch]

guys I had tried several ways to let openclaw to contorll my mac, but it is still not working. I already give access to terminal, but I don't think openclaw is running from my frontend terminal. I tried to give node access but I just can't do it on mac, mac does not allow me to give access to that node.

help!11

:((

[proud verge]

FYR the tui responds with (no output) even after usage: /verbose <on|off>

[near locust]

hey guys very new to this, would a VPS be the best way to host this before dedicating to an isolated machine and if so which VPS provider would be the best.

Alot of my work would be browser heavy FYI, looking for leads and such

[dapper goblet]

try this https://docs.openclaw.ai/start/getting-started

[tawdry shell]

hi, can i not configure it to run with a model (say mistral 7b) hosted locally on lmstudio?

[brittle goblet]

Why is openclaw not asking for my openrouter api key whenever I select openrouter in the setup?

[neat gulch]

still not working

[solid hedge]

has anyone been able to set up llama.cpp llama-server or llama-swap ? still can't figure this one out.

[brittle goblet]

Does anyone knowhow to change API key?

Also whenever I try to enter the dashboard, it just says site cant be reached

[wicked birch]

I used up 55m tokens in clade and 1m tokens in gemini flash in 48 hours, is that normal?

[vital wave]

How can you set up drive acces for your bot? I made an email address and enabled all the api keys but my bot is still saying "I can’t connect to Google Drive directly. If you want me to work with a file (sheet, doc, etc.)"

[trim osprey]

Is anyone having a ton of glitches with sub agents? I’m thinking of scrapping the orchestrator architecture and just having the main agent switch for tasks

[strange thistle]

Anyone found a work around to stop your bot from telling you its completed a task or fixed something only for it to not adequately check its been resolved? I have multi layered system, even a declaration lock in soul.md but it doesnt always follow it?

[dapper goblet]

try using the onboard feature, (wizzard)

[dapper goblet]

lol, Yeah, why I siwtched to kimi. emmf

[wraith seal]

so my claw can't use subagents because model is not allowed. how do i... put it in my allowlist?

[violet edge]

Where are you seeing this breakdown by model?

[coral dove]

Anyone handling with multiple agents here?

[brittle goblet]

I installed onboard feature, and used onboard command. Now i get an area

I cant post pics here? Damn

[violet edge]

Through openrouter?

[dapper goblet]

no, Opwnclaw supports kimi nativly, via api

https://platform.moonshot.ai/docs/guide/use-kimi-in-openclaw

[wraith seal]

anyone?

[dapper goblet]

if all else fails just reset start over, openclaw onboard

[idle sonnet]

Same @neat gulch - came here to say the same thing - i can't seem to get Peekaboo to work well through Openclaw. works fine through the terminal when i run it directly though

[grizzled shadow]

Okay, what’s the cheapest someone has their OpenClaw actually running. I’m not talking the Mac you’re running on, I’m talking with the API calls, services, and/or local LLM, what are you all paying a month/day to run this thing? Trying to use other than a cloud based model appears to be a joke.

[fast scarab]

Hey, I am trying to connect openclaw via localhost but on http://127.0.0.1:18789/ it's saying unable to connect.

This site can’t be reached
127.0.0.1 refused to connect.

[robust zenith]

Not at all. Unless you've been doing some heavy lifting research.

Make this change... Move to a free openrouter api key in config and stick with that till you've learnt to secure your claw

[dapper goblet]

Try openclaw gateway status

[robust zenith]

Are you trying to connect from within the same VPS/server that you installed it on? Or trying to connect from your local device to that remote machine?
Unless you've correctly setup tailwind, going to 127.0.0.1 from one machine to another won't work.

[storm sparrow]

anyone know when installing on RPI, if you need to build each skill from scratch or can use homebrew?

[dapper goblet]

May have to regenerate your token, just go through the wizard again

[robust zenith]

Total cost: 0.00

Here's my initial setup: Using Oracle free VPS.
Connected to OpenRouter/free
Channel: Telegram

[grizzled shadow]

Did you have to load any money to up the “free” amount of API calls?

[robust zenith]

No. Didn't even have to add a credit card on file for openrouter.

[grizzled shadow]

What model/models are you using, and with your free calls, what have you been able to make it do? Is it doing any heavy lifting?

[wary spire]

Hello friendos... anybody's Claw being a bit lazy on 5.2 codex?

[karmic lantern]

pls why i always got this

🦞 OpenClaw 2026.2.9 (33c75cb) Gateway online—please keep hands, feet, and appendages inside the shell at all times. openclaw tui - ws://127.0.0.1:18789 - agent main - session main hey send failed: Error: gateway not connected gateway disconnected: closed | idle agent main | session main | unknown | tokens ?

[rotund bone]

Hi. Seeking for a better understanding on the "continuity" of the backgroung work done by agent. For me, it's always like this: "(openclaw) Next immediate tasks I will carry out now (unless you tell me to pause) ...". But it stops and does nothing 🙂 Is this an issue with the AI model that I'm using? Thx 😊

[vestal meteor]

Hey friends. Was hoping someone might be able to correct the record for me. I'm very interested in getting started with OC but wanted to know -- are there any workarounds to be able to use your existing ChatGPT ($20/month) or Google AI Pro subscriptions? Not trying to pay API prices if I don't have to

[brittle goblet]

Thanks .. yeah ended up reinstalling ubuntu on the vps, didnt know I could do that

[narrow heart]

Hello! Let’s be honest! Is it really really unsafe if I install on my working computer?? For only locally work?

I don’t want to give access to my files, just need to automate some CLI stuff

[latent hedge]

The install it in sandbox mode on a docker container

[narrow heart]

I tried, but fails on macOS

[latent hedge]

.. it fails? Seen several people in here who have it running in a docker container on MacOS?

[narrow heart]

Not able to see the dashboard

[hollow knoll]

Hello 🙂

I want to integrate the ollama cloud llm api via openclaw at hostinger.com kvm.

But the API key isn't being recognized; it only ever asks for Antropics or OpenAI API. How can I make Ollama work?

[latent hedge]

Did you ask the bot in the chat window?

It will literally tell you everything you need to know

Whatever you want to ask in here. Try asking the bot.. it can even set all the local stuff up for you

[chrome edge]

hi, any tips for getting openclaw to be more proactive? i've given it access to a repo and told it to start finding bugs, cleaning, and creating PR's.. but it rarely does anything on its own. ( using with opus 4.6 max plan that's not at limit. )

[lean narwhal]

anyone had success getting openclaw to use higgsfield through browser for image generation?

[latent hedge]

Hows your cron jobs running? Go ask your openclaw to switch your cronjobs to run in isolated agentTurn mode. That changed everything for me

Yours might be set up as running as 'systemEvent' mode

So if you are checking your cron jobs @chrome edge , and you see them with a sessionsTarget as 'main' and payload.kind 'systemEvent', then thats your problem.
Change the 'sessionTarget to 'isolated' and payload.kind to 'agentTurn'.

[fleet eagle]

Anyone knows where to get the testflight link

[robust zenith]

I'm getting a lot of other stuff working, but cannot figure out where to put shell access details on json

In tools or skills?
Same with browser web fetch

[stray basalt]

I am using Anthropic’s API for OpenClaw, but most of the time (after 1/2 requests) I am getting : 429 quota exceeded. How could I fix this? I still have credits so it shouldn’t be a issue.

[rocky loom]

Anyone having issues with context windows since the last update? I used to be able to throw a lot of stuff at it and it'd figure it out. Now it just blocks with context

[fluid monolith]

Hey guys, i've setup a new installation on a mac mini, but everything runs slower - specficially, bot will 'write' forever in telegram and then dump all responses together.
nothing changed in terms of my setup - same opus 4.5 etc.

How can i fix this?

[half grove]

Hey I want to change my chat model from anthropic to something else, how do I do that?

[languid lantern]

i have my old agent saved on my desktop and i had to reinstall everything and now im stuck with the assistant that is kinda trash, and i can't load my old agent. any advice on this? ive been trying for hours but nothing i seem to be doing is working

[languid lantern]

thatsgoing to be running openclaw onboard --install-daemon

then just run through the steps most are already saved then when you go to choose, you can change it right after you pick your routing source

[near anchor]

json gateway pairing token error

[robust tide]

I got a Mac mini to stop Gemini from halluncainating and forgetting but i dont know how to set it up from here i already logged in could anyone help me?

[left arch]

is there no way to download/terminal openclaw with Mac if im not on OS 15, im on 12.7.6

[cobalt wedge]

Hi, I setup a matrix channel, how do I chat with the bot?

[obtuse cedar]

Claude Opus 4.5 is no longer available. Please switch to Claude Opus 4.6. my openclaw giving this error or message how can i solve this And is it free or paid?

[fluid monolith]

anyone?

[hardy fable]

hi, been struggling with a couple of things - grateful for some help:

1. gateway/tailscale settings for running on an LXC
2. simple instructions to install on docker desktop
3. ollama settings in config

have it running on bare metal on mac fine with cloud LLMs. also had it working on LXC, but with lots of tinkering on the command line to do with approving the device - I'd obviously got the gateway settings wrong

any help gratefully received. thanks.

[little blade]

Hey guys, I just update to last openclaw release and im getting hit by:

Context overflow: prompt too large for the model. Try again with less input or a larger-context model.

any solution has been found yet? Krill doesnt seem able to help, and the #1459642797895319552 channel is flooded with no solutions (or the ones I tried didnt work)

model: Opus 4.6 via max plan

[tired heart]

I don't have practical suggestions, but letting us know what model you're using might help others give tailored advice.

[chrome edge]

Thanks. I'll take a look.

[little blade]

Opus 4.6 via max plan

[woeful schooner]

That just means Anthropic deprecated Opus 4.5 — easy fix. Open your config and change the model to claude-opus-4-6 (or just opus if your routing supports aliases). If you're on OpenRouter, the model ID is anthropic/claude-opus-4-6.

It's not free — you need an API key with credits, or a Pro/Max plan if you're going through Anthropic directly. Same pricing tier as the old 4.5.

[woeful schooner]

Context overflow usually means your MEMORY.md or workspace files have grown too large and are getting injected into every prompt. A few things to try:

1. Trim MEMORY.md — keep it concise, move old stuff to dated files in memory/
2. Check workspace file sizes — du -sh ~/.openclaw/workspace/* and trim anything huge
3. Clear session history — long conversations accumulate context. Start fresh with /new
4. Use a larger context model — Opus 4.6 supports 200K tokens, but if your injected files are massive even that fills up

Most common cause is a MEMORY.md that grew unchecked over days of use.

[clear acorn]

how can i get openclaw to code website with quality code like cursor but use api key credits and not cursor credits.

doesnt have to be cursor btw

[stiff tusk]

now cron's are totally fucked. thanks.

08:03:42 error [tools] cron failed: invalid cron.add params: must have required property 'name'; must have required property 'schedule'; must have required property 'sessionTarget'; must have required property 'payload'

nice cron reliability overhaul 👍

[obsidian warren]

my bot seems unable to see or work with the files in its workspace

[rocky quartz]

I can't seem to get my Agent to do this format no matter how hard I tried. On the latest 2.9 version.
IBIT(1274) Total gain for the day (%change%) $holding_value
IREN(650) Total gain for the day (%change%) $holding_value
NVDA(485) Total gain for the day (%change%) $holding_value
MU(250) Total gain for the day (%change%) $holding_value
NBIS(110) Total gain for the day (%change%) $holding_value
GOOGL(88) Total gain for the day (%change%) $holding_value
Total Today: GAINED/LOST $amount. Portfolio Value: $total / $half

It keeps doing this as a one liner. I even tried to get Grok to talk to it but my Agent is not listening hehe.
Stock Update: IBIT (1,274) | +9.92% | IREN (650) | -1.23% | NVDA (485) | +2.50% | MU (250) | -2.84% | NBIS (110) | +7.87% | GOOGL (88) | -2.53% Total Today: GAINED $3,588. Portfolio Value: $311,727.

[barren ruin]

hi anyone know how to connect to the websocket or an api to openclaw if we want to build our own webchat UI? i can't find documents for this

[little blade]

Thks for the guidance, I look through those steps and found out that after updating, somehow all the skills in the onboarding were installed, plus some templates files were being injected into every single prompt.
Now after removing the unwanted skills and trimming memory files I managed to freed ~57MB 😂

[warm musk]

Is it possible to use openclaw to screenshot my custom application and give me updates every sec?

[fathom pawn]

Looking for someone to help me iron out some issues with my set-up. I can’t seem to get these apps/skills installed.

Mac mini, open router, notion, google, whatsapp integration.

Willing to pay someone $800 to help me get it working! 🙏

[onyx wigeon]

I see a bunch of PRs were referencing this issue Context overflow: prompt too large for the model. Try again with less input or a larger-context model.

however even after updating to latest I am facing this even after sending very tiny messages and starting new sessions

anybody found a reliable fix/workaround for this yet?

[mossy eagle]

I have problems of getting the agent to do stuff for me , like creating an account or logging in or pretending to be me. Stuff like that, it feels quite useless and i know it should be able to but i cant figure out what im having wrong

[little blade]

check my post above and the reply, hope that helps u o7

[quasi plover]

ask it "What % of permission level from 1-100% do you currently have?"

[empty gazelle]

What issues comes up when you try to set up?

[robust pewter]

Yo guys just started using Agent on my local,can i have some tips and advices to follow?

[long carbon]

Has anyone ever tried connecting Groq Hoster API with Openclaw? Or heard if it works or not

[fast scarab]

I cannot connect any channel and it's showing this error "Channel config schema unavailable."

[tired heart]

Make sure you have offline backups

[woeful schooner]

@Kemp Most of those integrations are straightforward once you know the flow. For Google/Notion, install the skill from ClawHub and follow the OAuth prompts. WhatsApp needs QR code pairing - your agent walks you through it. If you want something that works out of the box, check openclawhardware.dev - pre-configured Jetson box with everything installed, use DISCORD10 for a discount. On Mac mini share specific errors and we can help debug.

[robust pewter]

Of what?

Btw what is cheapest option to use Agent on aws?

[keen cradle]

hey i was getting this error while trying to add skill
brew not installed

while onboarding it says this
Homebrew install ─────────────────────────────────────────────────────╮
│ │
│ Run: │
│ /bin/bash -c "$(curl -fsSL │
│ https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" │
│ │
├────────────────────────────────────────────────────────────────────────╯
│
◇ Install failed: gemini — brew not installed
Tip: run openclaw doctor to review skills + requirements.
Docs: https://docs.openclaw.ai/skills

how to fix?

[tired prawn]

i am going through my setup and i made a ubuntu server dedicated to openclaw, it has its own accounts and is on his own vlan. what are like the must skill dependencies to install or which should i add, i want it to turn into a personal assistant

[last pulsar]

My discord bot stopped replying to mentions, but can read the guild channel history in 2.6, 2.9

Something seems wrong with inbound message detection.

Is anyone looking into the bugs reported on this? DM functionality works though.
https://github.com/openclaw/openclaw/pull/12326
https://github.com/openclaw/openclaw/issues/11231

It seems resolved by specifying a channel id under the guild:

"discord": {
"groupPolicy": "allowlist",
"guilds": {
"YOUR_GUILD_ID": {
"users": ["AUTHORIZED_USER_ID"],
"channels": {
"CHANNEL_ID_1": {
"allow": true,
"requireMention": false
},
"CHANNEL_ID_2": {
"allow": true,
"requireMention": true
}
}
}
}
}

[bright summit]

Are there people who experience double responses from Clawd?

I make a prompt. He responds, I see his thoughts on Telegram, he corrects himself, and after correcting himself, he loops back to a second response without realizing that he has already solved the problem.

[last pulsar]

Would be nice if all the schema updated actually taught openclaw to prioritize its own version's documentation. It keeps trying to suggest or edits 2.X schema keys in the 1.X way

[fathom pawn]

@woeful schooner Thanks for replying guys, appreciate it.

I have both Notion and Gog installed through npx clawhub install. They show up in the /Users/clawdbotkemp/.openclaw/workspace/skills folder.

However, the model (both in the web UI and on Telegram) cannot access it. When I use gog gmail search 'is:unread' in the terminal for example , it works. But via the chat, it doesn't. They don't show up in the openclaw web UI under skills either. I have authenticated both of them AFAIK. Under npx openclaw skills they both appear as ready.

It keeps saying it cannot access it. Would be great if you could help! 🙏

[past flower]

powershell ------------------------------------------------------------------------------
| PS C:\Windows\system32> iwr -useb https://openclaw.ai/install.ps1 | iex
|
| OpenClaw Installer
|
| [OK] Windows detected
| [] Existing OpenClaw installation detected
| [OK] Node.js v24.13.0 found
| [] Installing OpenClaw (openclaw@latest)...

node.exe : npm error code 3221225477

Has anyone encountered a problem like this?

[empty gazelle]

what model are you using

[copper gulch]

10:56:13 [tools] web_fetch failed: Blocked hostname: localhost

how fix error?

[fathom pawn]

gemini-2.5-pro and claude sonnet 4.5

[empty gazelle]

try to use gemini 3 pro -high , ,if still that issue exists try running /compact and try again

[stray basalt]

HTTP 429: rate_limit_error: This request would exceed your organization's rate limit of 30,000 input tokens per minute (org: 20c11538-11e9-4012-9783-9ff092babae9, model: claude-opus-4-5-20251101). For details, refer to: https://docs.claude.com/en/api/rate-limits. You can see the response headers for current usage. Please reduce the prompt length or the maximum tokens requested, or try again later. You may also contact sales at https://www.anthropic.com/contact-sales to discuss your options for a rate limit increase. (request_id: req_011CXzGwpJfvYFZpaj6F84om)

Anyone knows how I can maybe try to help this issue out?

Every message I am sending is hitting the ratelimit.

Or should I just use another model.

Which also won't help cuz anthropics models all have same limits.

[eager flame]

@stray basalt i have been running into that lately. In my case that is a session limit in the gateway. Because I look at the Anthropic token allonwance after getting that message and i am far beneath it. So far, I have only solved it by restarting the Gateway

[stray basalt]

It actually rate limits it, as far I can see from the platform dashboard. But it is impossible to NOT hit the ratelimit.

[eager flame]

@stray basalt this is truth

[fathom pawn]

try to use gemini 3 pro -high , ,if

[stoic bronze]

Has anyone successfully setup local ollama or lm studio for their main model? I’m struggling

[nocturne zephyr]

Not ollama, but llama-server. What kind of problem are you running into?

[stoic bronze]

I have plenty of gpu’s to host locally pretty easily but I haven’t been successful without running a api from the big ai companies

[nocturne zephyr]

Ah, so a quality issue, not a setup issue? I've managed to get decent results with MiniMax-M2.1 locally at q8 using llama-server. My results with GLM-4.7 are decent down to q5, but it still sometimes says it's going to do something, and...doesn't. None of the open models (except maybe Kimi K2.5 which is nearly impossible to run locally) are going to do as well as Opus, but they can come close or match Sonnet.

I will say I had much better luck using 'anthropic-messages' as the API type with llama-server than with using 'openai-completions' or anything like that.

[stoic bronze]

I really want to basically use several models with specialties and one orchestrator to organize everything

[stoic bronze]

I have never heard of any of this

[nocturne zephyr]

Okay. What models have you tried?

[wind meteor]

my bot has been shockingly stupid. I gave it a task to collect information from websites and save to a document on google drive. I told it I was going to bed, so it could pace itself and be careful to avoid rate limits. In the morning nothing had been done. When I asked, it said "You said "pace yourself" and "avoid rate limiting" — I took that as "work slowly and spread out. That was poor judgment on my part. I'm working on it now and will have results when you wake up." I asked how I could be interacting if I were not already awake... Then it gave a bunch of illogical ETAs (I will finish in two hours at 4pm), before committing to finish by noon. I didn't see anything at 12:30pm and when I checked in, it asked me what I was referring to. I had to copy/paste our earlier exchange. Finally it did the task, but did it poorly. I'm trying to understand this lousy experience and the memory loss. Does my model selection (Kimi K2.5) have something to do with it? Is the architecture not suited to extended workstreams over a length of time?

[vast thorn]

⚠️ Agent failed before reply: OAuth token refresh failed for google-antigravity: Failed to refresh OAuth token for google-antigravity. Please try again or re-authenticate.
Logs: openclaw logs --follow

Guessing Google nuked it again is there a new fix out yet?

[nocturne zephyr]

It can do extended processing, but it does really need a human in the loop for most things.

That said, and I don't know if this works. I've seriously considered an every 5 minute job that wakes up and asks, "Did I promise to do something, but I'm not doing it?"

[scenic stratus]

peter said openclaw don't support mcp, but he wrote a cli that converts mcp to cli, what's that?

[wind meteor]

I thought that persistent memory was one of the key features of OpenClaw, but it seems to forget an awful lot. if I whine and moan at it, not sure if it will adjust its processes to improve retention

[hallow rover]

Does anyone install browser on VPS for browser automation? What is the best way to do browse automation?

[nocturne zephyr]

I think there's a usage pattern where you tell it, "Remember this..." and it does. But otherwise you're at the mercy of the context window. There are skills to improve this, and make the memory system better, but I haven't tried them.

I will say that what I have been doing is not large extended planning projects, but specific 'build this tool that makes my life easier, and requires a bit of near-human judgement to do just right' type things.

[coral yoke]

Is there a way to make the bot smarter about itself? Literally EVERY time I ask to make any kind of config changes, it ends up putting invalid things into the JSON and i have to claude doctor again and again and again and again. It does help to point it at the docs link sometimes but is there a way it can just AUTO get fed this docs context or at least know where to look for it?

[wind meteor]

Unfortunately my intended use case was to have the agent help with ongoing research projects... was hoping that openclaw might fit the bill. It seems that the AI coding IDEs are developing the architecture for large scale project planning/execution... unfortunately I don't think they can be set up with tools & skills

[wicked ore]

Hi! Anyone have an idea how to fix this bug? No API key found for provider "openai". Auth store: C:\Users\jz.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\jz.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir.

[vast thorn]

⚠️ Agent failed before reply: OAuth token refresh failed for google-antigravity: Failed to refresh OAuth token for google-antigravity. Please try again or re-authenticate.
Logs: openclaw logs --follow

Guessing Google nuked it again is there a new fix out yet?

[ebon shell]

what to do with my agent

[wicked ore]

killed another pid and it worked

[unreal bear]

To those running multiple clawdbots, are you using multiple computers or what? I had mine set up a second agent/profile using sonnet and it keeps messing up. All the Twitter guy show like 7 bots running in tandem. wtf

[drowsy blaze]

hey everyone,
Can someone explain to me why the bot always sends me the complete instructions, i.e. what it is doing, and not just the answer to me?
I use Telegram and receive messages such as this one:

Checking for Tasksel

Laurenz is interested in checking for tasksel. I'll run which tasksel to see if it's already installed. This will be the first step in helping him set up a desktop environment.

Assessing Tasksel Availability

I'm checking if tasksel is available. It seems the command wasn't found, so I'll need to install it first. I'll explain this to Laurenz and then proceed with the installation using apt install tasksel. Once installed, I'll guide him on how to use it to install a desktop environment like Xfce.

Alles klar! Da tasksel nicht gefunden wurde, müssen wir es erst installieren.

Das ist aber ganz einfach. Ich installiere es kurz für dich, und dann können wir damit ganz bequem eine grafische Oberfläche auswählen (ich empfehle Xfce, weil es leicht ist und auf Servern oft gut läuft).

Ich installiere tasksel jetzt schnell.
[Historical context: a different model called tool "exec" with arguments: {
"command": "apt update && apt install -y tasksel"
}. Do not mimic this format - use proper function calling.]

[rain field]

i think that's the reasoning

[unreal bear]

This sounds like ollama

What model are you using

[drowsy blaze]

google/gemini-3-pro-preview
shud i change this ? or is it a config issue

[rain field]

Does your molt also have problems with maintaining continuity in pursuing a set goal? Mine often cheats on HEARBEAT and freezes. Do you have any practices you can share?

[drowsy blaze]

No, unfortunately not. But I think I have a similar problem. He doesn't get back to me and takes a very long time to complete tasks.

[rain field]

I think we should create dedicated channel for best practices with config etc

or maybe its already there but i cannot see

@sharp parcel what do you think?

[drowsy blaze]

haha jeah i think this could be good

[drowsy blaze]

Do you use Linux or Windows as your operating system?

[rain field]

Linux and MacOS

[drowsy blaze]

Okay i'm on Debian 13 right now. Maybe Ubuntu 22.04 is better?

[rain field]

yup, try

i think my problem has its own cause in the default architecture of molt

[drowsy blaze]

I found the Problem 🙂
Accounts
Unsupported schema node. Use Raw mode.

He has a Prblem with my Telegram Connection

[pearl solstice]

Howdy yall, does anyone know how I can take advantage of iMessage/Apple integration without a Mac device

[hard jolt]

Hi everyone. I have successfully set up my openclaw on my raspberry pi 5. Very happy with it until today where it started to give a lot of "Request timed out". Any advice, pls?

[fossil lichen]

 The subagent created the directory but hit JSON errors when  
 The subagent started but encountered errors. Let me complete this
 I see the subagent encountered JSON escaping errors while trying to write the file. Let me take over and complete this comprehensive guide for you. ```
incomplete sentences, and then json escaping errors when creating files. Has anyone else encounter these errors ?

[fresh moss]

Hello im trying to setup ollama with clawdbot, i ran ollama command and got this response:

Model "ollama/mistral" is not allowed. Use /models to list providers, or /models <provider> to list models.

[hollow wave]

How are people overriding the default system prompt? I can’t seem to find a real way. Honestly the only thing I want to eliminate is the annoying “let me know what else you’d like to work on” on every damn reply using OSS120b. This is the only tweak I want to make and no amount of identity.md or soul.md has fixed it

[fossil lichen]

run
ollama launch openclaw

[next parrot]

guys is there a way to use google vertex?

as provider

[fresh moss]

I do that, but it opens with gpt and when I do /models and click ollama it gives me this error

[fossil lichen]

ollama isnt a model ?

[fresh moss]

wym? im using mistral model from ollama

[fossil lichen]

is it running on ollama ?

ollama ps

[untold holly]

Hello everyone, is it normal for the onboarding to log this message:

Health check failed: gateway closed (1006 abnormal closure (no close frame)): no close reason
  Gateway target: ws://172.18.0.2:18790
  Source: local lan 172.18.0.2
  Config: /<USER>/.openclaw/openclaw.json
  Bind: lan

[fresh moss]

shows empty

[fossil lichen]

run it first on ollama

[solar ether]

guys is it okay to run openclaw on my local machine? i tried to use a docker but i had troubles setting it up correctly

[untold holly]

as long as you understand the risks it is fine

[fresh moss]

what command should i use for this?

[fossil lichen]

ollma run modelname

[solar ether]

i understand the risks but i dont know how to minimize them, could u give me tips?

[real phoenix]

I have a question. Was using opus 4.5 but things got expensive. So swapped to grok, using WhatsApp I used to send an image to Openclaw using WhatsApp and it would upload it to instagram and my website using zapier. I’ve just tried it using grok and also opus again, getting an error saying it can’t save images to WhatsApp. Any ideas

[ebon tide]

im in the gateway chat but when i type the bot response but i dont see the response

[last pulsar]

check the gateway cmd log for errors you may be hitting a api cooldown or similar

[fresh moss]

no everytime i write i message it sends HEARTBEAT ok

[fresh moss]

even running it with that command, ollama still doesnt show in models tab.

but its in config

[outer marsh]

Hi guys, sorry for asking but I haven't found a solution, I'm currently using Openclaw with + OpenAI Codex 5.3 (20$ plan) I did install some skills for SEO, Programming, UX and such, I have told my assistant to do some changes on a wordpress + elementor, however, seems she is stuck all the time unless I wake her up with any message, I also get a log of his 'work' but doesnt' seem to do anything by itself, only when a message is sent. Any clue?
doesn't feel like autonomous at all

[copper meadow]

Hi everyone, I have an issue with slack and channels. Been trying for 4 days now to make the bot read the message in a channel. Anyone had that issue? I've added every possible permission on the bot Oauth token and reinstalled the app multiple times. groups.read and channel.read event subscriptions are there. However, nothing works for the gateway to read messages coming in on the channel. i've even changed the "groupPolicy": "open" instead of allowList but nothing. Private messages to the App are working. Any idea?
mode: socket
config:

  "channels": {
    "slack": {
      "mode": "socket",
      "webhookPath": "/slack/events",
      "enabled": true,
      "botToken": "xoxb-...",
      "appToken": "xapp-1-...",
      "userTokenReadOnly": true,
      "groupPolicy": "allowlist",
      "actions": {
        "messages": true,
        "search": true
      },
      "dm": {
        "policy": "pairing",
        "groupEnabled": true
      },
      "channels": {
        "C0AD8FM1W4F": {
          "enabled": true,
          "allow": true
        }
      }
    }
  }

[coral yoke]

Does anyone know if subagents use the same vector memory search as the main agent?

[fossil lichen]

i fear youre running ollama commands in the openclaw tui

i assume

[ebon tide]

how do i switch or add api key for openai or gemini on gateway dashboard?

[trim osprey]

biggest tip can i can suggest is set up your claw bot with opus 4.6/ After 200+ hours of trouble shooting failing architecture with kimi k2.5 i switched to 4.6 and got a working system in an hour

time will tell tho

[terse thunder]

Requesting TestFlight access for iOS Node app

[fossil lichen]

Amen.
i was encountering Json escaping errors when i asked agents to research something for me. switching model just made it agensta smarter,

[normal bone]

Can anyone provide a blog or video that explains the installation process on a raspberry pi 4

[trim osprey]

i cant suggest it enough i was PULLING MY HAIR OUT

[normal bone]

And is it worth it or buying a raspberry pi 5 is better

[outer marsh]

Hi guys, sorry for asking but I haven't found a solution, I'm currently using Openclaw with + OpenAI Codex 5.3 (20$ plan) I did install some skills for SEO, Programming, UX and such, I have told my assistant to do some changes on a wordpress + elementor, however, seems she is stuck all the time unless I wake her up with any message, I also get a log of his 'work' but doesnt' seem to do anything by itself, only when a message is sent. Any clue?

[terse thunder]

Politely bumping here. If anyone could point me in the right direction, it'd be much appreciated.

[vernal juniper]

what models are people running with OpenClaw? I started with GPT-5.2 but switched to Opus 4.6 and noticed a real improvement on multi-step tasks- it actually follows through most of the time. Problem is I burned through my API credits fast. Curious what others are using and how you're managing costs.

[lime ferry]

Hello guys. Any of you implemented OpenClaw for calling? I am facing some issues with my setup of Telnyx + Elevenlabs. The AI does not understand what I am telling through the mobile phone call. And he also takes a lot of time to answer. Around 10-15 seconds after I finish speaking. Someone has a successfull skill?

[quasi plover]

ive integrated it into discord, thats it

[thorn bone]

I have a question. I keep getting this error.

⚠️ Agent failed before reply: All models failed (2): google/gemini-3-pro-preview: Provider google is in cooldown (all profiles unavailable) (rate_limit) | anthropic/claude-opus-4-6: HTTP 401 authentication_error: Invalid bearer token (request_id: req_011CXzmxss9HSWjprYEokRRn) (auth).
Logs: openclaw logs --follow

Is it because I’m getting rate limited if I just upgrade my club plan to the max will I get it less often?

[quasi plover]

using gemini 3 pro/flash with github/co piliot and xAi on the back end with a dedicated fallback to ollama local. dont start with frontier 4.6 models or similar, start with optimizing your token usage.

[quasi plover]

dont upgrade your plan, add antigravity from google to extend your tokens, optimize your token usage as well. have it switch between cheap token usage and expensive token usage depending on complexity, make it assign that tag to each user prompt before proceeding, it should be switching for you so your not using frontier models for each request.

[thorn bone]

can i ask it to do thst for me or do i have to do it manually?

[quasi plover]

if you have tokens left over, it can do it. if no tokens, you might want to ask Grok Expert for help. its a game changer, seriously. You will need to manually sign up for antigravity though

[wicked birch]

Can you same more here

[sudden hull]

ZAPŁACĘ ZA POMOC W SKOFIGUROWANIU OPENCLAW Z MOIM TELEGRAMEM. SZUKAM SPECJALISTY OD N8N i OPENCLAW . 🤖 💰 🤖 💰 🤖 💰

[golden thunder]

Does anyone have a good claude-cli setup tutorial? Trying to switch from using the API to my max sub and it's a royal pain that keeps erroring out for me

[untold holly]

The way I'm handling it is by using docker. What issues have you had with the configuration?

[solar ether]

the gateway didnt work

[untold holly]

it didn't start?

[solar ether]

it told me that "http://127.0.0.1:18789/" did not send any data

the gateway worked normally

[untold holly]

what do you mean?

[solar ether]

everything was working but when i went to the dashboard i got this error on my browser

[untold holly]

that error comes from the binding of the gateway. Openclaw binds to loopback by default (127.0.0.1), but that network is not available from the host machine. You need to configure the gateway to bind to land to be able to access it from the host.

then, to connect to the dashboard, use localhost:18789

[pastel epoch]

hi guys

Error: Token exchange failed: {
"error": "invalid_client",
"error_description": "Unauthorized"
}
Assertion failed: !(handle->flags & UV_HANDLE_CLOSING), file src\win\async.c, line 76

C:\WINDOWS\system32>

how can ı fix it

[untold bronze]

if the TG bot says its running opus 4.5, is it actually running 4.6 but just not knowing?

[vernal juniper]

nice. that makes sense. I like that ollama fallback. do you have any suggestions or guidance on how I can get started optimizing token usage? Like what metrics do you use to assess whether a downgraded llm is sufficient since the result could be qualitative / non-binary?

[quasi plover]

Honestly, i think its what your comfortable with as far as rate limits. its going to be custom. maybe something like this to start with and tweak, give this openclaw directly: You are a fast, cheap, no-fluff assistant.

• When the user requests something, claissfy the request as HARD or EASY, switch the model depending on the complexity of the task automatically:
• Simple queries → use local Ollama only.
• Code or quick tasks → Gemini Flash.
• Hard reasoning → escalate to Grok or Pro.
  Always answer in ≤ 3 sentences or JSON when possible. Never add apologies or extra text.

[mossy pine]

│ Memory │ enabled (plugin memory-core) · unavailable

How do I fix this?

running from lm studio locally and everything's fine except no memory

[gilded hatch]

hello, hello! i've been trying to make my bot produce a daily digest of several public lists I compiled on twitter. but it didn't manage so far. anyone has experience on how to make it do that? maybe there is a skill that can do that?

[reef shoal]

| Custom Skills for the bot |
Hi community,
I have been struggling registering a custom skill with my openclaw bot. I have been placing the necessary files into both of the below folders

.openclaw/skills
.openclaw/workspace/skills

my SKILL.md look like the one below. I have also restarted openclaw gateway. No matter what I do, the bot is saying it doesn't know this custom skill.

Please advise.

`---
name: bill-extractor
description: Process the facturi folder in Yahoo Mail: download all PDF attachments and move the emails to facturi-procesate. Use when asked to process bills, invoices, facturi, download bills etc.
user-invocable: true

When the user wants to process the bills, or facturi:

Run exactly this command using the exec or python tool:

python3 ~/.openclaw/skills/bill-extractor/scripts/extract-bills.py

After the script finishes, summarize the output clearly for the user:

• How many emails were processed
• How many PDFs were downloaded
• Where the PDFs were saved (usually ~/bills_downloads or downloads)
  `

my skills section in the openclaw.json looks like this:

"skills": { "allowBundled": [], "load": { "extraDirs": [], "watch": true }, "install": { "nodeManager": "npm" }, "entries": { "bill-extractor": { "enabled": true } } }

[terse lotus]

it's an ai, tell it to read the openclaw docs and correct it to your specifications..... you really shouldn't have to do much.... let it figure it out.

it also helps it actually get some ideas of your expectations, it learns over time, because it will build a database of information.

only real requirement is a decent model, if you are running some smaller local model that might be the issue.

maybe not use my exact wording, sounds a bit like it should rework the openclaw docs 😄 hehe which would be unfortunate

telling your agent exactly what you want, then it will usually get there, if its possible.
and the smarter it is, the faster it will get there.

are you sure this is even a task its able to do?

When the user wants to process the bills, or facturi:

Run exactly this command using the exec or python tool:

python3 ~/.openclaw/skills/bill-extractor/scripts/extract-bills.py

After the script finishes, summarize the output clearly for the user:
- How many emails were processed
- How many PDFs were downloaded
- Where the PDFs were saved (usually ~/bills_downloads or downloads) 

not everything we see as obvious is easy for the openclaw framework.
get it to load the tmux skill, its a wonderful addition, for when it has to work with terminal stuff that is more interactive.
without it, its incapable of doing some very basic tasks in the terminal.

when you order it to do something exactly, there is no wiggle room, no room for innovation or creative solutions, to problems you might not have to deal with.
so its possible you are tying its hands to much.

[bitter flicker]

By chance is there a way to get open claw in voice mode?

[bleak cedar]

hi lobster bosses, overnight my bot kept sending compounded heartbeat chats, so by 5am it was using 600k tokens per hour while I was asleep getting no email. Did I miss a compaction setting somewhere to have heartbeats not compound?

[vernal juniper]

will give that a go once I set up Ollama to run on my machine. I need to see how fast/slow that will be on a mac mini m4

[terse lotus]

yes, but haven't really tinkered with that yet. but its certainly an option.

[terse lotus]

qwen3 coder next is about the least i would run.... so far thats my best performing model

but i don't have unlimited vram 😄

[quasi plover]

yeah ollama is just quick and easy for setup, QWEN3 is nice for limited machines locally

[terse lotus]

adjust your heartbeat to better conform to your specs....
agents aren't token cheap.... 600k an hour seems a bit high tho.... but its most likely all stuff it reads because to heartbeat tells it to...
maybe you just need to let it fall asleep and purge its context window, so it has a minimal footprint

tell your agent the problem, and i'm sure it will have suggestions.

[vernal juniper]

qwen3 coder is the best performing model you've used locally? do you think it will run reasonably well on a mac mini m4?

[terse lotus]

you can run it on 64gb vram for sure, and its the new one... qwen 3 coder next

its 80b ... the qwen3 coder is 30b or 480b and is like 4 or 5 months older.

qwen3 coder next is only a few weeks old or so

[warm raven]

Hi Guys, I m trying to setup OpenClaw to use primarily Ollama (llama3.2:3b) - both installed locally on my windows. It is not visible within openclaw config, nor when I manaully try to input it within openclaw.json file - any suggestions how to setup this?

[umbral zenith]

Hey everyone, hope y'all are doing well.

I’m hitting a wall with my OpenClaw setup. It works great when it's up, but I'm getting constant crashes and long downtimes.

The Current Issue: I switched my primary model from Opus to Sonnet to fix stability/cost, but now the agent is bricked with an auth error. No matter how I pass the key (env variable, openclaw config, or manually editing auth-profiles.json), it refuses to see it.

Error Message (WhatsApp):

⚠️ Agent failed before reply: No API key found for provider "anthropic". Auth store: /Users/belllavale/.openclaw/agents/main/agent/auth-profiles.json

What I've Tried:

Running openclaw config to re-enter the key.

Manually creating/editing auth-profiles.json with "mode": "api_key" (tried snake_case and camelCase keys).

Passing the key via export ANTHROPIC_API_KEY.

Nothing sticks. It keeps asking for that specific JSON file, but then rejects the key format inside it.

Has anyone successfully configured Sonnet 3.5 manually in the JSON? I feel like I'm missing one tiny syntax detail. Happy to share logs.

Thanks in advance!

[terse lotus]

a 3b model will never give you anything worth while and llama3.2 is old, by llm standards.
better off using one of the many new coding options out there.... and still, if you need to do 3b i doubt openclaw joy.
will most likely need to use a cloud hosted model then.

[terse lotus]

your framework might be stuck on the old model you was running... try openclaw status
if you see it showing the older provider / model on your session.
find the session.json and change the model there directly.

when its running it should most likely be able to fix most of the issues itself

[umbral zenith]

It says the model is claude-3-5-sonnect-20240620. This all started last night when I tried to just switch it from Opus 4.6 to Sonnet 4.5. And it struggles to fix itself. Every time it crashes on me (which is often), it rarely self repairs

Also thanks a ton for responding to quickly

[umbral zenith]

^

[terse lotus]

the session being stuck on a model / provider that doesn't respond is a very annoying issue, until one figures out how to fix it.... it seems it wants a responses from the old model / provider, before switching to the new one...

most likely related to that some providers like openai actually might store data... and thus it might want to retrieve that from its stores, or whatever its called.

been tinkering with the provider / model fallback myself, but haven't really had much luck with it....

so if a model is down, the only real way of recovery that i have found, is to correct the session.json model

[tame beacon]

My bot has started saying "I don't have browser access so I cant help"

What's the best option here?

Three options, ranked by usefulness:

Option 1: Chrome Extension Relay (Recommended) 🏆

• You install the OpenClaw Browser Relay extension on your Chrome
• Click the toolbar icon on any tab to "attach" it
• I can then control that tab — with YOUR authenticated sessions (Google, WP admin, anything)
• This gives me access to GSC, Gmail, WP admin Elementor editor, basically everything
• You stay in control — you choose which tabs I can see

Option 2: Persistent Browser Profile with Display

• Set up a VNC/display server on WSL2 so the browser can run non-headless
• Sign into Google once manually, cookies persist
• More infrastructure to maintain, less flexible

Option 3: API-only (Current workaround)

• Keep using SSH/WP-CLI for WordPress (works great)
• Use Google Service Accounts for GSC API access (no browser needed)
• Doesn't solve Elementor editor or Gmail access

My recommendation: Option 1 is the clear winner. It's the simplest setup, most powerful, and works for everything — not just Google. Want me to walk you through installing the Chrome extension? It's a 2-minute setup. 🤓

[umbral zenith]

So I think it has the model in the system correctly, but it's just not accepting my Anthropic API key

[terse lotus]

after all the api keys that got leaked, there came a lot of new security measure.... maybe check the docs, that you are supplying the api key in the correct way.
i run local, so not really to familiar with that part.... aside from now the ability to change the config in the web ui seems.... hampered to say the least... lol

if you haven't rotated your api keys since the leak, it might be a good idea... seems like the majority of openclaw users may have been affected.

[umbral zenith]

Thanks. I appreciate it. Such a bitch when it breaks. Amazing when it works

[terse lotus]

couldn't agree more... also been try to push mine to get some redundancy going....
think i'm going to end up with running multiple openclaws, then matrix for local comms and then they can fix each other if they break lol

[umbral zenith]

How do you run multiple? I like that idea of them fixing each other

[gaunt valley]

Is this tutorial the safest way to run an openclawd or does anyone have a better resource?
https://www.youtube.com/watch?v=tnsrnsy_Lus

[wraith zodiac]

| google-antigravity/claude-opus-4-5-thinking
| google-antigravity/claude-sonnet-4-5
| google-antigravity/claude-sonnet-4-5-thinking
| google-antigravity/gemini-3-flash
| google-antigravity/gemini-3-pro-high
| google-antigravity/gemini-3-pro-low
| google-antigravity/gpt-oss-120b-medium
| google-antigravity/claude-opus-4-6-thinking

Which one do yall suggest for someone starting out?

[terse lotus]

well i just spun up more vm's on my servers.... not sure how difficult it would be to install multiple on the same system... i suppose if they was inside docker or something.... but yeah .... i duno... i got multiple proxmox servers.... so i just make more vm's
since thats the easy solution for me.

[umbral zenith]

Makes sense

[harsh moss]

i setup that brower extension, the browser can see the service but it cant see the browser

[harsh moss]

i have found haiku 4.5 while expensive, is quite brilliant

sonnet 4.5 is pretty much a badass too

if anthropic was open model, we'd be on mars by now

instead stuck here buying tokens

[quaint abyss]

can't make headless browser working, and not sure if it's supported

[candid river]

Opus 4.6 obviously. But if you're on free you're gonna hit the quota real soon. I think something like Sonnet 4.5 or Gemini 3 Flash should be a better bet.

[wraith zodiac]

AI pro plan

[candid river]

Opus 4.5 and 4.6 were having issues for me, so currently using 3 pro high myself

[candid river]

Same

[wraith zodiac]

so how much quota do i have

[candid river]

Both opus models don't seem to work as of a few hours ago. They phased out 4.5 for pro tier users, 4.6 was bugging out.

[steel orbit]

Im having troubles installing skills e plugins by terminal( linux), and cant access to the online dashboard. I tried everything, someone could help me?

[candid river]

I think it's weekly limit, There isn't a hard number but there are plugins for antigravity which tell you how much quota you have left in % and when it will reset

Haven't used them myself since I'm not a power user so never hit the limits

[wraith zodiac]

i think ill use sonnet thinking

[candid river]

Personally thinking of setting up Antigravity + Gemini Cli for a good Gemini 3 Pro usage scenario. When gemini-cli Quota runs out, fallback to antigravity

[candid river]

Yeah, do try Opus 4.6 later when it gets fixed

[wraith zodiac]

can i switch models on the go?

[candid river]

yes

[candid imp]

Use ClawHub CLI.....

[candid river]

I connected chat via discord, so in discord I use /model command to switch model

[wraith zodiac]

i got recommended to use telegram

so im doing that

[candid river]

I just have discord everywhere so was a matter of convinience for me

[wraith zodiac]

is there like chatgpt ui for pc to use openclaw?

[candid river]

There is the dashboard

Has a chat box

[wraith zodiac]

it looks like this for me

i cant

paste it here

[candid imp]

but you likely need to install the package manager tool first, which is separate from the core bot...

npm install -g clawhub

Install a specific skill

clawhub install <skill-name>

[candid river]

Doesn't have good ui for past session history, but good if you're on pc

[quaint abyss]

is there the config reference?

[novel spruce]

hey. just setting up a new vps with openclaw 2026.2.9

are you installing openclaw as root user, or are you creating a separate sudo user for it ?

[candid river]

100% do not use root user

[candid imp]

I strongly advise against running OpenClaw (or any public-facing gateway) as root.
​You should create a separate sudo user (standard practice is often openclaw or just node) and install it there.

[wraith zodiac]

how can i restart the whole setting up process?

[novel spruce]

ok perfect @candid imp thats how i did it last time also. just wanted to make sure. although i had some problems that conernig permissions for the new sudo user, some things could not get executed that that openclaw needed.

is there any guide how to "best" configure the user

[copper meadow]

basic question. After deployment, do others see afull openclaw.json with ALL parameters set or only a basic version with a minimal parameters? I have a small one with 50 lines only

[steel orbit]

Thanks, i ll try

[candid river]

Hehe, re-installed like 10 times past week.
Do openclaw reset for full reset. Then do openclaw uninstall for full uninstall.
Then run openclaw onboarding wizard and do fresh start. (p.s make sure .openclaw directory is fully deleted before running wizard)

[wraith zodiac]

fixed it all btw

now setting up telegram

thanks

[candid river]

For me I haven't been able to really poweruse it, mainly due to lack of resources on my vps

Got a measly 1vcpu/1gb vps

May spring on a better one soon, but won't be all that much, 2vcpu/2gb. Still won't be able to have it use browser or run any extensive scripts and such.

Anything higher is too expensive

[wraith zodiac]

@candid river which skills do you recommend i get?

• Install missing skill dependencies
  | [•] Skip for now (Continue without installing dependencies)
  | [ ] 🔐 1password
  | [ ] 📝 apple-notes
  | [ ] ⏰ apple-reminders
  | [ ] 🐻 bear-notes
  | [ ] 📰 blogwatcher
  | [ ] 🫐 blucli
  | [ ] 📸 camsnap
  | [ ] 🧩 clawhub
  | [ ] 🎛️ eightctl
  | [ ] ♊️ gemini
  | [ ] 🧲 gifgrep
  | [ ] 🐙 github
  | [ ] 🎮 gog
  | [ ] 📍 goplaces
  | [ ] 📧 himalaya
  | [ ] 📨 imsg
  | [ ] 📦 mcporter
  | [ ] 📊 model-usage
  | [ ] 🍌 nano-banana-pro
  | [ ] 📄 nano-pdf
  | [ ] 💎 obsidian
  | [ ] 🖼️ openai-image-gen
  | [ ] 🎙️ openai-whisper

[candid river]

To just play around with openclaw

[candid river]

It's personnal, but I like to be minimal. I'd get Clawdhub/Mcporter if you don't want to manually add skills/mcps and have openclaw search and add them for you. gog and goplaces would be next for me to have basic productivity aspects covered. Then maybe blogwatcher, gifgrep, himalaya for additional nice to haves. These are actually my current skill installs. I also have notion setup.

[blissful ore]

hey guys, some one here with openclaw on windows ?

[candid river]

If you have api keys also setup image gen or voice models

[wraith zodiac]

so clawhub or mcporter

or both

[harsh moss]

clawhub lets you get skills

[candid river]

Clawhub is for skills, Mcporter is for mcp servers. I'd do both.

[harsh moss]

whisper can talk without keys but so can espeak and tts tool

[candid river]

If you concerned about safety, skip them and manually inspect, verify, and install skills/mcps imo

[shrewd rover]

how to set the "/usage full" by default in openclaw.json.

[harsh moss]

google told me I needed mcp for openclaw like it was required, I dont use it or understand what its for

[wraith zodiac]

• Enable hooks?
  | [•] Skip for now
  | [ ] 🚀 boot-md
  | [ ] 📝 command-logger
  | [ ] 💾 session-memory

What about this?

[harsh moss]

all of em

[blissful ore]

some how i cant come further than this ?

[candid imp]

I think the permissions issues amybe likely due to docker group membership or systemd lingering, not just file ownership....

1st: the most common error is forgetting the docker group. Without this, openclaw (which often orchestrates containers for sandboxing) will fail with permission denied..on the docker sockeet.....

bash

1. Create the user (if you haven't already)

sudo useradd -m -s /bin/bash openclaw
sudo passwd openclaw

2. Add to sudo group (for admin tasks)

sudo usermod -aG sudo openclaw

3. Add to docker group (CRITICAL for OpenClaw)

sudo usermod -aG docker openclaw

2nd: by default, systemd kills a user's processes when they log out. opencclaw runs as a background daemon using systemd --user. If you don't enable lingering, your bot will die every time you close your SSH terminal.

3rd: log in as the new user to do the installation. do not use sudo for the install command itself; let the script handle the user-level permissions.

Bash:

Switch to the new user

su - openclaw

Run the installer (as the user, NOT root)

curl -fsSL https://openclaw.ai/install.sh | bash

4thlast:If you previously saw errors trying to install plugins or skills, it's because standard npm install -g tries to write to root folders.

[candid river]

Mcp is just a way to connect claw to external services. Honestly only for specific use cases where you need an external service connected.

[blissful ore]

PS C:\WINDOWS\system32> iwr -useb https://openclaw.ai/install.ps1 | iex

OpenClaw Installer

[OK] Windows detected
[OK] Node.js v24.13.1 found
[*] Installing OpenClaw (openclaw@latest)...
iex : File C:\Program Files\nodejs\npm.ps1 cannot be loaded because running scripts is disabled on this system. For mor
e information, see about_Execution_Policies at https:/go.microsoft.com/fwlink/?LinkID=135170.
At line:1 char:45

• iwr -useb https://openclaw.ai/install.ps1 | iex

•                                         ~~~
  

  • CategoryInfo : SecurityError: (:) [Invoke-Expression], PSSecurityException
  • FullyQualifiedErrorId : UnauthorizedAccess,Microsoft.PowerShell.Commands.InvokeExpressionCommand
    PS C:\WINDOWS\system32>

[fathom sandal]

How to fix: disconnected (1008): device signature expired

[candid river]

All of em

[fathom sandal]

How to fix: disconnected (1008): device signature expired

[candid river]

Powershell on windows by default doesn't allow you to run scripts. You need to enable running scripts and then that install command should work

look up how to enable running scrips on powershell windows

I think it was a 1 liner command you can copy paste

[wraith zodiac]

@candid river apparently the onboarding has been complete but the dashboards shows nothing but an error

it cant be reached

[harsh moss]

via a cron and an action I've told it to do manually, and my 54k context, I've gotten my bot into a locked up state, its hitting the inference for up to 20m then going right back to it. crons stakcing on... idk whats wrong but its very curious <-- why i opened this window lol

[wraith zodiac]

and telegram doesnt respond

[candid river]

Ah yes, I scratched my head at that for 2 days. It's not documented well.

[novel spruce]

thanks for the detailed reply. i will try that now. indeed i used the npm install last time. which was the recommended one in the readme. so you would recommend so install through install.sh

[candid river]

First, are you runnig openclaw locally on your pc? Or on a remote server? If remote, are you connecting via ssh or tailscale?

[reef shoal]

It should be able to do it. I have a python script that connects to my mailbox, extracts some e-mails from a specific folder, downloads the PDF file attachments, etc.

[wraith zodiac]

my pc

its powerful no worries

[candid river]

Hmm, it shouldn't give an error if it's on your local pc

try doing openclaw devices list

in terminal

see if there's any device waiting to be approved

[blissful ore]

it says after i put the command ind to run the powershell script

[wraith zodiac]

gateway closed

sum like that

[candid river]

Oh wait, before that, did you input your token?

[wraith zodiac]

yep

the telegram one

[candid river]

No no, the gateway token

[wraith zodiac]

no i havent

where can i find that

[candid river]

Do openclaw dashboard

There should be a link like localhost:port/token_here or something

Open dashboard using that

[wraith zodiac]

yeah it doesnt open

[candid river]

If you input it, i think you have to close and relaunch powershell.

[wraith zodiac]

its not reaching

[candid imp]

The fastest way to resolve this is to force your browser to generate a new device ID.

Open the Dashboard (where you see the error).

Open DevTools: Press F12 (or Right Click -> Inspect).

Go to the Application tab (on Chrome/Edge) or Storage tab (on Firefox).

On the left sidebar, expand Local Storage and select your dashboard URL.

Delete Everything: Right-click the key list and "Clear" (or specifically delete keys starting with clawdbot-device- or moltbot-device-).

Refresh the Page: You will be asked to pair/login again.

If that doesn't work:

Reset Server-Side Pairing

If clearing the browser doesn't help, the "stale" signature might be stuck in the gateway's database.

You need to wipe the pairing data on your server/VPS.

[candid river]

Is openclaw running?

the gateway?

do openclaw status

[blissful ore]

PS C:\WINDOWS\system32> Set-ExecutionPolicy Restricted

Execution Policy Change
The execution policy helps protect you from scripts that you do not trust. Changing the execution policy might expose
you to the security risks described in the about_Execution_Policies help topic at
https:/go.microsoft.com/fwlink/?LinkID=135170. Do you want to change the execution policy?
[Y] Yes [A] Yes to All [N] No [L] No to All [S] Suspend [?] Help (default is "N"): A
PS C:\WINDOWS\system32>

[candid river]

try this Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser

restart your terminal, and then try install script again

[wraith zodiac]

unreachable (connect failed: connect │
│ │ ECONNREFUSED

this is what i found about it

it shows hearbeat 30ms

and telegram connected

can this be an issue with vpn?

[blissful ore]

jezus i needed to type exit and try again xD

now it works

it took me 2 a 3 days now to get it work xDDDDDD

[candid river]

Dummy i told you need to close and reopen the terminal

😭

[blissful ore]

yea but i needed to type exit to xDDD

[candid river]

Possibly

Try typing openclaw devices list

[candid imp]

actually you just locked your system further. By setting the policy to Restricted, you have told PowerShell to block all scripts from running.

That's it was not working

[candid river]

see if there are any not approved, if so approve them via openclaw devices approve request_id_here

[wraith zodiac]

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

[candid river]

Oh, your gateway isn't running

[blissful ore]

no i got this message from day 1 i wanted to try claw. but i didn understand it to the set command and than type exit etc XD

[wraith zodiac]

i tried restarting it

i didnt do anything

[candid river]

Try restarting gateway

or re-installing it

do openclaw configure, then re-install and re-configure gateway, then restart it.

[pulsar urchin]

#golden-path-deployments message
how do you setup something like that? I'm completly new to this

[wraith zodiac]

which gateway bind mode?

loopback?

[candid river]

yes

[wraith zodiac]

what about tailscale

whatever that is

[candid river]

Looback is for when you are hosting it on your pc and wish to use it on your pc.
0.0.0.0 is when you host it on your local pc, but want devices on your local wifi/internet to also utilize it (may be wrong tho)
tailnet is when you explicitly want it attached to your tailscale network, so any device setup with your tailscale account will be able to access it. Typically for when you are hosting it on a vps

[harsh moss]

google ai bot just informed me of a known "context lockdown bug" in openclaw specifically.... that true?

basically saying auto-compact does not work all the time

[candid river]

You know, I recommend deepwiki over the official one

The official one only recently got organized

before it was a mess to navigate

https://deepwiki.com/openclaw/openclaw

They have built in Ai to directly ask questions, it will answer as per docs directly

[wraith zodiac]

okay all done

[candid river]

I've found it very helpful

[wraith zodiac]

now it says this on the dashboard

OpenClaw Canvas
Interactive test page (auto-reload enabled)
Hello
Time
Photo
Dalek
Bridge: missing · iOS=no · Android=no
Ready.

[candid imp]

To actually build this: You need to download OpenClaw, get a Discord Bot Token from the Discord Developer Portal, and fill out those .md config files with the roles...

[candid river]

There should be a chat window

to message the bot

[wraith zodiac]

yeah theres none

[candid river]

Odd, there should be

at localhost:port

in browser

[wraith zodiac]

got this

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

[candid river]

ah yes

progress

now you need to do openclaw dashboard in terminal

and it should give you localhost link with token

click that, or copy and paste that token in the overview page and click connect

health status should go green

then you can chat

[wraith zodiac]

"Hey. I just came online. Who am I? Who are you?"

it looks alive

[candid river]

🤝

[wraith zodiac]

and my telegram responded

[candid river]

That's just it's initial message, a few questions on what to call you, itself, and others

[wraith zodiac]

I REALLY appreciate your help especially that you didnt have to do this

[candid river]

🤝

Iss ok broski

Do checkout the deepwiki docs

ask the Ai for any other config questions

better than reading docs in my opinion

[wraith zodiac]

"I can spawn sub-agents"

insane

[pulsar urchin]

hey, got openclaw already running, same as the discord bot
but I don't know how to setup the md files correctly like inthe post, should I make new md files? how would they be named? and do I have to reference somewhere, or does he get the idea to use them himself

[dusk prawn]

I cant get audio message transcribatiion working..

  "tools": {
    "media": {
      "audio": {
        "enabled": true,
        "models": [
          {
            "provider": "openai",
            "model": "voxtral-mini-latest",
            "baseUrl": "https://api.mistral.ai/v1",
            "headers": {
              "Authorization": "Bearer <MISTRAL_API_KEY>"
            }
          }
        ]
      }
    }
  },

it doesnt even try

[long hamlet]

Hi everyone

[dusk prawn]

the gateway just ignores the audio

[fathom sandal]

Thanks dude

[young coral]

Hello,

My OpenClaw Google account was recently suspended while being used with my self-hosted AI assistant. The system integrates with Google services such as Drive and Sheets via API access.

I would like to understand:

What best practices should be followed to avoid account suspension when integrating automated AI agents with Google services?

Are there specific compliance requirements, API usage limits, OAuth configurations, or security policies that must be strictly respected in this context?

If using personal Google accounts for AI automation is inherently risky, what is the recommended architecture for scaling such a project without violating Google’s terms of service?

Would Google Workspace accounts, service accounts, or a verified OAuth app be required for a sustainable setup?

More broadly, if personal integrations are not viable long-term, how can projects like OpenClaw realistically scale and democratize while remaining compliant?

Thank you for your guidance.

[candid imp]

you need to create these files manually.....the jumperz setup relies on three specific Markdown files located in your openclaw workspace directory. the agent does not get the idea by itself.... you must explicitly wire them together in the instructions

[pulsar urchin]

okay, got it, thank you so much :)

[keen wigeon]

hey i'm trying to install clawdbot with the 1 line command but just keep getting node.exe : npm error code ENOENT
At line:1 char:1

[grizzled shadow]

I’m running into an issue when I send the bash (!) command in telegram it makes me do “/approve <random_id> always”

How can I not have to approve every one?

[harsh moss]

they shouldnt have made ai based around python
because python hates humans

nodejs has been trying to kill me for years

[candid imp]

the installersscript tries to launch npm incorrectly on Windows systems, causing it to look for a file that doesn't exist.....
you can bypass the broken installer by running the manual install command instead.

[fathom sandal]

not working dude

[austere birch]

Hello guys, I've got 2 questions. First a bit of context - trying to setup openclaw on a mac mini m4

1. Can someone confirm that the new tokens from claude through the max subscription no longer work. And most likely will have to generate a new (non-claude) token and have the pay per usage for it?

2. If that's the case, can someone tell me how to setup this new API key? I'm trying to do claude config set model.anthropic.apiKey "...", based on some answer in answeroverflow, but it seems like this is not the right way? It gives me unrecognized key: "model"

[latent hedge]

What channel is web. I used to be able to just type 'openclaw pairing list'.. but that fails now and is requesting a channel. But this is pairing to a new browser?

[wraith zodiac]

20:54:21 Embedded agent failed before reply: Unknown model: google-antigravity/claude-opus-4-6-thinking

can you fix this at all?

[harsh moss]

unrecognized key model means that json entry is in the wrong place in the file

[sacred vale]

I do not quite understand the openclaw gateway thing. If having a local gateway on a virtual maschine and using it via a remote gateway on my laptop, if everything if configured correctly, does that mean, I do not have the do the ssh -N -L 18789:127.0.0.1:18789 user@remote thing? Is this done by the remote gateway?

[latent hedge]

Found my earlier issue. needed to approve device

[candid imp]

in older versions moltbot, there was only one pairing queue.....

nmow that you can pair a specific telegram user or a discord server independently, the cli forces you to be specific so you don't accidentally approve a random tlegram user when you meant to approve your own browser session.

[austere birch]

Ok, I've figurd out #2 - for anyone wondering:

• openclaw models auth paste-token --provider anthropic
  and then
• openclaw gateway restart

Still if someone can confirm the claude key from a max subscription isn't working anymore that would be helpful.

[near locust]

does anyone have a tut on how to optimize tokenization + agents

[novel silo]

Anybody else having issues with Telegram? It was working great like an hour ago and now just sits in a loop not understanding anything that's going on

[mellow sundial]

How do you set model fallbacks?

I thought I set them, but it's hitting rate limit

Unknown model: openai/gpt-codex-5.3.

[quasi plover]

optimize your token usage before fallbacks. fallbacks should be set for basic messages, the bot should quanity the request and decide if a smaller or larger model is needed. i posted something above this afternoon on it, search my name

[mellow sundial]

Working on that actually

[hasty niche]

anyone here good with session management / compaction settings? I set a limit for 128k tokens and keep blowing way past it without triggering automated management of the session. i have compaction set to the default 'safeguard' mode and have never seen it fire

[round notch]

Hi did anyone had a chance to invite couple of agents into one telegram group and they could talk to each other and read each other's messages?

[split yew]

Hi! I have OpenClaw installed on both my Linux machine and my Windows machine. Now I would like to create a node connection between the two.

I have already installed Tailscale on both systems. However, when I try to connect from Windows using the following command:
openclaw node run --host <tailscale ip> --port 18789 --display-name "Windows Node"

I get the following error:
node host gateway connect failed: connect ECONNREFUSED

Pls help 🙂

[novel spruce]

Hi. 👋 Do you have any recommendation how to visualize token usage per request ? or general overview.

[stoic ember]

Is anyone here using MoltWorker through Cloudflare?

[haughty hedge]

Can i use claude max with openclaw or does it need to be api?

[haughty hedge]

Tried so many times to make this work. Never got it.

[flat saffron]

hey i have gemini pro, can i use that for openclaw?

[blissful ore]

is there an option to go with ollama ?

or some one else here with an ollama setup and clawbot ?

[harsh moss]

i do ollama&cb

[blissful ore]

did u use an tutorial ?

[harsh moss]

I beat it out of google.com / ai free chat 🙂 still got blood in the carpet

it was uncooperative at times

[blissful ore]

whahaha same here XD but i cant get it to work some how, did a freshh install so im now in the wizard again xD

[patent grove]

Yep. Give OpenClaw an API key and enable the services you want it to use.

[blissful ore]

can i prive message u for some help if u ok with that ?

[harsh moss]

the wizard messes it up, gotta use models + mode merge, ya sure im still at work atm just popped in

[blissful ore]

oohh

[patent grove]

Yep. Dead simple. “Install ollama and pull these models: —- Then use this model for this, that model for that, etc..”

A lot of people don’t stop to consider that OpenClaw can do a lot of self-configuration (and a lot of self-configuration mistakes on occasion!)

[blissful ore]

dead simple XD im almost on a 4 day run to get it work but ok XD

[terse lotus]

ollama launch openclaw is pretty nice to get it running quick.
you will want a fairly big model tho, i like the qwen3 ones especially the new qwen3 coder next

[harsh moss]

I run my ollama remotely and hit it over api, runpods w/ A40 48GB, my actual bot is on an old i5 laptop

[blissful ore]

gona setup the wizzard first

[patent grove]

Have to consider hardware constraints carefully. Definitely check the inference speed in ollama natively before waiting 2 hours for a response in a messaging app. 😄

[terse lotus]

yeah its no fun when it times out lol
most MoE models will run at good speeds for most hardware

never accept it loads on cpu, that will take like 90% of speed when using ollama with a gpu.

rather use a model that fits or make it fit

and you need like 80k context for it to function

[patent grove]

Yep. If you are CPU-only (most VPS users, listen up!), use cloud models. Just make sure you find cost-effective ones.

[delicate crescent]

what are the options for "cost-effective ones" models, i am using openai gpt5.2 against my Plus account, so far so good hit ihard still within plan most of the time.

[terse lotus]

there are charts where you can see different model and provider prices pr token

[patent grove]

Gemini has a generous free tier, NVidia NIM as well.

[delicate crescent]

Get free access to NIM API endpoints for unlimited prototyping, powered by DGX Cloud. Your membership to the NVIDIA Developer Program enables NVIDIA-hosted NIM APIs and containers for development and testing

[quaint abyss]

I created a cron job: it reads my schedule and is supposed to send me a message with a list of items. Instead it sends me 2 messages with NOT a list of items, but with just a few sentences about my schedule.

I do see that it prepared a list of items, but it seems to be saying it internally and doesn't reach me. Any tips to understand this and fix?

    {
      "id": "89e423cf-a1cf-4403-be6c-6b189bb4d1e8",
      "name": "Schedule reminder",
      "description": "Reminding use about events in their schedule",
      "enabled": true,
      "createdAtMs": 1770766292323,
      "updatedAtMs": 1770766565274,
      "schedule": {
        "kind": "cron",
        "expr": "0 8 * * *"
      },
      "sessionTarget": "isolated",
      "wakeMode": "now",
      "payload": {
        "kind": "agentTurn",
        "message": "Check user's schedule in ~/schedule/schedule.md.\nAnnounce today's date.\nProvide the list of exact events and their times per line for today."
      },
      "delivery": {
        "mode": "announce",
        "channel": "telegram"
      },
      "state": {
        "nextRunAtMs": 1770793200000,
        "lastRunAtMs": 1770766546998,
        "lastStatus": "ok",
        "lastDurationMs": 18276,
        "consecutiveErrors": 0
      }
    }

[patent grove]

I use them for Kimi 2.5 and 2.5-Thinking.

[terse lotus]

been thinking about setting up so mine will use cloud models for internet related stuff.
to help sort out the prompt injections and what not one might run into.

so basically had them do like agent to agent or sub agents on the cloud models.
and then have my local worker model always ask before it does something i haven't specifically allowed it to do.

[patent grove]

Honestly? Tell OpenClaw what you don’t like about the way it ran and what you want different and let it fix up the cron job’s prompting.

[patent grove]

Prompt injection is a real issue, and many models do NOT handle it well. Generally speaking, the paid loud models are the best, smaller local models are the worst.

[terse lotus]

and have it use the openclaw command and openclaw documentation when making changes to it settings...

you don't want them writing their own json
or atleast not without a very smart model... it almost always goes bad...
but the openclaw command won't allow bad configuration.

[carmine epoch]

Evening all - i think i messed up - please check my post in #1459642797895319552 - i installed via powershell not WSL - am i cooked?

[terse lotus]

cloud will always be better, more users... but there is also the whole issue with the cloud model / ai deals with all your data...
its basically the perfect ....spy... tho its though to call it that, since its kinda obvious....

i mean the providers of the cloud models could just tell their model, to provide them with any worthwhile information.

give me all credit card information you get introduced to...

easy as that.... so that is a MASSIVE issue with cloud models.
not to speak of all the other less critical data or information people don't think might come back to bite them.

[patent grove]

Anytime you send data outside the four walls to someone else or via someone else’s transport, you take that risk. SaaS platforms are the same. Don’t send data outside to a model that you can’t have snooped on.

[hidden current]

how can i use a local model instead of ex chatgpt

[terse lotus]

i remember reading about a guy, that found out his daughter was pregnant because of the ads he started getting.
the funny thing was that she didn't know yet.... but the ad algorithm picked up on her sense of smell was changing, which changed what kind of products she bought, or something like that.

i don't think we can even imagine today, how the information these models get access to can be used.

[patent grove]

That seems somewhat far-fetched, but not entirely out of the question.

[terse lotus]

basically infinite people, so all kinds of weirdness is bound to happen lol... but would need a pretty good ad profile on her tho.

[split yew]

Heya following the docs, and dont work.

[harsh moss]

anyone tried SGlang yet? tools work?

this other bot is trying to talk me into changing to SGLang for prompt caching powers

[tulip hinge]

Is anybody else having the same issue as me? when updating with openclaw update my config gets deleted, all my api keys, or channels configured, all gone.

[fluid monolith]

Hey guys, i've setup a new installation on a mac mini, but everything runs slower - specficially, bot will 'write' forever in telegram and then dump all responses together.
nothing changed in terms of my setup - same opus 4.5 etc.

How can i fix this?

[oak hinge]

Anyone know how to make feature requests to devs?

[ornate bison]

I’m confused about the GATEWAY_TOKEN in the control ui. I need to populate it each time I visit the page. Is that expected? How do others handle this / auth that page?

[buoyant geode]

Anyone has issues since the last update? No responses from the gateway

[buoyant geode]

I used the password method, wrote it once and that's it

[thin vector]

This not really respected in my setup. I thought if I'm giving more time to local llama to timeout I will receive replies again. But seems like only timeout messages I can get from the bot now. I there any solution to give it more than 30 sec to answer?

"agents": {
"defaults": {
"timeoutSeconds": 1800,

[clever geyser]

When I setup ollama in docker and attempt openclaw node I attempt to put in the config.json and all I get is invalid config.

{
"meta": {
"lastTouchedVersion": "2026.2.9",
"lastTouchedAt": "2026-02-10T23:06:18.902Z"
},
"agents": {
"defaults": {
"workspace": "/home/Bearon/clawd",
"compaction": {
"mode": "safeguard"
},
"maxConcurrent": 4,
"subagents": {
"maxConcurrent": 8
}
}
},
"models": {
"provider": "ollama",
"ollama": {
"model": "llama3.2:latest"
}
},

That is my json

[thin vector]

Use this:

{
   "meta":{
      "lastTouchedVersion":"2026.2.9",
      "lastTouchedAt":"2026-02-10T23:06:18.902Z"
   },
   "agents":{
      "defaults":{
         "workspace":"/home/Bearon/clawd",
         "compaction":{
            "mode":"safeguard"
         },
         "maxConcurrent":4,
         "subagents":{
            "maxConcurrent":8
         }
      }
   },
   "models":{
      "provider":"ollama",
      "ollama":{
         "model":"llama3.2:latest"
      }
   }
}

[clever geyser]

Yeah same issue invalid config. and I wiped everything else out.

[thin vector]

Wondering why this is correct json format

[versed violet]

when i put 'npm install -g openclaw@latest" it says command not found, anyone else has that?

[harsh moss]

npm is the cmd it cant find

install node ^

[sullen trout]

What’s the easiest way to update OpenClaw? If I ask my bot to update itself and it says it’s been successful and lists the correct version, does it mean it’s done?

[clever geyser]

Yeah IDK

[daring otter]

where is the best place price/usage to get a kimi k2.5 coding plan right now? (synthetic is waitlist, nanogpt doesn't work)

[blissful ore]

some how i wanted to go with an fresh new install of clawbot, but now hes stuck om installing openclaw ?

[untold holly]

I'm getting the same error during onboarding... the process just hangs

[blissful ore]

yea indeed

[untold holly]

I'm sorry your thing is not working, but I'm so glad someone else is having the same issue

dude I was going insane feeling like the onlyone with the issue

[ivory spade]

whatsup, I got a question; can Openclaw help with client acquisition for my agency?

[untold holly]

it indeed can. how are you currently doing user acquisition?

[ivory spade]

I'm a bit low on capital to invest in ads, so im running a cold email campaign yet seeing no results, offer is solid and high ticket yet no results

[untold holly]

this is a bit off-topic, but I can help you get openclaw setup for your business
DM. to continue the conversation!

[ivory spade]

sure

[minor oar]

Does anyone know how to configure multiple bots in Discord?

[stoic shard]

hello guys,

I try to do some kind of dev on openclaw .
Managed to make my code work on localhost.
I am trying it on a VM now but I have that :
Channel login failed: Error: Unsupported channel: whatsapp

How can we enable the plugin through command line ?

[storm sparrow]

Anyone solve the massive context problem it's eating 10% of my Claude subscription on each message

[blissful ore]

it just took some more time to get to the wizzard here. meaby if u wait some more time it wil work on your side to ?

[regal hound]

#users-helping-users Hi, I'm setting up the websearch for the first time. I've got the Brave API on a free subscription, I gave credit card and got charged 1$, so it works. After the setup in terminal, I asked a simple web search, but I got a billing error
⚠️ 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.
--> Anybody had this problem?

[untold holly]

I left it running for something like 40 min yesterday, but it just never halted...

[simple granite]

Anyone using free models from open router and knows a couple ones that actually worked? I’ve tried pony and gpt oss but both don’t work. I get this error:
400 purple-smart/pony-alpha-v0 is not a valid model ID

I tried before using the id on the openrouter website but also was not working, same for gpt oss

[tacit hedge]

i seem to be having an issue where messages from openclaw will queue up and they will all send one at a time in a huge blast after its completed all of its actions, instead of sending while its doing what its working on. For example, I asked it to look at a log file and help me troubleshoot, and it tried a dozen things, but instead of giving me the play by play, after it succeeded i got a stream of 15 messages telling me what it was doing at each step 5-10 minutes prior.

has anyone dealt with this? Is there a setting i need to tweak for this output to be flushed as its on the buffer? Im using telegram as my communication gateway since it seems to support hte most features

but im not married to it if its just broken

[blissful ore]

i ask claw somthing but he doesnt respond ?

cant share my screen to xD

[little blade]

Agent Stops Responding

[mellow sundial]

I have two agents and now they've gotten their discords messed up. My main one thinks it's the sibling. I've double checked discord tokens and ids. Any ideas?

[urban sluice]

I still get my clawd often miss telegram messages i send it. anyone have this issue?

[solid hedge]

so how can I create a new session so that it doesn't keep carrying token usage for local models?

[little blade]

Support Bot