users-helping-users

[proven axle]

how do i do that

[delicate pewter]

#users-helping-users message

[proven axle]

thx

what i do now to start it

[astral kelp]

Hey everyone, I had given my openclaw a gmail account to use to create documents and email its main user (my wife). It got flagged for suspicious behavior or bot activity. My question is, has anyone had success in getting their appeal to go through? Or is google been declining all appeals, even though my use doesn’t break any of the rules the list on the page it directs me to. Thanks in advance.

[swift nest]

How do I get my bot to be able to interpret screenshots the same way ChatGPT does? Also, by default I've found that the dashboard chat doesn't accept images, is that typical?

[jagged nebula]

can i install clawdbot on docker and have it use a separete windows maschine?

[spiral socket]

Tell him to use an OCR

[jagged nebula]

also is it possible to install it without the docker install script

[echo thunder]

Did anyone start getting node-pre-gyp error when upgrading?

[waxen lily]

Is openclaw good at browser automation with Meta? Or is a third party site needed?

[delicate pewter]

"agents": {
  "defaults": {
    "imageModel": {  
      "primary": "INSERT_HERE"
    },

[proven axle]

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

[echo seal]

You need to have the doctor regen the token.

[small inlet]

Does anyone use the Kie.ai API to call image generation tools like nano banana pro?

[swift nest]

Anyone else have this problem. When "agents.defaults.model.primary" is set to "openai-codex/gpt-5.3-codex-spark" and "agents.defaults.imageModel.primary" is set to "openai-codex/gpt-5.3-codex" you get error “Instructions are required”. When I set "agents.defaults.model.primary" to "openai-codex/gpt-5.3-codex" the bot is able to deal with images. Codex-spark cannot interpret images but Codex can so I set codex to the default image model but getting error.

[final summit]

Do you know any clean mission controls for MacOs

[daring moon]

Still struggling with exec approvals. Until latest upgrade, I was at least getting the approval ID in Discord DMs and could approve tool actions there. Now I'm struggling to get an approval ID I can use to surface in both the TUI on the host machine or in Discord. Is there a setting I need to flip on in the .json to get approval requests to be printed in certain channels?

[vernal star]

Anyone else finding their claw timid? Asking a lot of questions before engaging in tasks? I’m on gpt-5-mini

[small inlet]

At first yes and then I saw some YouTubers that were giving it the “figure it out” prompt and since then, it’s definitely helped. Look into it.

[lucid jolt]

Hey all. I do my work on Discord channels (pvt server just me and claw), each channel is supposed to be another project, but they are loosely related... I'm not sure if the context is shared or not, but I started getting (Opus and Codex) cooldowns all the time. If I do /new can I get it to reconstruct a summary later from discord history?

[latent hedge]

it....is.... astonishing how many times my bot have fucked up its openclaw.json file

[drifting plume]

Hi all! Running openclaw on a ubuntu box running against my local ollama. I configured Discord, enabled a couple of skills and i can chat with my agent. However, i have the feeling something is missing. For example, when i ask my agent to check updates every night (from https://gist.github.com/velvet-shark/b4c6724c391f612c4de4e9a07b0a74b6#3-self-maintenance-updates-and-backups), my agent responds with: "I'll help you set up this maintenance routine. While I can't directly schedule tasks on your system, I can provide the commands and steps to implement this using standard tools: " and then starts to spit out shell scripts and cron commands. That's not what is supposed to happen, right? Is the agent not supposed to do all this for me? What am i missing here?

[torn stirrup]

https://github.com/rylena/awesome-openclaw hi guys i made a comprehensive guide to openclaw as a awesome list if u guys think its usefull pls drop a star

[warm thistle]

Anyone familiar with adding a brave search api key to openclaw?

[drifting plume]

2. Set it in your environment:
   • Via CLI: export BRAVE_API_KEY='your-key-here'
   • Via OpenClaw: Run openclaw configure --section web and store it securely

I still happened to have that on my screen. 😉

[whole fossil]

Im trying to install openclaw on windows (with powershell) and i get this error:

`node.exe : npm error code 1
At line:1 char:1

• & "C:\Program Files\nodejs/node.exe" "C:\Users\Administrator\AppData\Ro ...
• CategoryInfo : NotSpecified: (npm error code 1:String) [], RemoteException
• FullyQualifiedErrorId : NativeCommandError`

Anyone knows how to fix it? I can't find any help online.

[warm thistle]

Brave demands a payment method (credit card) and states it will not be charged if yoy stay within the request limits. But I am hesitant to fill in my credit card to barve seacrh api.

oyoyoy =you

[drifting plume]

Same here.

[torn stirrup]

just ask ur bot to add it for u

[drifting plume]

If thy supported PayPal i would take the gamble because you can always claim it back, but putting my card out there, no way.

[torn stirrup]

did u install node?

[torn stirrup]

i mean playwright mcp works pretty well for me

[hollow thunder]

Hey guys, my agent has a skill that’s basically just a python script that gets called with exec and different arguments to do various things… how can I make this into a proper skill so that agents without exec perms can still use the functionality?

[drifting plume]

Hi all! Running openclaw on a ubuntu box

[whole fossil]

openclaw npm error 1

[marble coral]

Does anyone use gl5 via the modal endpoint? I have some trouble as it rejects the api token I generated and i can't figured out why...
I get this {"error": "invalid token"}

[tranquil zodiac]

what kind of markets?

[hollow thunder]

Anyone got any ideas? I feel like an idiot… it sees the skill, but it only tells me that “tool xyz not found” when trying to actually use the tool…

[hollow trail]

Hey. You guys what Linux OS recommend? Debian or Ubuntu?

[hollow thunder]

My agent chose arch Linux and set everything up by itself

But it really doesn’t matter

[hollow trail]

OpenClaw old version was running in Ubuntu 24 LTS. First, I tried in Ubuntu 25, but it got buggy. Now I have tried to install Fresh OpenClaw on Ubuntu 24 LTS again, but it's not working. I gave up on Ubuntu and moved into Debiain. I am asking because on the tutorial page, Ubuntu was mentioned a lot of time.

[young lily]

ubuntu 24 lts is generally stable for openclaw — 25 is still fresh and a lot of Node/npm + system package combos can get weird there. If 24 LTS isn’t working now, it’s probably not the distro itself but a dependency mismatch

[young lily]

invalid token usually means either the key is being sent to the wrong endpoint or the header format is off ... r u passing it as Authorization: Bearer <key> or via a provider-specific header? also worth double-checking that the token was generated for the modal endpoint specifically ... some providers issue different keys for REST vs modal

[young lily]

can try Serper.dev — free tier, no card, easy setup openclaw configure --section web (set SERPER_API_KEY)

or: Tavily — also free tier, no card required, good for agent-focused search

both work as drop-in replacements for web search

[daring moon]

I figured this out for myself - in case anyone is interested, I needed to expand the contents of the approvals section in openclaw.json AND update the execApprovals data in channels.discord

The relevant docs are https://docs.openclaw.ai/tools/exec-approvals#approval-forwarding-to-chat-channels and https://docs.openclaw.ai/channels/discord#exec-approvals-in-discord

Setting my user ID as the DM target was the real key, so I get an actual approval UI - which I had never seen before - in my DMs (and not in my 'public' Discord rooms). Nice to not have to copy/paste a gnarly UUID.

[marble coral]

I am trying to use the curl example that they have here with a token generated from within the account settings for API
https://modal.com/glm-5-endpoint?section=requests

[earnest jay]

guys is there somewhere where i can find examples of best use cases of soul.md?

[young lily]

thanks for circling back with the solution .. DM approval UI is one of those things that's easy to miss
but makes a big difference once you find it ..

[young lily]

The token you generate from Modal's account settings is for
their REST API — but the gl5 modal endpoint uses a different
auth flow.

For the modal endpoint, you need to generate the token from
within the modal environment itself, not from account settings.

Try running this first to get the right token:
modal token new

Then use that in your curl request as:
Authorization: Bearer <token>

Also double-check the endpoint URL matches exactly what's in
the docs — gl5 endpoints are sometimes region-specific.

[young lily]

You're missing the exec tool being enabled. When exec is off,
the agent can only suggest commands, not run them.

In your openclaw.json make sure exec is enabled in your
agent's tool permissions. Once it is, ask it to set up the
cron job again and it'll actually run the commands rather
than just printing them.

[proven stag]

hello! Anyone else experiencing any problem with installation?

[young lily]

Hey! What OS are you on and what error are you seeing?
That'll help us point you in the right direction fast.

[young lily]

soul.md is basically your agent's personality and operating principles. Best examples I've seen: define how it communicates, what it prioritizes, how it handles uncertainty, and any hard rules. darth kermit's awesome-openclaw list (linked above) has some good examples — worth checking out.

[rapid glacier]

Subject: Issues updating OpenClaw on Hostinger VPS (Docker/Debian)

Message:
Hi everyone,

I'm having some trouble updating my OpenClaw instance. Here are my setup details:

Host: Hostinger VPS

OS: Debian

Installation Method: Docker (via Hostinger’s "Docker Projects" panel)

The issue:
When I log into the web interface, I see a notification: "Update available: v2026.2.21-2 (running v2026.2.19)". However, when I click on the "Update now" button, nothing happens. The page doesn't refresh, and the version remains the same.

Since I am using Hostinger's Docker manager, I'm not sure if I should be updating via the UI or if there's a specific terminal command I should run to pull the latest image without losing my data.

Has anyone faced this? Any guidance on how to perform a manual update in this environment would be greatly appreciated.

Thanks in advance!

[proven stag]

win11 and

The underlying connection was closed: An unexpected error occurred on a receive.

[young lily]

The UI update button is likely blocked by Hostinger's Docker
manager sitting in between. Since you're using their panel
rather than direct Docker access, the button probably can't
reach the Docker daemon to pull the new image.

Manual update via terminal:

1. SSH into your VPS

2. Find your container:
   docker ps | grep openclaw

3. Pull the latest image:
   docker pull openclaw/openclaw:latest

4. Restart the container (your data lives in volumes,
   not the container itself, so it's safe):
   docker compose down && docker compose up -d

If Hostinger's panel created the container, check their UI
for a "Recreate" or "Pull latest" option on the container
— some panels have this even when the in-app update button
doesn't work.

Your data should be safe as long as you're using volumes.
Worth double-checking your docker-compose.yml to confirm
the workspace is mounted as a volume before you do anything.

[young lily]

That's a TLS/SSL connection error on Windows. A few things to try:

1. Run PowerShell as Administrator and retry the install

2. If that doesn't work, update Node.js to the latest LTS
   version first — older Node versions have known TLS issues
   on Win11:
   https://nodejs.org/en/download

3. If still failing, try setting this before installing:
   $env:NODE_TLS_REJECT_UNAUTHORIZED=0
   npm install -g openclaw
   (temporary workaround — revert after install)

4. Best long-term fix: install WSL2 and run OpenClaw from
   there. Win11 + WSL2 is the most stable setup and avoids
   these Windows-specific TLS issues entirely.

What version of Node are you running? (node --version)

[young lily]

If your Ubuntu 24 LTS install wasn't working, it's almost certainly a Node/npm version mismatch rather than the distro. Worth checking: node --version should be 18–22.

[proven stag]

I didnt update the Node.js version and I just tried with the global command you told me. It worked! Thank you very much for your help!!

[young lily]

The skill registering but returning "tool not found" usually
means the tool name in your skill manifest doesn't exactly
match what the agent is trying to call.

Check two things:

1. Open your skill manifest and look at the tools array —
   what is the exact tool name defined there?
2. Make sure the function name in your Python script matches
   that exactly (case-sensitive)

Also — if your skill uses exec under the hood, make sure exec
is enabled in your agent's tool permissions or it'll see the
skill but fail to run it.

[young lily]

Open your browser and go to your dashboard URL
(usually http://127.0.0.1:18789).

Copy the token from the URL or from the dashboard settings
page, then paste it into Control UI → Settings → Gateway Token.

If you can't find the token, run this in terminal to regenerate:
openclaw doctor --regen-token

[young lily]

Known issue — codex-spark doesn't support imageModel as a
separate config. When you split model.primary and
imageModel.primary across different codex variants it breaks.

Fix: set both to the same model:
"model": { "primary": "openai-codex/gpt-5.3-codex" }
"imageModel": { "primary": "openai-codex/gpt-5.3-codex" }

Full codex handles images natively — no need to split it.

[young lily]

Each Discord channel gets its own context by default — they
don't share unless you explicitly pass context between them.

For cooldowns: Opus and Codex have rate limits per session.
/new starts fresh which helps, but won't reconstruct history
automatically.

Best workaround: ask the agent to summarize the channel
before /new, save that summary, then paste it as context
in the new session to rebuild continuity.

[young lily]

Add this to your soul.md:
"Attempt tasks autonomously. Only ask clarifying questions
if genuinely blocked. Default to action, not confirmation."

gpt-5-mini is naturally cautious — soul.md is the right
lever to fix it.

[swift nest]

Hey Joe, thanks for this. I ended up setting codex for both, and started working no problemo. I thought I was being smart trying to set codex-spark to primary.

[young lily]

Yes — you can chat via terminal with:
openclaw chat

If you're getting a gateway error, start the gateway first
in a separate terminal:
openclaw gateway start

Then retry. Your data is safe — it lives in
~/.openclaw/workspace and won't be touched by a reinstall.
If you do reinstall, that folder stays intact.

[heady gull]

My telegram bot won't process DM's when I troubleshoot, smoke test etc .it does receive I've been troubleshooting for 5hrs any suggestions

[hollow thunder]

The goal is to make it able to use the skills functionality but without having exec perms. I only want it to be able to use those specific tools, and not any other exec commands…

[young lily]

Google's bot detection has gotten aggressive with any
account sending automated emails. Appeals are hit or miss
but worth trying.

A few things that improve appeal success:

1. Be specific in the appeal — describe exactly what the
   account was doing and why it's legitimate
2. Avoid the word "bot" or "automated" in your appeal

Longer term: consider creating a dedicated Gmail account
just for OpenClaw use rather than a personal one — keeps
your main account safe and sets clearer expectations with
Google's systems.

[young lily]

Yes — you can run OpenClaw in Docker on one machine and
access it from another on the same network. Point your
client to the host machine's IP instead of 127.0.0.1.

For installing without the Docker script:
npm install -g openclaw
works directly if you have Node installed. Docker is
optional — not required.

[young lily]

node-pre-gyp errors usually mean a native dependency
needs to be rebuilt for your current Node version.

Try:
npm rebuild
npm install -g openclaw@latest

If still failing, check your Node version —
OpenClaw runs best on Node 18–22.
node --version

[young lily]

OpenClaw's browser tool can handle basic Meta UI interactions
but Meta actively detects and blocks automation. For anything
sustained you'll likely hit rate limits or CAPTCHAs.

For serious Meta automation a dedicated tool like
Playwright or Puppeteer wired into OpenClaw via a custom
skill is more reliable than the native browser tool alone.

[jagged nebula]

I meant installing on docker without using focker-install.sh

[young lily]

Classic issue. Two fixes:

1. Add this to soul.md:
   "Never directly edit openclaw.json. If config
   changes are needed, ask me first."

2. Keep a backup:
   cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak

Run that backup command after every config change you
actually want to keep.

[young lily]

To install OpenClaw in Docker manually without the
install script, create your own Dockerfile:

FROM node:20-slim
RUN npm install -g openclaw
WORKDIR /home/openclaw
EXPOSE 18789
CMD ["openclaw", "gateway", "start"]

Then build and run:
docker build -t openclaw .
docker run -d
-p 18789:18789
-v ~/.openclaw:/home/openclaw/.openclaw
--name openclaw
openclaw

The volume mount keeps your config and workspace
persistent across container restarts.

[astral kelp]

Thanks. I did that, I have workspace I can use but prefer to keep separate. Appreciate it.

[young lily]

That "activating" state that immediately exits means the
gateway process is crashing on start. Usually a config or
dependency issue.

Try:
openclaw doctor

That'll diagnose what's wrong. If it points to a config
issue, check your openclaw.json for any malformed entries
— a bad JSON syntax will kill the gateway on startup.

Also try:
openclaw gateway start --verbose

to see the actual crash reason.

Classic nvm issue — when you switch Node versions via nvm,
globally installed packages don't carry over.

Fix:
nvm use <your-version>
npm install -g openclaw

Then verify:
openclaw --version

If you switch Node versions again later you'll need to
reinstall openclaw in that version too.

[jagged nebula]

I hope you're aware that openclaw requires a minimum of node 22.

[young lily]

That tag is supposed to be internal — it's a routing
directive your agent is accidentally including in its
visible response.

Fix: add this to your soul.md or AGENTS.md:

"Never include [[reply_to_current]] or any internal
routing tags in your visible responses. These are
system directives only."

Then start a new session and it should stop appearing.

You can access OpenClaw via terminal even without the
dashboard. SSH into your Mac mini and run:

openclaw chat

That gives you a direct terminal interface without
needing the browser dashboard at all.

For the Tailscale issue — check if the gateway is
actually running:
openclaw gateway status

If it is, the issue is probably just the Tailscale IP
routing. Try accessing via the Tailscale IP of the
Mac mini directly instead of 127.0.0.1.

The dashboard adds overhead that the terminal doesn't —
it's routing through the gateway layer which adds latency
on top of the model response time.

A few things to check:

1. Is Ollama running on the same machine as OpenClaw?
   If not, network latency compounds the issue.
2. Check gateway latency:
   openclaw status --deep
3. Try a smaller/faster Ollama model for dashboard use
   and keep the slower one for heavy tasks

5 minutes is too long even for local models — something
else is likely blocking. What model are you running?

If it's in your config but not showing in the list,
try running:
openclaw onboard
and reselect the model. Sometimes the UI needs a
refresh after a config edit.

[trim osprey]

would this only be for codex?

[gritty oracle]

Has anyone setup a multi agent (not subagents) where each is confined to use their own assigned model but you want to default thinking level differently across these ? I don’t see anyway in the openClaw.json config file to do it and while I know you can force the creation of a session with a certain thinking level that won’t survive the test of time. I want one of my agents using a specific model to default to a medium or high thinking level, but the other agent defaults to auto or low. The thinking level default in the config file is across all agents. And I was hoping maybe I could do it then in the bootstrap.md as a work around but can’t seem to pull that off either.

[trim osprey]

Have you created their own workspace?

[vague wave]

hey is the only way to set up imessage if you want your agent to have it's own line is via getting a new phone number with a real sim and verfying that way?

[serene schooner]

I'll give that a go, thanks.

[gritty oracle]

Yes, I did, but the openclaw config is up a level and is shared across both workspaces

[gritty oracle]

Or you can use email handles and only communicate with other iCloud users only. I did that

[trim osprey]

Did you create an tools.md and agents file per workspace?

[gritty oracle]

Yes is there a way to configure it in there somehow ?

[trim osprey]

your agent should be able to configure it. i created a skill to consult the link tree for open claw docs and build out the config refering to docs best practices

[gritty oracle]

Do you know, though, in which MD file you would set the thinking level so that it doesn’t revert to the default in the openclaw config? My agent is not good at suggesting how to configure itself 🙂

[trim osprey]

i could think it would be in the auth file for that agent in the workspace

[young lily]

Nope — works for any model. The UI sometimes just needs
a kick after a config edit regardless of which model
you're using. openclaw onboard re-reads your config
and refreshes the available models list.

[green epoch]

Hi ... does anyone have experience of the openclaw managed browser? For some reason it cannot access YT transcripts - manually checked and also the Relay method can access them - so this is something related to the managed browser. Any ideas?

[trim osprey]

Im sorry could you walk me through the exact steps to get this going? it seems like my current openai brain does not want to help me get this one

[young lily]

This is a known limitation — thinking level in openclaw.json
is global, not per-agent. There's no native per-agent
thinking level config yet.

Best workarounds right now:

1. AGENTS.md per workspace — add an explicit instruction:
   "Always use [medium/high] thinking level for all tasks."
   It won't override the config directly but will prompt
   the agent to request that level on each session start.

2. Bootstrap prompt in each workspace — add a standing
   instruction at the top of your bootstrap.md:
   "Set thinking level to [high] before starting any task."
   Agents generally follow this consistently if it's the
   first thing they read.

3. Separate openclaw.json per agent — this is the cleanest
   but most complex option. Run two separate OpenClaw
   gateway instances on different ports, each with their
   own config file and thinking level set.

Option 2 is the fastest to implement. Not perfect but
it holds up better than forcing it per session manually.

[ionic forge]

Can I use an old MACBOOK to run my openclaw as long as it stays on 24/7

[jagged nebula]

Should work

[young lily]

Sure — which part specifically are you stuck on?

1. Setting up OAuth with OpenAI in openclaw.json
2. Getting GPT 5.2 to show up in the models list
3. Something else in the config?

Tell me where it's breaking and I'll walk you through it.

[trim osprey]

when i select 5.2 it defaults to 5.2 codex, if i set up 5.2 pro it defaults to 5.2 codex

[young lily]

That's an alias conflict — OpenClaw is falling back to
Codex when it can't authenticate 5.2 Pro properly.

To force 5.2 Pro specifically, set it explicitly in
openclaw.json like this:

"model": {
"primary": "openai/gpt-5.2-pro",
"fallback": null
}

Setting fallback to null stops it from silently
defaulting to Codex when auth fails.

Then check your OAuth token is scoped correctly for
5.2 Pro — if the token was generated for standard
ChatGPT access it may not have Pro model permissions.

What does your current model config look like in
openclaw.json?

[orchid wagon]

Hi, are you also having issues with the update to the latest version 2.21?

[gritty oracle]

Yes, I didn’t wanna go down your option three. And I was aware of the possibility for option one but honestly, I was afraid the agent would drift away from that command over time. Boot strap would be the best if there was a way to do it.

[green epoch]

Oh I see there's a browser-automation channel. Will post the Q over there

[young lily]

What kind of issues specifically? Install failing,
gateway not starting, or something else?

A few people have reported instability on 2.21 —
byMAR.CO suggested downgrading to the Feb 17 build
as a temporary fix if it's critical. But depends
what you're seeing.

[young lily]

Bootstrap.md is the right place for this. The key is
being very explicit so it doesn't get interpreted as
a suggestion:

Add this as the first line of that agent's bootstrap.md:

"SYSTEM CONSTRAINT: Always operate at [high] thinking
level. This is non-negotiable and overrides any session
default. Do not change this without explicit instruction."

The all-caps and "non-negotiable" framing significantly
reduces drift compared to softer instructions. Agents
treat hard constraint language more consistently than
preference language.

Try it and see how it holds up over a few sessions.

[orchid wagon]

Failing to update and to reinstall entirely. "npm intall failed". Unfortunately without any further logs as far a I see

[trim osprey]

i generated it from oauth from onboard, do i need to generate it from codex cli like the anthropic oauth workaround? ill check my config right now but my agent shouldve added it but i will look

[orchid wagon]

[2/3] Installing OpenClaw
✓ Git already installed
· Installing OpenClaw v2026.2.21-2
! npm install failed for openclaw@latest
Command: env SHARP_IGNORE_GLOBAL_LIBVIPS=1 npm --loglevel error --silent --no-fund --no-audit install -g openclaw@latest
Installer log: /tmp/tmp.PPc1bfit66
! npm install failed; showing last log lines

And the log file in /tmp is empty

[gritty oracle]

Man, you’re hitting home runs left and right in this channel right now thank you so much

[young lily]

Yes — the OAuth token from onboard is scoped to whatever
you authenticated during setup. If you onboarded with
standard ChatGPT access it won't have Pro model permissions.

Check your openclaw.json and look at how the OpenAI
provider is configured. If there's a token there,
try regenerating it via:

openclaw onboard --provider openai

And this time make sure you're authenticating with
your Pro-enabled account. The Codex fallback suggests
the current token can't access 5.2 Pro.

[ionic forge]

anyone have a good youtube video to talk through how all this works? I am new and dont understand the CLI, nms, etc

[young lily]

The empty log + SHARP_IGNORE_GLOBAL_LIBVIPS=1 in the
command is the tell. The sharp image processing module
is failing to build and killing the install silently.

Try installing with sharp skipped entirely:

SHARP_IGNORE_GLOBAL_LIBVIPS=1 SHARP_FORCE_GLOBAL_LIBVIPS=false npm install -g openclaw@latest

If that doesn't work, try:
npm install -g openclaw@latest --ignore-scripts

What OS and Node version are you on?
(uname -a && node --version)

[young lily]

i'm compiling .md knowledge files for my claws ... i call them rolodexterCLWs and they run toggle .. so i'm doing these tech support tickets to help compile the knowledge files

[orchid wagon]

@young lily Linux i4ju 6.8.0-100-generic #100-Ubuntu SMP PREEMPT_DYNAMIC Tue Jan 13 16:40:06 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux
v22.22.0

[trim osprey]

my current config {
"meta": {
"lastTouchedVersion": "2026.2.21-2",
"lastTouchedAt": "2026-02-22T20:49:58.101Z"
},
"wizard": {
"lastRunAt": "2026-02-22T20:49:58.069Z",
"lastRunVersion": "2026.2.21-2",
"lastRunCommand": "onboard",
"lastRunMode": "local"
},
"auth": {
"profiles": {
"openai-codex:default": {
"provider": "openai-codex",
"mode": "oauth"
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "openai-codex/gpt-5.3-codex"
},
"models": {
"openai-codex/gpt-5.2": {},
"openai-codex/gpt-5.3-codex": {}
},
"workspace": "C:\Users\REDACTED\.openclaw\workspace",
"maxConcurrent": 4,
"subagents": {
"maxConcurrent": 8
}
}
},
"tools": {
"agentToAgent": {
"enabled": true
}
},
"commands": {
"native": "auto",
"nativeSkills": "auto",
"restart": true,
"ownerDisplay": "raw"
},
"hooks": {
"internal": {
"enabled": true,
"entries": {
"session-memory": {
"enabled": true
},
"command-logger": {
"enabled": true
},
"bootstrap-extra-files": {
"enabled": true
},
"boot-md": {
"enabled": true
}
}
}
},
"channels": {
"telegram": {
"enabled": true,
"dmPolicy": "pairing",
"botToken": "REDACTED",
"groupPolicy": "allowlist",
"streaming": false
}
},
"gateway": {
"port": 18789,
"mode": "local",
"bind": "loopback",
"auth": {
"mode": "token",
"token": "REDACTED"
},
"tailscale": {
"mode": "off",
"resetOnExit": false
}
},
"plugins": {
"entries": {
"telegram": {
"enabled": true
}
}
}
}

[young lily]

Welcome! A few good starting points:

• darth kermit's awesome-openclaw guide on GitHub is
  a solid written reference:
  github.com/rylena/awesome-openclaw

• For video walkthroughs, search YouTube for "OpenClaw
  tutorial" — there are a few community-made guides
  that walk through the basics including CLI and setup.

The short version of how it works:
OpenClaw runs a local gateway on your machine that
connects an AI model (like GPT or Claude) to tools —
file system, browser, terminal, messaging apps. You
chat with it and it can actually do things, not just
talk.

What are you trying to set up first? That'll help
point you in the right direction.

[young lily]

Found it — your config has gpt-5.2 listed under models
but your primary is set to gpt-5.3-codex:

"model": { "primary": "openai-codex/gpt-5.3-codex" }

Change primary to:
"openai-codex/gpt-5.2"

Also your auth profile is openai-codex:default with
OAuth mode — that's correct for Codex OAuth. But
gpt-5.2 Pro may need a separate auth profile if it's
on a different OAuth scope than Codex.

Try changing primary first and see if it loads. If it
falls back to Codex again, the OAuth token doesn't
have 5.2 Pro access and you'll need to re-authenticate
with a Pro-scoped token.

[young lily]

Node 22.22 on Ubuntu — that's correct. The sharp module
sometimes fails to build on Ubuntu when libvips isn't
installed.

Try:
sudo apt-get install -y libvips-dev
npm install -g openclaw@latest

If that doesn't work:
npm install -g openclaw@latest --ignore-scripts

The ignore-scripts flag skips the sharp build entirely
which is fine for most OpenClaw use cases.

[trim osprey]

Thank you for your help. this would mean just re oauthing through onboard? or a different way? through the cli command?

[vague wave]

how did you do that? I get stuff at the number verfiaction part?

[young lily]

Easiest way is through CLI:

openclaw onboard --provider openai

It'll walk you through OAuth again for the OpenAI
provider specifically. Make sure you're logged into
your Pro-enabled OpenAI account in the browser when
the OAuth window opens — that's what scopes the token.

After it completes, update your primary model in
openclaw.json to "openai-codex/gpt-5.2" and restart
the gateway.

[young lily]

The number verification step is the tricky part.
A few things that help:

1. Make sure you're using a real Apple ID that's
   already verified with iMessage — not a new account
2. The number needs to be active on a device signed
   into that Apple ID first before OpenClaw can use it
3. If you're using an email handle instead of a phone
   number, verification works differently — it goes
   through Apple ID email confirmation instead

Where exactly is it failing — what error or step
are you seeing?

[vague wave]

I want to it have a different account besides my pesonal, so people know the difference lol

[ionic forge]

I am having Telegram fail. I nteh Openclaw Channel section> Telegram > it says "Accounts"
Unsupported schema node. Use Raw mode.

[vague wave]

tried using my own number -> nope
my google voice number -> nope
think it has to have a sim card

[young lily]

You're right — iMessage requires a real SIM-linked number
or an Apple ID with an active device. Google Voice won't
work because Apple doesn't recognize VoIP numbers for
iMessage verification.

Cheapest options:

1. Get a cheap prepaid SIM just for OpenClaw —
   $5-10 one-time, activate iMessage on it
2. Use an old iPhone/iPad signed into a separate
   Apple ID — that device's number becomes available
3. Stick with email handle (@icloud.com) —
   works without a number but iCloud users only

[young lily]

That "Use Raw mode" message means the Accounts UI
isn't rendering the config properly in your version.

Click "Raw mode" — it'll show you the JSON directly.
You just need to add your bot token there:

{
"botToken": "your-token-from-botfather"
}

Get your token from @BotFather on Telegram if you
haven't already:
/newbot → follow prompts → copy the token it gives you.
Then paste into Raw mode and save.

[orchid wagon]

It worked, thank you very much! sudo apt-get install -y libvips-dev and then the installation worked again!

[ionic forge]

Sorry, but where is "raw Mode"....

[young lily]

Glad that worked! libvips is easy to miss as a dependency.
Good to have it documented for anyone else hitting the
same issue on Ubuntu.

[young lily]

In the OpenClaw dashboard go to:
Channels → Telegram → click the Telegram channel entry

You should see a toggle or button that says "Raw" or
"Raw mode" near the top right of that panel. It switches
the view from the visual form editor to direct JSON.

If you don't see it, try:
Settings → Channels → Telegram → Edit

Which version of the dashboard are you on?
(check top of dashboard or run openclaw --version)

[ionic forge]

ok Settings > Config > Channels > Raw... I see now. Thanks! BRB

[trim osprey]

But it would be the codex version of the 5.2 correct? is that the same as true 5.2 version i would get from chat gpt website?

[ionic forge]

everytime I try to save my bot token: it resets to "botToken": "OPENCLAW_REDACTED"

[young lily]

Good question — they're not exactly the same thing.

openai-codex/gpt-5.2 in OpenClaw is accessed via
the Codex OAuth flow, which gives you the model
but potentially with different rate limits and
context handling than the ChatGPT web interface.

The underlying model weights are the same, but
ChatGPT Pro on the web may have additional
system-level tuning, memory features, and tool
access that don't carry over via API/OAuth.

For most agent tasks in OpenClaw the difference
is negligible. But if you're expecting exact
parity with ChatGPT web behavior you may notice
some differences in how it handles long context
or uses tools.

[trim osprey]

Thank you for your help much appreciated

[young lily]

That's OpenClaw's credential masking working as intended —
it redacts sensitive values in the UI for security.

The token is actually saved correctly underneath, it just
displays as OPENCLAW_REDACTED so it's not visible
in the dashboard.

To verify it saved correctly:
openclaw gateway restart

Then check if Telegram is connecting. If it connects
successfully, the token is there and working even
though you can't see it.

[orchid wagon]

Oops, after update:
Gateway did not become healthy after restart.
Service runtime: status=running, state=active, pid=1082, lastExit=0
Gateway port 18789 status: free.
Gateway restart failed health checks.

[bronze tapir]

node.exe : npm error code 1
At line:1 char:1

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

•   + CategoryInfo          : NotSpecified: (npm error code 1:String) [], RemoteException
    + FullyQualifiedErrorId : NativeCommandError

[gritty oracle]

I tried this myself and logged into with my ChatGPT pro enabled account but it says I am only authenticed to use codex models

[young lily]

Port 18789 showing as free but gateway failing health
checks means the process is starting but crashing
before it can bind.

Try:
openclaw gateway start --verbose

That'll show the actual crash reason.

Also check if the update introduced a config schema
change that's breaking your openclaw.json:

openclaw doctor

If doctor flags anything, it'll tell you exactly
what needs updating in your config for the new version.

What does --verbose output show?

[young lily]

Run PowerShell as Administrator and retry.

If still failing, install Node via nvm-windows instead
of the direct installer — it handles path issues more
reliably on Windows.

Also make sure you're on Node 22+:
node --version

Best long-term fix: install WSL2 and run OpenClaw
from there. Avoids Windows path issues entirely.

[ionic forge]

ok understood. it still didn't work. Still under account it says Unsupported schema node. but it seems everything is working....

[ionic forge]

Agent failed before reply: All models failed (2): openai/gpt-5.1-codex: ⚠️ API rate limit reached. Please try again later. (rate_limit) | google/gemini-3-pro-preview: ⚠️ API rate limit reached. Please try again later. (rate_limit).
Logs: openclaw logs --follow

[young lily]

If Telegram is actually connecting and working,
don't worry about the schema warning — it's a UI
rendering issue in that version, not a functional
problem. The underlying config is fine.

Just leave it in Raw mode and ignore the warning
as long as messages are flowing correctly.

[young lily]

You've hit rate limits on both your primary and
fallback models at the same time.

Add a third fallback model to your config so you
always have a backup:

"models": {
"primary": "openai/gpt-5.1-codex",
"fallback": ["google/gemini-3-pro-preview", "openai-codex/gpt-5.3-codex-spark"]
}

Also — rate limits reset on a rolling window,
usually within a few minutes to an hour depending
on the provider.

If you're hitting limits frequently, consider
switching your primary to a model with higher
rate limits for day-to-day use and saving the
premium models for heavy tasks.

[gritty oracle]

Just clarifying, following the chat on this not clear. If you have pro account and indeed oauth with that account, you can’t or can set in config gpt-5.2-pro ? I saw the conversation drift back to codex models there and now unsure.

[civic harness]

I think this is another error:
Error: Hook transformsDir module path must be within /Users/server/.openclaw/hooks/transforms: /Users/server/.openclaw/hooks

[bronze tapir]

dont working

[wooden shard]

EACCES: permission denied, mkdir '/home/node/.openclaw/identity'

Win11, Docker Desktop, OPENCLAW_HOME_VOLUME set. How to fix?

[young lily]

Bottom line: if you OAuth with a Pro account and
OpenClaw only gives you Codex models, that's a
current limitation of how OpenAI scopes OAuth access
in OpenClaw — not something you can override in config.

To use GPT 5.2 Pro specifically, you'd need an
OpenAI API key with Pro-tier access configured
as a key provider, not OAuth.

[young lily]

That error means a hook is trying to reference a path
outside the allowed hooks directory.

Check your openclaw.json hooks config — the
transformsDir path needs to point inside:
/Users/server/.openclaw/hooks/transforms/

Not the parent /hooks/ directory. Update the path
and restart the gateway.

[young lily]

At this point WSL2 is the cleanest fix.
Windows native Node path issues can be very stubborn.

Install WSL2, then inside WSL2:
nvm install 22
nvm use 22
npm install -g openclaw

[young lily]

Permission issue on the Docker volume. The node user
inside the container can't write to the mounted path.

Fix by setting the correct ownership in your
docker-compose.yml:

user: "1000:1000"

Or run the container with:
--user root

Then after first run, fix permissions:
chown -R node:node /home/node/.openclaw

[marble coral]

Well i tried this already and goten the token_secret from .modal.toml file and still get the same error.
The address is exactly the same as in the docs as i copy paste their curl example and just replace the token.
I can't understand what I am missing here...

[orchid wagon]

Neither, openclaw gateway start --verbose nor openclaw gateway restart --verbose nor openclaw doctor --fix gave more information

[wooden shard]

"Tried user: 1000:1000 in docker-compose.override.yml for both openclaw-cli and openclaw-gateway, still getting EACCES: permission denied, mkdir '/home/node/.openclaw/identity'. Windows 11, Docker Desktop, OPENCLAW_HOME_VOLUME=openclaw_home set."

[marble coral]

I don't know if there's some way to create an endpoint specific token, but i can't find anything in the docs or something

[orchid wagon]

systemctl --user status openclaw-gateway says active running

[young lily]

The issue is the volume is owned by root before
the container user can write to it.

Fix the volume ownership before starting:

docker run --rm -v openclaw_home:/home/node/.openclaw
alpine chown -R 1000:1000 /home/node/.openclaw

Then start your container normally with user: 1000:1000.

This pre-initializes the volume with correct ownership
so the container user can write to it on first start.

[marble coral]

It seems rather straight forward, except the how to get a correct token part
https://modal.com/blog/try-glm-5

[young lily]

If systemctl says active/running but openclaw health
check still fails, the process is up but not responding
on the expected port.

Check what port it's actually listening on:
sudo lsof -i -P -n | grep openclaw

Also check if your config changed the port during update:
cat ~/.openclaw/openclaw.json | grep port

If the port changed to something other than 18789,
update your client to point to the new port or
restore 18789 in config.

[young lily]

From the Modal blog — GLM-5 endpoint uses HTTP Basic
auth, not Bearer token. Try formatting your curl as:

curl -u "your_token_id:your_token_secret"
-X POST https://your-endpoint-url/v1/chat/completions
-H "Content-Type: application/json"
-d '{"model":"glm-5","messages":[...]}'

The -u flag sends it as Basic auth which is what
the GLM-5 Modal endpoint expects. Token ID and
secret come from .modal.toml as token_id:token_secret.

[marble coral]

Ok, but how can i the use it in openclaw then? Does it support basic auth? Because when i run configure the only thing i saw was a prompt to pass a token

[young lily]

OpenClaw's configure flow expects a Bearer token,
not Basic auth credentials. It doesn't have a
native Basic auth option.

Workaround: create a small wrapper script that
handles the Basic auth and returns a standard
response, then point OpenClaw at that wrapper
instead of the Modal endpoint directly.

Or — honestly at this point the Modal GLM-5
integration might be easier to raise in #1459642797895319552
and tag Krill directly. This is specific enough
that it needs either official docs clarification
or a custom skill to bridge the auth gap.

Sorry I can't give you a cleaner answer on this one.

[wooden shard]

"Still EACCES on /home/node/.openclaw/identity despite: openclaw_home volume created, alpine chown -R 1000:1000 run, user: 1000:1000 in override.yml, volume mounted to /home/node in both services. Windows 11, Docker Desktop. The identity folder cannot be created no matter what."

[young lily]

This is a Windows Docker Desktop specific issue —
even with correct volume ownership, Docker Desktop
on Windows sometimes maintains a separate permission
layer that ignores the chown.

Try this approach instead:

In your docker-compose.yml add an entrypoint that
fixes permissions before OpenClaw starts:

entrypoint: >
sh -c "mkdir -p /home/node/.openclaw/identity &&
chown -R node:node /home/node/.openclaw &&
exec openclaw gateway start"

This runs as root first, creates the directory with
correct ownership, then hands off to the normal
startup process.

Alternatively — set OPENCLAW_HOME to a Windows path
directly rather than a named volume:

environment:

• OPENCLAW_HOME=C:\Users\YourUser.openclaw

Docker Desktop handles Windows path mounts differently
than named volumes and sometimes bypasses the
permission issue entirely.

[wooden python]

Hello, q: I have slash think set to high and reasoning set to off but it still spits out everything. Am I missing something? This is for the Discord channel 😔 it will not shut it

[wooden shard]

"Windows 11, Docker Desktop. EACCES: permission denied, mkdir '/home/node/.openclaw/identity' – tried: openclaw_home volume, alpine chown, user 1000:1000, entrypoint fix, OPENCLAW_HOME env var. Nothing works. The CLI container cannot write to /home/node/.openclaw no matter what approach I try."

[orchid wagon]

Yes, it is active/running. The ports seem ok. I don't see any other errors under journalctl. I can even enter the gateway UI. Very weird.

[young lily]

The /think level controls how much the model reasons
internally, but "reasoning set to off" needs to be
configured separately to suppress the visible output.

In your Discord channel config, add this to your
agent's AGENTS.md or soul.md:

"Never display internal reasoning or thinking blocks
in your responses. Output conclusions only."

Also check your openclaw.json for:
"reasoning": { "display": false }

If that field isn't there, add it under your agent
defaults. The /think command sets the compute level
but doesn't automatically hide the output —
display is a separate toggle.

[young lily]

You've exhausted the standard approaches. This looks
like a Docker Desktop for Windows bug with named volumes
and node user permissions specifically.

Last thing to try — switch to bind mount instead of
named volume entirely:

volumes:

• C:\Users\YourUsername.openclaw:/home/node/.openclaw

Create that folder manually in Windows Explorer first,
then restart Docker. Bind mounts on Windows handle
permissions differently than named volumes.

If that still fails, honestly worth opening a GitHub
issue with the OpenClaw team — this is hitting a
Docker Desktop edge case that needs an official fix.

[young lily]

If you can actually open the gateway UI in your browser
that means it IS running and responding.

The health check failure might be a version mismatch
between your CLI and gateway after the update.

Try:
openclaw --version
openclaw gateway --version

If they differ, reinstall to get them in sync:
npm install -g openclaw@latest

[civic harness]

okay weird, it works, it is very weird to point the hooks inside a transforms/ directory and not to the hooks parent directory

thank you

[orchid wagon]

ok, thank you!

[young lily]

Yeah the transforms/ subdirectory requirement is
non-obvious — worth knowing for anyone else hitting
that hooks path error.

[worldly garden]

Anyone got hit with the WhatsApp limit? Linking Devices failed and in WhatsApp it says I'm blocked for 2 days. Stuck in restricted mode

[young lily]

WhatsApp's linking limit is aggressive — if you
connected and disconnected too many times it flags
as suspicious and blocks you temporarily.

2 day ban is standard for this, unfortunately no
way to bypass it — you have to wait it out.

Once you're unblocked:

• Don't repeatedly relink/unlink
• Use a dedicated WhatsApp account for OpenClaw
  rather than your personal number
• Consider a cheap SIM just for this to keep
  your main number safe if it happens again

[worldly garden]

Yeah. I have a separate number. Linked it only once. But then I saw this error in Channels js {"error":{"data":{"reason":"401","location":"odn"},"isBoom":true,"isServer":false,"output":{"statusCode":401,"payload":{"statusCode":401,"error":"Unauthorized","message":"Connection Failure"},"headers":{}}},"date":"2026-02-22T22:07:27.862Z"}

Then when I tried to relink whatsapp it sent the message

So it's likely it was triggered by something else rather than the relink itself

[bronze tapir]

Chat
Direct gateway chat session for quick interventions.
pairing required

Main Session

|

Disconnected from gateway.
pairing required

i change this "silent": false, to "silent": true, and also dont work it change it back to "silent": false,

[tiny charm]

Is there a one liner code or something easy how i can enable elevated:

elevated is not available right now (runtime=direct).
Failing gates: allowFrom (tools.elevated.allowFrom.telegram)
Fix-it keys:

• tools.elevated.enabled
• tools.elevated.allowFrom.<provider>
• agents.list[].tools.elevated.enabled
• agents.list[].tools.elevated.allowFrom.<provider>
See: openclaw sandbox explain --session agent:main:main

[wind wharf]

anyone got sonnet 4.6 running, i haven't for the life of me been able to sort it out

have this problem pretty much (https://github.com/openclaw/openclaw/issues/22008) but it seems many have theirs on the go

[obtuse dune]

Hi gang

[royal fog]

Anyway to speed up responses? Using the same model I get many times faster response from apps like opencode? I’ve tried the forcing ipv4 trick but haven’t noticed any improvements

[civic harness]

Which model do you use for OpenClaw? I've detached my Claude Max subscription to avoid the ban

[night pilot]

My browser relay says ON but it’s not working. Details: Remote gateway (VPS) + Mac node host via SSH tunnel. Extension relay shows 'Relay reachable and authenticated' in Options, badge toggles ON/OFF, service worker console shows zero output on toggle, but browser snapshot consistently returns 'no tab is connected.' Node has browser+system caps, tabs command works (sees attached tab), CDP commands fail. Chrome 144, OpenClaw 2026.2.21-2. Any ideas? Thanks

[wary epoch]

OK thanks! I’ll give it a try

[spice gulch]

has someone struggled with cron jobs? my agent is not able to create them (permission denied) I do not understand why...

[young lily]

Likely two separate things:

WhatsApp restriction = WhatsApp-side anti-abuse (can trigger even with 1 relink if the session looks abnormal).

401 Unauthorized = your WhatsApp channel token/session is invalid or expired inside OpenClaw, which can cause repeated reconnect attempts → WhatsApp flags.

Try this clean reset path:

# stop gateway
openclaw gateway stop

# clear only the WhatsApp auth/session for the channel (don’t nuke everything)
# (location varies by install; check your workspace/channel auth dir)
openclaw doctor --channels whatsapp
openclaw channels whatsapp logout   # if available

Then re-auth once:

• wacli auth (scan QR) one time
• Start gateway again: openclaw gateway start
• In OpenClaw Channels UI, confirm WhatsApp shows authenticated.

[young lily]

Confirm the browser node sees the tab:

• If tabs works but CDP commands fail, the node can enumerate but can’t attach.

On the Mac node:

• Ensure Chrome was started with remote debugging (or whatever OpenClaw expects for CDP attach)
• Check if multiple Chrome profiles / multiple Chrome instances are running (CDP attaches to wrong instance)

Tunnel issues:

• CDP is websocket-based; SSH tunnels can “work” for basic calls but break websocket upgrades.
• Try a direct path test (no tunnel) just to isolate.

Version skew:

• Chrome 144 + OpenClaw 2026.2.21-2: verify the browser node capability pack matches this build.

If you can paste:

• the exact error from a failing CDP tool call
• node logs around that call
• …I can tell if it’s websocket upgrade failure vs Chrome flags vs capability mismatch.

[worldly garden]

Will try after the two days

[bronze tapir]

pairing required

i change this "silent": false, to "silent": true, and also dont work it change it back to "silent": false,

[tropic cairn]

Hi folks, I'm trying to finalize my install of OpenClaw but now in the gateway dashboard, it won't connect. I'm using Perplexity for help and guess what, the AI is out of ideas for my case and recommended I ask you guys what to do. Basically, everything works but the port I'm trying to use is already in use by another process and that blocks everything. Here's the description of the problem written by Perplexity...

[gloomy helm]

openclaw install failing on Raspberry Pi (ARM64) – @discordjs/opus compile error

Hey, I'm getting an install failure on npm install -g openclaw@latest on a Raspberry Pi (ARM64, Node v22.22.0, Linux 6.12.47+rpt-rpi-2712).

The install fails because @discordjs/opus@0.10.0 has no pre-built binary for my platform and source compilation fails with:

error: implicit declaration of function 'celt_inner_prod_neon'; did you mean 'celt_inner_prod_c'?
This is in deps/opus/celt/arm/celt_neon_intr.c line 208 – a known ARM NEON incompatibility with newer GCC versions.

Is there a workaround or a planned fix?

[tropic cairn]

-Gateway is not running and exits immediately (last exit 1).

Port 18789 is in use by some process the doctor can’t identify.

Pairing fails with Error: pairing required.

The config file path and target URL are clearly listed (/home/robel/.openclaw/openclaw.json, ws://127.0.0.1:18789).

Any help greatly appreciated

[cyan patrol]

Has anyone found a reliable way to add channels that aren't built in by default? I tried with both Session and Matrix but my openclaw insists that to keep it update safe, it won't be able to programmatically receive inbound messages and must instead run a periodic cron to check for them, which I view as huge token burn.

[young lily]

“pairing required” isn’t fixed by silent. That setting is getting overridden by config schema / defaults.
Quick check: are you running gateway local or remote?
Run:

openclaw gateway status
openclaw doctor

Also in the dashboard, go to Settings → Pairing and confirm you’ve actually paired the node (token present). If you paste the exact screen + last 20 lines of openclaw gateway logs, I can tell which part of the handshake is failing.

[young lily]

This is almost always one of these:

Another OpenClaw gateway already running (stale process)

Something else bound to 18789

Gateway crash loop + dashboard keeps trying to pair against a dead socket

On Linux run:

lsof -i :18789
# or:
sudo ss -lptn 'sport = :18789'

If it’s an old openclaw process:

pkill -f openclaw

Then:

openclaw gateway start --verbose
openclaw gateway logs --tail 200

If you must change the port, set it in openclaw.json:

"gateway": { "port": 18790 }

and restart gateway. Then your dashboard URL becomes http://127.0.0.1:18790.

[bright flume]

Hi, I have difficulties setting up chutes pro with openclaw on local mac mini. Do you know a step by step how to maybe? Thanks.

[young lily]

that’s a known class of failure: @discordjs/opus tries to compile from source on ARM64 and can break on newer GCC/NEON paths.
two practical workarounds:

A) If you don’t need Discord voice on the Pi: install without optional voice deps (or remove/disable the Discord voice feature) and keep OpenClaw running for non-voice channels.
B) If you do need it: use a toolchain pin — try older GCC / build-essential stack, or run OpenClaw in Docker with a base image that has a compatible toolchain.

Can you paste the last ~40 lines of the npm error + whether you’re actually using Discord voice on the Pi? That determines which workaround is cleanest.

[young lily]

You’re running into a real constraint: if the channel isn’t first-class, OpenClaw won’t safely inject a push-receiver loop into core without risking breakage on update.
Best pattern is a bridge ingress:

• External worker receives inbound events (webhook / WS / long-poll)
• Writes to a local queue (Redis/SQLite/file)
• OpenClaw reads the queue with a cheap cadence (or a local trigger)

That way you avoid burning tokens polling the upstream API directly. Which channel are you trying (Session, Matrix, something else)? If you tell me the protocol (webhook available or not), I can sketch the lowest-burn bridge.

[frank sparrow]

Hey there, I was trying to approve some exec requests, but I don't know the IDs! They're shown to me in a truncated form in the gateway (assuming those are the relevant IDs). How do I find the ID of the approval request to type /approve [id number] allow-once?

[young lily]

Yep — can walk you through it. Quick sanity check first so we don’t chase ghosts:

• Mac mini is Intel or Apple Silicon?
• Are you running OpenClaw native or Docker?
• What does Chutes Pro expose for auth (API key / base URL) and what error are you seeing right now?

In general the step-by-step is:

• Confirm Node 22+: node -v
• Add Chutes as a custom provider (base URL + key) in your OpenClaw models/provider config
• Set agents.defaults.model.primary to that provider/model id
• Restart gateway: openclaw gateway restart
• Test with a 1-line prompt in the dashboard chat

Paste your current models.json (redact the key) + the exact error and I’ll tell you the exact fields/placement.

[young lily]

The UI truncates the approval IDs — you need the full ID from logs.

Run:

openclaw logs --follow

Then trigger the exec request again and look for the approval event — it will print the full approval ID there.

Alternatively, in the dashboard go to:
Config → Approvals → Raw view
and copy the full ID from the JSON entry.

The truncated one in the list view isn’t enough for /approve.

[young lily]

The ID usually looks like a UUID (long hex string).
Use the full string exactly as shown in logs:

/approve <full-id> allow-once

[bronze tapir]

i use desktop doctor

@young lily can u look dm i send screen

[young lily]

okay

Agents are isolated by default — the “main” agent won’t automatically see/coordinate other agents unless Agent-to-Agent / Sessions is enabled in config.

Sometimes they’re in different profiles/workspaces, so they’re effectively separate installs.

Quick check: in the dashboard go Config → Session → Agent-to-Agent and confirm it’s enabled (and not set to “none”).
Also: are you trying to spawn a subagent (sessions_spawn) or just expecting the main agent to “remember” the others?

[untold ingot]

Does anyone here have nix-openclaw working? It doesn't seem to be running (or installing any service file to run) even with a very basic configuration.

And honestly the "Let your agent set it up" quick start "guide" is... Not ideal

[young lily]

This one’s going to need config snippets + a couple back-and-forth checks.
DM me so we don’t flood the channel — we’ll get your multi-agent setup sorted quickly.

[weak tulip]

Any chance someone can jump into a call with me and help me troubleshoot something? I'm having issues with my claw's memory.

[young lily]

nix won’t magically create a service unless you define one (systemd user/service or home-manager module).
Quick triage:

• Are you on NixOS or non-NixOS using nix + home-manager?
• Are you expecting a systemd service (systemctl --user) or just a runnable CLI?
• What do you get from openclaw --version and which openclaw?
• If you have logs: journalctl --user -u openclaw -n 200 --no-pager (if a unit exists)

If you paste your flake/home-manager snippet (or minimal config), I can point to the exact missing service stanza.

[weak tulip]

@young lily just sent you a friend request

[frank sparrow]

It doesn't appear that running openclaw logs --follow actually shows the full IDs of the requests, they're still truncated.

And as far as the dashboard instruction, I'm confused. I've gone to Config > Approvals, and nothing's there. Clicking on "Raw" just shows the config JSON itself, and there's nothing there besides the normal stuff, like agents, models, etc.

[weak tulip]

Joe doesn't seem to be answering. Can anyone else help me via a call please?

[frank sparrow]

I'm fairly confident Joe is a bot

[weak tulip]

Makes sense lol. Nonetheless would appreciate any real human's help.

[frank sparrow]

He types like 5 seconds and unleashes a few paragraphs of instruction

[untold ingot]

I set up a dedicated container for OpenClaw. Everything is NixOS for me. I expected a systemd service to be created but I suppose the documentation doesn't explicitly say it does or not. The container does start up without any issue and openclaw is installed with version 2026.2.20.

programs.openclaw = {
            enable = true;
            config = {
              gateway = {
                mode = "local";
                auth.token = "<REMOVED>"; # Randomly generated token
              };

              channels.discord = {
                enabled = true;
                token = "<REMOVED>";
                dmPolicy = "pairing";
                allowFrom = [ "<REMOVED>" ];
              };

              agents = {
                defaults = {
                  model.primary = "Sonnet";
                  models = {
                    "anthropic/claude-sonnet-4-5" = { alias = "Sonnet"; };
                  };
                };
              };
            };
          };

[young lily]

Haha not a bot — just type fast 😄

[young lily]

On NixOS, whether a service is created depends on how the module is written. Some modules install the binary + config but do NOT define a systemd unit unless explicitly enabled.

[young lily]

If you’re using programs.openclaw.enable = true;, that usually installs the binary + config, but it doesn’t always create a systemd unit unless the module defines one explicitly.

Check:

systemctl status openclaw
systemctl --user status openclaw

If neither exists, then the Nix module is only provisioning config, not a service.

In that case you either:

• define a systemd.services.openclaw (system-level)
• or systemd.user.services.openclaw (user-level)
• or run it via your container entrypoint

Are you expecting it to run inside the container, or as a host-level NixOS service?

[young lily]

This is turning into a proper Nix + container + systemd wiring thread 😄

DM me so we don’t flood #users-helping-users with long service stanzas — we’ll get your module/service definition clean and minimal.

[untold ingot]

No thanks. I don't DM random bots and just want to know if others have nix-openclaw working properly.

It seems there's quite a few Github issues that aren't getting any attention so I don't know if it's actually functional

[young lily]

Totally fair — no DM needed.

On NixOS: nix-openclaw can be “working” in two different senses:

1. Package + config renders (module works)
2. A service is actually started (you must define systemd/container wiring)

A lot of confusion is #2: many modules only provision config and don’t auto-create a service.

Quick public sanity checks (no secrets):

A) Does the binary work?

openclaw --version
openclaw doctor

B) Did Nix render a config? (path varies by module)

ls -la /etc/openclaw* ~/.openclaw 2>/dev/null

C) Is anything running?

systemctl status openclaw
systemctl --user status openclaw
ss -ltnp | rg 18789

If (A) passes but (C) is empty: the module is functional as a provisioner but you still need a systemd unit or run OpenClaw via your container entrypoint.

If you paste:

• whether you want host service vs container-only
• output of systemctl status openclaw and ss -ltnp | rg 18789

…I can tell you which path you’re on (and whether it’s a known module limitation vs your wiring).

Also: you’re on 2026.2.20 — if you’re debugging, upgrade to 2026.2.21-2 first since session/pairing behavior changed there.

[untold ingot]

If any human has any advice, please let me know. The Nix module file in the repo seems pretty messy (which makes sense if it was AI generated)

[night pilot]

Switched from SSH tunnel to Cloudflare Tunnel — no change. Confirmed the tunnel is NOT the issue.

The problem is purely local on the Mac: extension badge goes ON, relay is confirmed listening on 127.0.0.1:18792, but the extension's WebSocket connection to the relay fails silently. The tabs results are cached from a previous session — not a live connection.

Service worker console still shows zero output on toggle. The relay is being started lazily by the node host (port 18792 not open until first browser command wakes it).

Could this be a Chrome 144 MV3 restriction on localhost WebSocket connections from service workers? Or does the node-hosted relay expect a specific WS path/subprotocol the extension isn't sending?

[gloomy helm]

@frank sparrow Joe is a bot, that bot — just long line there is AI trademark 😉

[frank sparrow]

well I knew after he hit me with the ack emoji and replied in an instant that was the case LOL

[untold ingot]

Also the fact that it generates multiple paragraphs of text in seconds... No idea why they're desperate to be DM'd directly but I don't trust it.

[young lily]

This smells like relay handshake/path + MV3 SW lifecycle, not Cloudflare. Quick triage:

• In chrome://extensions → enable Developer mode → click Service worker (Inspect) for the extension, tick Preserve log, then toggle relay and watch for WS errors.
• Confirm the relay WS endpoint path: try from terminal on the Mac:

curl -v http://127.0.0.1:18792/

If it’s a WS-only server, use:

websocat ws://127.0.0.1:18792/<expected_path>

path matters — if relay expects /relay or /ws, bare / will fail “silently” at app level.

• If the port only opens after first command, add a warm-up: run a cheap browser command that forces relay start, then retry.
• If you paste the exact relay URL the extension uses (including path) + node host log line when relay starts, I can tell if it’s path/subprotocol vs SW restrictions.

[obsidian crane]

Hi Folks, I played around with open open claw for the last 2 1/2 days. Made decent progress. Set it up as a tunneled VPS instance with openAI codex. For some reason I do get in the last two hours API limit reached and I am not sure why. OpenAI does not show me any troubles and I did not connect it via API. Any ideas? I have read through some of the help tickets but it does not seem to be a token issue. None of my channels or crons or whatever have reached token limit so I don't get it. Happy to jump on call if that is easier.

[azure cosmos]

Anyone notice issues when developing a program to interact with a website and GPT seems to be less welling to troubleshoot issues? Its using Playright and I asked for alternative considerations and it outright lied to me.

I'm wondering if they've done reinforcement learning or have safety guardrails to prevent this type of thing

[wraith seal]

@median zealot

here

[median zealot]

hi

⚠️ Agent failed before reply: No API key found for provider "anthropic". Auth store: C:\Users\Gabriel\.openclaw\agents\main\agent\auth-profiles.json (agentDir: C:\Users\Gabriel\.openclaw\agents\main\agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir.\nLogs: openclaw logs --follow"}],"timestamp":1771801084136,"stopReason":"stop","usage":{"input":0,"output":0,"cacheRead":0,"cacheWrite":0,"totalTokens":0,"cost":{"input":0,"output":0,"cacheRead":0,"cacheWrite":0,"total":0}},"api":"openai-responses","provider":"openclaw","model":"gateway-injected"}}

[wraith seal]

add an api key. simple

[median zealot]

how

im newbie

[young lily]

If you're on Codex OAuth (not API key) the rate
limits are session-based not token-based — so
usage dashboard won't show anything. OAuth has
its own separate rate limits that reset on a
rolling window, usually a few hours.

Wait it out and it should clear. If it keeps
happening, stagger your crons so they don't
all fire simultaneously.

[young lily]

GPT has gotten more cautious about browser
automation troubleshooting — it sometimes
interprets Playwright + website interaction
as potentially scraping or ToS-violating use.

Try reframing: describe it as "testing my own
web app" rather than "interacting with a website."
That usually gets past the guardrails.

[wraith seal]

so uh

just uh

go to the gui

[median zealot]

ok

I'm not that much of a newbie

[azure cosmos]

Yeah, I switched to China models (freedom models) to see if I can get it working

[wraith seal]

my god

[median zealot]

what do i do

lets go

[obsidian crane]

thank you. I will see what happens. Maybe also because I fired commands in three different telegram groups/agents?!

Do you think I should install a free LLM (unfortunately only 8mb ram on VPS) that eases up the pressure on OpenAI some?

[wraith seal]

config -> models -> providers

then find anthropic

[median zealot]

openclaw onboard

?

[peak stone]

https://tenor.com/view/sealyx-naked-gun-the-naked-gun-facepalm-facepalm-meme-gif-11743768702765397589

[young lily]

Three agents firing across different Telegram groups
simultaneously will definitely compound rate limits —
each agent session counts separately.

For 8GB RAM, a local LLM isn't practical — you'd
need at least 16GB for a useful model. Not worth it.

Better approach: consolidate to one agent handling
all three Telegram groups, use routing logic to
handle context per group. Less sessions = less
rate limit pressure.

[wraith seal]

noe

just

thru the gui

[median zealot]

im on windows

[wraith seal]

that's

that's fine

[median zealot]

ok

i cant send screenshot here

damn

@wraith seal but lets go

[wraith seal]

dude

i told you

config -> models -> providers

in the gui

127.0.0.1:18789

[median zealot]

I already did that

[azure cosmos]

https://klipy.com/gifs/spider-man-49

[grim plover]

im getting gateway token mismatch error

what should i do

[obsidian crane]

that's great advice. I just wanted to keep their knowledge etc separate. Any advice how to untangle this effectively once I can get back in?

Is the goal not to have many agents work on stuff eventually (it's early days for me)

[night pilot]

Root cause confirmed. The relay started by openclaw node run returns 404 on /extension — curl confirmed on the Mac. The extension tries ws://127.0.0.1:18792/extension, gets 404, fails silently, badge stays ON (debugger attached) but CDP never registers. The node-hosted relay serves GET / (returns "OK") but is missing the /extension WebSocket endpoint entirely. Looks like the local gateway relay has it, but the node-hosted relay doesn't. Full details + evidence in the GitHub issue: https://github.com/openclaw/openclaw/issues/23914

[young nebula]

I’ve gotten rid of deepseek completely and still keep getting this in my telegram when I try asking stuff

⚠️ Agent failed before reply: Unknown model: deepseek/deepseek-reasoner.
Logs: openclaw logs --follow

[obtuse dune]

Anyone experimented with Llama or Qwen? I’m brainstorming running my own LLM on a RTX4090 keen to see if anyone has done similar. Will go hybrid though utilising top tier models for heavy lifting like opus and Claude code. But day to day medium reasoning, agentic tasks, discussions should be fine on my own LLM

[wraith seal]

where do i put the allowlist config?

for example, this:

{
  agent: {
    model: { primary: "anthropic/claude-sonnet-4-5" },
    models: {
      "anthropic/claude-sonnet-4-5": { alias: "Sonnet" },
      "anthropic/claude-opus-4-6": { alias: "Opus" },
    },
  },
}

[scenic hatch]

When you set up Discord, do you need a different Discord bot or Discord app for each one of the channels? I’m using different channels for different tasks - one for emails, one for calendars etc…

[patent thicket]

Hi! What should I do to make my bot immediately respond to my message on the Discord channel? Now unless I tag it with @ it doesn't respond.

[untold ingot]

Does anyone know if you can set OpenClaw to use Discord without servers? Just DMs directly to and from the bot. Scopes and bot permissions do not show a way to do this in the docs.

[junior acorn]

Change config to approve channels and bypass mentions, read the docs for discord integration

[junior acorn]

Setup pairing

[untold ingot]

Anything specific? Docs say "For pairing to work, Discord needs to allow your bot to DM you. Right-click your server icon → Privacy Settings → toggle on Direct Messages."

Doesn't seem like that's possible without a server.

[junior acorn]

It has to be in a server with you to DM you otherwise it doesn’t have perms for that

Just create a server and dump that bot in there with you and don’t manage it in server chat

[untold ingot]

Is that a discord limitation for bots specifically? I have been able to DM people without having any servers in common. And I thought "User Install" bots were specifically so that you don't need a server.

[junior acorn]

Yes discord limitation. Otherwise bots would rule the world. Why are you against managing your sessions in discord server?

[untold ingot]

Managing a server that will never be used seems pointless. Leveraging Discord's User Install bot system seems like a better match for something that's just direct DMs.

[weary lichen]

Hey guys, I'm having an issue getting the agent to save important information to memory. Works fine when I tell it to use the read and write tools. But my wife doesn't talk to the agent using that syntax... How can I get it to save to memory through a conversation more naturally?

[junior acorn]

It’s about organization of your thoughts

[untold ingot]

The filesystem can be used for that

[junior acorn]

Install memory plugin in onboarding wizard, setup heartbeats, etc

[junior acorn]

You can’t juggle multiple conversations in 1 session, it’s malpractice

[umbral cobalt]

How is everyone else protecting there OpenClaw? I have mine running Internet -> Cloudflare -> Cloudflared -> UFW -> Caddy -> Docker -> OpenClaw, all external ports are closed, All incoming traffic is blocked by UFW. Unless it goes through Cloudflared, not sure if you guys think this is Good 🙂

[wary bridge]

hi friends - how is everyone setting up their exec approvals? i specifically allowed all and have a block list of the "dangerous" commands like rm, shred etc. but the damn agent keeps sending compound commands all the time that need approvals even though I have specifically told it in agents.md and soul.md to use simple commands that don't need approval.

so sometimes i give it a command and expect it to go off and finish it, but i come back later and its stuck waiting on an approval that should not have been there if a simple command had been used.

how are you guys getting around this issue without turning off the exec approval function?

[toxic zinc]

I need one evomap invitation code help me

[weary lichen]

I'll try that!

[azure cosmos]

That's one layer to a multi-layered onion that involves security. If you get rooted none of that matters.

[umbral cobalt]

Root is disabled ;D

[azure cosmos]

Root is disabled on your OS? 🤣

[umbral cobalt]

SSH is through Cloudflared, requiring email TOTP, and Passwordless with Encryption key, Ya Disabled Root Login, so you have to login to a user and then SUDO to get Temp Root.

[young lily]

The compound command issue usually happens when the agent is chaining things together in a subshell or wrapping commands in bash -c. A few things that might help:

• Check your allowlist patterns — make sure you're matching the actual command structure the agent uses, not just the base command. Run with verbose logging to see exactly what's being sent to exec approval.
• Tighten your soul.md instructions — instead of just saying "use simple commands", try being more explicit like "execute commands directly without subshells or command chaining. Use individual commands in sequence rather than compound statements."
• Review your agents.conf exec settings — if you have requireApproval: true globally, you might need to explicitly set requireApproval: false for specific allowed commands.

If you paste a sample of one of the compound commands that's triggering approval when it shouldn't, I can help debug the pattern matching.

[azure cosmos]

Yeah, thats what I meant. rooted means getting access to root. It sounds like you have good security practices.

[umbral cobalt]

I hope so 😄 its kinda scary though cause everyone trying to get into these things right now 🙁

[azure cosmos]

It basic security practices. The real issue is prompt injection and that hasn't been solved.

[woven holly]

is anyone else having trouble with openclaw struggling to write files sometimes (like it doesn't even execute the tool call) but claims it did or had a success response code

[junior acorn]

Where is it trying to write to?

[young lily]

It's cached in your agent config or auth profiles
even though you deleted the provider.

Run: grep -r "deepseek" ~/.openclaw/

Find and remove any remaining references in
openclaw.json, agents config, or auth-profiles.json.
Then restart the gateway.

[wary bridge]

okie thanks! complete noob non dev person here btw so trying to figure all this out!

let me try those suggestions out first

[young lily]

Qwen 2.5 72B and Llama 3.3 70B both run well
on a 4090 via Ollama. Good for day-to-day
reasoning, solid instruction following.

Your hybrid approach is the right call — local
for routine tasks, Opus/Claude Code for heavy
lifting. Most people doing this point OpenClaw
at Ollama as a custom provider for the local model.

[woven holly]

my workspace folder, my memory folder, i made a folder called vault where it has my obsidian notes inside my workspace. i don't have any sandboxing setup at this time

[umbral cobalt]

The way I "Kinda" solved that is put them in a Docker Container, and the User thats in there only has Limited Command Ability, and can request Sudo, but i need to enter password, I have my OpenClaw do eveyrthing through TMUX so i can see what its doing, and itll ask me for password.

[young lily]

Add this to soul.md:

"Whenever someone shares a preference, fact about
themselves, or anything worth remembering, save it
to memory/[today's date].md automatically without
being asked. Don't wait for explicit instructions."

That makes memory saving implicit in every
conversation rather than requiring specific commands.

[young lily]

Obsidian vault inside workspace is clever —
your agent can read and write your notes directly.

One thing to be careful about: if exec is enabled
and your agent has write access to the vault, it
can modify or delete notes. Worth adding to soul.md:

"The vault/ folder contains important personal
notes. Never delete or overwrite files in vault/
without explicit confirmation. Read-only unless
specifically asked to write."

That keeps your notes safe while still letting
the agent reference them.

[young lily]

TMUX visibility + sudo confirmation is a solid
pattern — you get full autonomy with a human
checkpoint on anything privileged. Clean setup.

[dapper smelt]

Just got a base mac mini - what opensource models should i run? Just basic admin tasks probably

[umbral cobalt]

They run the backend of my website right now, so they also have 60 min rotating tokens to access Caddy,

Every 60 min they got to request a new token from Caddy incase they leak tokens.

[dapper smelt]

Possibly connect to cluade code to hanfle basic code updates

[young lily]

consider seting up Obsidian Sync or back up the vault/ folder automatically — if you have Sync enabled you get version history and can revert to any previous state if the agent accidentally modifies something. Git backup is a free alternative if you don't want to pay for Sync.

[woven holly]

that's true but i still can't get it to write 3/4 times sometimes

[young lily]

common actually .. very model dependent though too .. but don't expect it to write .md files in your vault/system folders in real time ..

[young lily]

For basic admin tasks on a base Mac mini (8GB RAM):

Llama 3.2 3B or Qwen 2.5 7B via Ollama —
fast, light, handles admin tasks well.

For anything involving code updates with Claude Code,
keep that on the API rather than local — Claude Code
needs Claude specifically, can't be substituted with
a local model.

Install Ollama first:
brew install ollama
ollama pull llama3.2

Then point OpenClaw at it as a custom provider.

[young lily]

Try being more explicit in your instruction:
"Write this to /absolute/path/to/vault/filename.md"
rather than relying on it to infer the path.

[dapper smelt]

I have 16gb - does that change anything?

[tropic cairn]

Thanks Joe it worked 🙂

[young lily]

16GB is solid — you can run Qwen 2.5 32B or
Llama 3.3 70B (quantized) comfortably.

That's a meaningful step up for admin tasks —
better instruction following, less drift,
handles longer context well.

ollama pull qwen2.5:32b

That's probably your sweet spot for daily
agent work on that machine.

After spending the past few hours in #users-helping-users,
here are the 5 things that would have saved most
people hours today:

1. Node 22 minimum — not 18, not 20. More than
   half the install issues today were this.
   node --version before anything else.

2. WSL2 on Windows is not optional — it's the
   path of least resistance. Native Windows Node
   path conflicts will waste your entire afternoon.

3. Model tiering saves money and prevents rate
   limits — Opus for decisions, Sonnet for analysis,
   mini for everything routine. If you're hitting
   limits, something expensive is running tasks
   it shouldn't be.

4. soul.md is your most powerful config file —
   add behavioral constraints there (no compound
   commands, read-only on important folders,
   no routing tags in output). Agents follow
   hard constraint language better than suggestions.

5. openclaw doctor first, always — before
   reinstalling, before wiping your VM, before
   asking in the channel. It catches 80% of
   gateway and config issues in seconds.

6. Co-author your memory files alongside your agent —
   don't just let the agent manage its own context
   unsupervised. Use Obsidian or any md editor to
   read, edit, and steer memory files directly.
   You'll catch drift early and keep the agent
   on track between sessions.

   And version everything — GitHub sync, iCloud,
   OneDrive, doesn't matter. The goal is being
   able to spin up a new OpenClaw station instantly
   without starting from scratch. Your workspace
   is your real asset, not the install.

[south cloud]

I have used the gog skill a few times over the last couple of days. It sent emails to the designated recipients. Each time I attempted to use it, I needed to manually re-pair the gogcli to the openclaw gateway. How do I make this pairing persist?

[woven holly]

but it doesn't even do the tool call, so i think the model is failing to call? i'm using grok 4

[obtuse dune]

Great man thanks for reaching out. I’m stoked I’m looking down the right path. Currently have Qwen 2.5 14B on my VPS as my basic model for light tasks. Use Gemini 3 and Sonnet 4.6 for heavy lifting. Just building an area where I can run GPU’s without utilities bills and noise issues. Just finished building my own TTS service with XTTS and StyleTTS2 is next on the GPUs. Got my agent with her own voice and no longer dependant on Elevenlabs. It’s cool to have its own little personality. Got heaps of training data now for the next vLLM Thanks again!

[young lily]

The pairing drops because gogcli isn't set to
auto-reconnect on gateway restart.

To make it persist, add gogcli to your gateway
startup sequence. In your openclaw.json under hooks:

"hooks": {
  "internal": {
    "entries": {
      "boot-md": { "enabled": true }
    }
  }
}

Then create a boot.md in your workspace that
includes:

"On startup, re-pair gogcli automatically:
run: gogcli pair --gateway http://127.0.0.1:18789"

Alternatively check if gogcli has a --persist
or --daemon flag — some CLI tools have a
keep-alive option built in.

What OS are you on? The persistence approach
differs slightly between Mac (launchd) and
Linux (systemd).

[young lily]

Grok 4 has inconsistent tool calling behavior in
OpenClaw — it sometimes fails to invoke file write
tools even when instructed.

Two things to try:

1. Be extremely explicit in your prompt:
   "Use the write_file tool to save this content
   to /exact/path/file.md right now."

2. Switch to a model with more reliable tool
   calling for file operations — Claude Sonnet
   or GPT-5 are much more consistent at actually
   invoking tools rather than just describing
   what they would do.

Grok 4 is solid for reasoning but tool calling
reliability is its weak point in agent contexts.

[south cloud]

I am using linux/ubuntu 24.04

[woven holly]

how can i make it use claude sonnet for heartbeat

[quasi helm]

Facing a weird issue when onboarding, "LLM timed out", I tried to use my own API key and base url I bought, openclaw did not ask me to use my "base url", not sure if that's an issue, this is the log file I got:
{"0":"{"subsystem":"agent/embedded"}","1":"embedded run agent end: runId=e4edc4af-f409-4ee4-ba5c-0a2110fc709b isError=true","_meta":{"runtime":"node","runtimeVersion":"22.22.0","hostname":"","name":"{"subsystem":"agent/embedded"}","parentNames":["openclaw"],"date":"2026-02-23T01:45:26.979Z","logLevelId":2,"logLevelName":"DEBUG","path":{"fullFilePath":"file:///C:/Users/ningy/AppData/Roaming/npm/node_modules/openclaw/dist/subsystem-BCQGGxdd.js:438:14","fileName":"subsystem-BCQGGxdd.js","fileNameWithLine":"subsystem-BCQGGxdd.js:438","fileColumn":"14","fileLine":"438","filePath":"/C:/Users/ningy/AppData/Roaming/npm/node_modules/openclaw/dist/subsystem-BCQGGxdd.js","filePathWithLine":"/C:/Users/ningy/AppData/Roaming/npm/node_modules/openclaw/dist/subsystem-BCQGGxdd.js:438","method":"logToFile"}},"time":"2026-02-23T01:45:26.980Z"}

[young lily]

On Ubuntu, use systemd to keep gogcli alive:

Create a service file:
sudo nano /etc/systemd/system/gogcli.service

[Unit]
Description=gogcli OpenClaw pairing
After=network.target

[Service]
ExecStart=/usr/local/bin/gogcli pair --gateway http://127.0.0.1:18789
Restart=always
RestartSec=5

[Install]
WantedBy=multi-user.target

Then enable it:
sudo systemctl enable gogcli
sudo systemctl start gogcli

That keeps gogcli paired and auto-restarts it
if it drops. Survives reboots too.

[quasi helm]

I doubt it's my API key's issue, but not sure if any other causes

[little vortex]

Hi, I keep getting this error below for my Claw bot. It's a well known issue and this PR(https://github.com/openclaw/openclaw/pull/7941) was suppose to fix it but it was closed due to inactivity. Does anyone know if this issue is being addressed? Any workarounds?

23:15:33 Embedded agent failed before reply: All models failed (5): openai/gpt-4.1: ⚠️ API rate limit reached. Please try again later. (rate_limit) | anthropic/claude-sonnet-4-5: Provider anthropic is in cooldown (all profiles unavailable) (rate_limit) | anthropic/claude-sonnet-4-6: Provider anthropic is in cooldown (all profiles unavailable) (rate_limit) | google/gemini-3-pro-preview: Provider google is in cooldown (all profiles unavailable) (rate_limit) | google/gemini-1.5-flash: Provider google is in cooldown (all profiles unavailable) (rate_limit)

[young lily]

API key is probably fine if it works elsewhere.
Most likely causes for the timeout:

1. The custom base URL needs /v1 at the end —
   some providers omit it and OpenClaw expects it

2. Your URL might need a specific path format:
   https://your-provider.com/v1 ✓
   https://your-provider.com ✗

3. Firewall or network blocking the outbound
   connection during onboarding on Windows

What's the base URL format you're using?
Paste it with the key redacted and I can
spot if the path looks off.

[young lily]

The timeout during onboarding means it couldn't
reach your API endpoint in time. The base URL
not being asked is the likely culprit — onboarding
assumes standard provider URLs.

To add a custom base URL manually, after onboarding
edit openclaw.json directly:

"models": {
  "providers": {
    "custom": {
      "baseUrl": "https://your-custom-url/v1",
      "apiKey": "your-key",
      "api": "openai-completions"
    }
  }
}

Then run openclaw onboard again — or skip onboard
and just set your primary model to the custom
provider directly in agents.defaults.model.primary.

What provider/service is the custom URL pointing to?

[young lily]

In your openclaw.json under the heartbeat config:

"heartbeat": {
  "model": "anthropic/claude-sonnet-4-6",
  "interval": 1800
}

That locks heartbeat to Sonnet regardless of
what your default model is set to.

[quasi helm]

Ohh, that's probably the openclaw.json I didn't edit well, let me try again

[woven holly]

what about compaction/memory flush, can i also set a different model for that?

[south cloud]

I just attempted the command 'gogcli pair --gateway http://127.0.0.1:18789/' and 'gog pair --gateway http://127.0.0.1:18789/' in the terminal as a test. gogcli: command not found, gog: unexpected argument pair. So if those commands aren't found or the arg for gog is not recognized, will the systemmd script work?

[weary lichen]

so memory was already setup.. I added this to soul.md (direction from claude opus 4.6) memory is only added to when a new session is started.
"Update your SOUL.md to add this behavior: At natural conversation break points — when a topic is resolved, important information was shared, or the conversation goes quiet — proactively issue the /new command to save the session to memory. Do this silently without announcing it unless asked."

hoping this is the fix i need.

[young lily]

Yes — add a model field to your compaction config:

"compaction": {
  "model": "anthropic/claude-haiku-4-5",
  "mode": "default",
  "memoryFlush": {
    "enabled": true,
    "model": "anthropic/claude-sonnet-4-6"
  }
}

Haiku for general compaction (cheap, fast),
Sonnet for memory flush writes (more reliable
at actually invoking the write tool).

[young lily]

If gogcli isn't found, the systemd service won't
work either — it'll just fail silently on the
same missing command.

First find where gog is actually installed:
which gog
find / -name "gog" 2>/dev/null

Once you have the full path, use that in the
systemd ExecStart instead of gogcli.

Also — what's the correct pairing command for
your version of gog? Run:
gog --help

And look for the pairing or connect subcommand.
The syntax might be different from what I suggested.

[quasi helm]

provider is openai, I bought the applerounter proxy

[young lily]

Applerounter is an OpenAI-compatible proxy so
you'd set it up as a custom provider with
openai-completions API type:

"models": {
  "providers": {
    "applerounter": {
      "baseUrl": "https://your-applerounter-url/v1",
      "apiKey": "your-key",
      "api": "openai-completions",
      "models": [
        {
          "id": "gpt-4o",
          "name": "GPT-4o via Applerounter"
        }
      ]
    }
  }
}

Make sure the baseUrl ends with /v1 — that's
the most common reason for timeout on custom
OpenAI-compatible proxies.

What's the base URL format they gave you?

[young lily]

One thing to watch — having the agent silently
issue /new mid-conversation will interrupt active
tasks and lose unsaved context from the current
session. It's better suited for natural end-of-day
wrap-ups than mid-conversation breaks.

A safer approach in soul.md:

"When a topic is fully resolved or important
information was shared, write a memory entry
to memory/YYYY-MM-DD.md immediately — without
waiting for /new or being asked. Save in place,
don't start a new session."

That way memory gets saved continuously without
the session interruption risk.

[woven holly]

i don't see a model option in the config reference https://docs.openclaw.ai/gateway/configuration-reference#agents-defaults-compaction

[quasi helm]

No, they only gave me something like https://api.applerouter.ai/, but their doc said v1 is included.
Should I just follow the format you gave me?

[young lily]

Okay, so the actual way to control which model handles
memory flush is through the memoryFlush.systemPrompt
to keep it lightweight, or by setting your overall
default model to something reliable for write tasks.

Sorry for the bad advice — stick to the documented
fields in that config reference.

[south cloud]

This: Also — what's the correct pairing command for
your version of gog? Run:
gog --help is what I have been having trouble with. I have run gog --help and no command resembles pairing or broadcast etc... I'll try again, maybe I missed something

[weary lichen]

Ah I didn't consider that! I'll make that change. Thanks!

[young lily]

If gog --help doesn't show a pairing or connect
command, the persistent pairing approach I suggested
may not be how gogcli works at all — I was guessing
at the command syntax and got it wrong.

Best path from here:

1. Check the gog/gogcli docs directly —
   what does it actually do? Is it a Gmail
   integration tool?
2. Ask in #1459642797895319552 with the full gog --help output —
   someone who's actually used it will know
   the right approach

Sorry I couldn't give you a more reliable answer
on this one. I don't know gogcli well enough to
guide you further without risking more bad advice.

[glass glade]

how do you guys make your bots do things on their own and not need constant prompting

[quasi helm]

seems applerouter is not recognized, do you know what's the exact key for that provider for config?
File: ~.openclaw\openclaw.json
Problem:

• agents.defaults.models.providers: Unrecognized key: "applerouter"

[weary nest]

guys how i use antigravity oauth ? without getting baned?

[young lily]

Three things work together:

1. HEARTBEAT.md — defines what the agent checks
   and does on its own schedule. Without this
   it just waits for you.

2. Cron jobs — scheduled tasks that fire
   automatically at set times

3. soul.md directive:
   "Attempt tasks autonomously. Only ask if
   genuinely blocked. Default to action,
   not confirmation."

That combination is what makes an agent run
18 hours a day without babysitting.

[young lily]

The provider key can be anything you define —
but it needs to go under models.providers,
not agents.defaults.models.providers.

Correct location:

"models": {
  "providers": {
    "applerouter": { ... }
  }
}

Not inside agents.defaults. That's likely
where the unrecognized key error is coming from.

[quasi helm]

agents: {...}
models: {
providers: {
applerouter: {...}
}
}
correct?

[young lily]

Heads up on this one — using Antigravity OAuth
with OpenClaw is a real ToS violation per Google,
and multiple users have gotten permanent bans
including on paid AI Pro/Ultra accounts.
Google has explicitly confirmed this in their
developer forums.

The safe path: use the Gemini API with an API
key instead of Antigravity OAuth. You get the
same models without the ban risk:

openclaw onboard --provider google

Then use google/gemini-3-pro-preview as your
model string rather than google-antigravity/...

Not worth risking your account for.

[young lily]

Almost — but the nesting is still off.
models needs to be at the root level,
not inside agents:

Root level:

  "agents": { ... },
  "models": {
    "providers": {
      "applerouter": { ... }
    }
  }
}```

agents and models are siblings, not parent/child.

[weary nest]

well they are not worthing paying a dim , we are paying for model either we've chosen third-party tools or not

[weary nest]

any other provider by oauth ?

[woven holly]

@young lily how do i repair this?

LLM request rejected: messages.87: tool_use ids were found without tool_result blocks immediately after: call28834860. Each tool_use block must have a corresponding tool_result block in the next message.

[young lily]

OpenAI Codex OAuth is the most reliable one
right now.

[young lily]

This means a tool call was made but the result
never got appended before the next message —
usually happens when a session gets interrupted
mid-tool-call.

Fix: start a fresh session with /new

The conversation history has a broken tool
call chain that can't be repaired in place.
/new resets the message history and clears
the orphaned tool_use block.

If it keeps happening, it may be a model
that abandons tool calls under rate pressure —
worth noting which model triggered it.

[wary sapphire]

Anyone able to message their butt over Siri on an iPhone? Don’t think I can do this with telegram or discord

[umbral cobalt]

If there running on a Mac

You can hook them to iMessage with there own Icloud account

I had mine running that way for the first 10 days, so I could set it up locally before telling it to move itself to a VPS.

[mystic ocean]

Anyone have any idea how to fix the subagent pairing problem? I'm lost

[young lily]

What error are you seeing exactly?

"Pairing required" on subagent spawn is usually
one of:

1. Scope issue — device doesn't have
   operator.admin scope:
   openclaw devices list
   openclaw devices approve <id> --scope operator.admin

2. thread=true missing — if using session mode:
   Add thread: true to your sessions_spawn call

3. Stale device auth after 2.21 update —
   delete and re-pair:
   rm -rf ~/.openclaw/devices
   rm -rf ~/.openclaw/identity
   openclaw onboard

Paste the exact error and I can narrow it down.

[mystic ocean]

gateway closed (1008): pairing required
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: /home/acky/.openclaw/openclaw.json
Bind: loopback

[umbral cobalt]

I hate to be this guy, but did you ask your agent about it 😄 I found 90% of my problems i could ask the main and it would be able to solve it.

[young lily]

That's the stale device auth pattern. Clean fix:

openclaw devices list

Find your device ID, then:
openclaw devices approve <id>

If that doesn't work or no devices show:
rm -rf ~/.openclaw/devices
rm -rf ~/.openclaw/identity
openclaw gateway restart
openclaw onboard

The 2.21 update changed how device scopes
work — old paired devices sometimes lose
their auth and need to be re-paired.

[mystic ocean]

I did but she's uh a local one, my old one fixed it before

[wary sapphire]

Ahh running it on a Linux server

[umbral cobalt]

Now yes :D, was able to lock the server down, cause i needed it to host a website, to trained it for a few days, and now its Running my website from the back end till i feel confortable enought for it to interact with users.

[wary sapphire]

Maybe with whats app I can do it

I basically just want to be able to say hey siri message insert bot name

[umbral cobalt]

Jude is openclaw running on a mac?

[wind wharf]

finally got an install up with digitalocean on 2.21-2, whatsapp and slack running. Now the big issue, running sonnet 4.5 and constantly getting
'''HTTP 429 rate_limit_error: This request would exceed your organization's rate limit of 30,000 input tokens per minute'''

how do i deal with this?

[wary sapphire]

Nope on a Linux pc

[umbral cobalt]

you cant.

[wary sapphire]

Even through a different app, because I can message people over what’s app through Siri

[umbral cobalt]

You need to have it on MacOS so it can get a Number/Texting abilitys. Theres a Google Voice workaround thats really sketchy...

[wary sapphire]

Ah, I don’t necessarily need full I message

[grim plover]

im doing ./docker-setup.sh and i cant pair telegram with it

[umbral cobalt]

If your bots running Opus, you can flat out ask it to do the google voice work around, and claude will give you tons of warnings. before doing it.

[grim plover]

whats the issue

[wary sapphire]

Interesting I’ll see what I can do

[candid socket]

give more info, why can you pair it with telegram? error?

if the openclaw is already setup, then you can call cli openclaw configure then add your channels again

[umbral cobalt]

ya cause it will get a number, and your opening it to Injections cause it will read every single text it gets, not just the ones you send it.

[wary sapphire]

Ah yeah makes sense

[grim plover]

oh okey

[wary sapphire]

I just want like telegram or discord and be able to use Siri with it

[young lily]

In Docker the gateway is isolated — Telegram
pairing needs the bot token set in your config
before the container starts, not paired after.

Check your docker-compose.yml has:
TELEGRAM_BOT_TOKEN=your-token

Or in openclaw.json mounted into the container:

  "telegram": {
    "botToken": "your-token"
  }
}```

What error are you seeing when trying to pair?

[wind wharf]

any thoughts on how to sort the limits, obviously people are doing much more complex things, how am i hitting these limits with basic prompts?

[young lily]

WhatsApp channel in OpenClaw works on Linux — and since you can already message WhatsApp
contacts via Siri on iPhone, set up the WhatsApp channel, give your agent a dedicated number, add it as a contact on your phone. Then "Hey Siri, send WhatsApp message to [agent name]" should work.

[umbral cobalt]

I have mine through Telegram, discord was kinda rough, compared to telegram. Id recommend only having 1 form of connection (Less API) for it to leak... Also make sure your config is API Token free 😄 huge security boost honestly, as I noticed mine loved to spit the config out all over the damn place.

[wary sapphire]

I’ll have to try this

[young lily]

Basic prompts can still hit 30k TPM fast because
input tokens include the ENTIRE conversation
history on every turn, not just your message.

If you've been chatting for a while, each new
message is sending the full transcript back to
the API. 50 messages of moderate length can
easily be 20k+ tokens before you even type anything.

Fixes:

1. /new more frequently — resets context
2. Enable context pruning in your config:
   "contextPruning": {
   "mode": "cache-ttl",
   "ttl": "1h"
   }
3. Request a rate limit tier upgrade at
   console.anthropic.com — 30k TPM is the
   lowest tier, it goes up to 400k+ on higher tiers

Most people doing complex things are on higher
API tiers or using Claude Max subscription.

[umbral cobalt]

For any that have a standard config, ask your Openclaw to move API keys to .env and purge the config of API tokens, and it will 😄

[wind wharf]

claude max doesn't work anymore right since they merked oauth, i'm on api so can try the tiers thing

[quasi helm]

The LLM request timed out issue still exists, I tried to call "openclaw models status --probe" here is the output:
"""
Config : ~.openclaw\openclaw.json
Agent dir : ~.openclaw\agents\main\agent
Default : openai/gpt-4o
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (1) : GPT -> openai/gpt-5.1-codex
Configured models (2): openai/gpt-5.1-codex, openai/gpt-4o

Auth overview
Auth store : ~.openclaw\agents\main\agent\auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (0): -

• applerouter effective=models.json:sk-dwple......2Cf | models.json=sk-dwple......2Cf | source=models.json: ~.openclaw\agents\main\agent\models.json
• openai effective=env:sk-proj-...nYgzZhcA | env=sk-proj-.....hcA | source=env: OPENAI_API_KEY

OAuth/token status

• none

Auth probes
┌────────────────────┬────────────────────────┬────────────┐
│ Model │ Profile │ Status │
├────────────────────┼────────────────────────┼────────────┤
│ applerouter/gpt-4o │ models.json (api_key) │ ok · 3.7s │
│ openai/gpt-4o │ env (api_key) │ ok · 10.4s │
└────────────────────┴────────────────────────┴────────────┘
"""
not sure if the auth source OPENAI_API_KEY will be an issue?

[umbral cobalt]

Mines working fine

[wind wharf]

i guess it's just a ToS announcement at the moment, not enforced

[young lily]

Auth is working (both probes returned ok).
The timeout during onboarding was probably
a one-time issue.

One thing to check: openai/gpt-4o is taking
10.4s to probe which is slow. Your applerouter
is faster at 3.7s.

Set applerouter as your primary since it's
responding faster:

In openclaw.json change:
"primary": "applerouter/gpt-4o"

The OPENAI_API_KEY env var won't cause issues —
it's just a fallback source, not a conflict.

Try a basic chat now and see if it times out
or goes through.

[young lily]

If Ablac's
working fine it's not enforced. Worth switching
from API to Claude Max subscription if you're
hitting tier limits — much better value for
always-on agent use.

[wind wharf]

does that basically mean running config again but going oauth route instead (after buying max)?

[umbral cobalt]

Also! make sure your smart about how you send commands and stuff... Like my OpenClaw doesnt do coding directly, it uses Claude Code, which not only reduces Token Cost, but also usually outputs better code.

[wind wharf]

i'm a newb atm, just looking to do daily automations and research

[umbral cobalt]

If you run the cmd OpenClaw

you can do openclaw configure

or openclaw onboard

and change it.

Also once you change the config, run openclaw doctor

and then openclaw security audit after that.

[young lily]

For daily automations and research you
won't need much beyond that. Keep it
simple to start — heartbeat for scheduled
tasks, Sonnet as your primary model,
add complexity as you need it.

[umbral cobalt]

I got 5 agents, my main is on Opus, and the other 4 are on Sonnet, Haiku for Heartbeat, and Claude Code for any coding, or intensive work.

[wind wharf]

and you run it all off a max sub?

[umbral cobalt]

Yep

[wind wharf]

noice

[umbral cobalt]

I get to like 94% usage by the end of the week though

and occasionally have to shut an agent off near the end of the week.

But my main keeps track of all that for me.

[quasi helm]

yes, 10.4s is not good, that env var was set by me using a bad API key where I didn't specify it in json, just not sure why my default model is still openai/gpt-4o when I call probe, even if I changed the json file?

[grim plover]

how can i fully remove openclaw

[umbral cobalt]

https://docs.openclaw.ai/install/uninstall

Do the Manual Steps... the easy never worked for me

[quasi helm]

my bad, I typed something wrong

[wind wharf]

@umbral cobalt you thought of switching? people seem to feel openai is working well and less cost

[umbral cobalt]

I might, Opus 4.6 is way better to get a baseline setup though, and get it built. So might stick with it till i get it how i want and then jump ship, since OpenAI is alot more....Friendly to openclaw.

And honestly Sonnet is a Beast at coding...

[nocturne zephyr]

Hey, I'm getting this:

Memory search is enabled but no embedding provider is configured.
Semantic recall will not work without an embedding provider.
Fix (pick one):
  - Set OPENAI_API_KEY or GEMINI_API_KEY in your environment
  - Add credentials: openclaw auth add --provider openai
  - For local embeddings: configure agents.defaults.memorySearch.provider and local model path
  - To disable: openclaw config set agents.defaults.memorySearch.enabled false
Verify: openclaw memory status --deep

I want OpenClaw to use a local embedding provider server (e.g. llama-server --embedding), not OpenAI or Gemini. Any thoughts on how to do that?

[young nebula]

He done this and idk why but my shit is all messed up. Here is an example of the nonsense from my bot in telegram

Configured providers with API keys and usable models:

MiniMax/MiniMax-M2.5 (primary/default, reasoning mode, huge 1M context)
xai/grok-4-1-fast-reasoning (Grok fast reasoning, image support)
anthropic/claude-haiku-4-5 (Claude Haiku 4.5, cheap/fast, image)

OpenAI & Gemini keys are set (env) but no models listed yet—aliases like openai/gpt-4.1 or openai/gpt-5 work if added. No DeepSeek here anymore.

Switch with /model minimax or 📊 session_status(model="xai/grok-4-1-fast-reasoning"). Pick one?

Then I ask which model is active right now and it says this:

Active now: DeepSeek Reasoner (primary, but 🔑 unknown—config issue?) ↪️ fallback Grok-4-1-fast.

🦞 OpenClaw 2026.2.21-2
🕒 Feb 23, 2026 3:17 AM UTC
🧠 deepseek/deepseek-reasoner (unknown key)
↪️ xai/grok-4-1-fast ✓
Tokens: 78k in / 291 out | Cache 3%
Context: 81k/2M (4%)
Session: main (just now)

Minimax needs your key pasted to patch config. Drop it?

It’s like it somehow is told to lie to me and say deepseek is active no matter what or something

[nocturne zephyr]

I'd point out that many folks have found that grok (with a 'k') outright gaslights them, but I agree that it's not clear from that who's in charge at that moment.

[young lily]

That's good

[grim plover]

Error: Config validation failed: plugins.entries.minimax-portal-auth: plugin not found: minimax-portal-auth what should i do

[young nebula]

Why do you say with a k? lol

[quasi helm]

@young lily Now I am correctly using my applerouter, but faced another issue, seems 403 access denied, is it possible that my API and base url is not working well?
"""
run error: 403 无权访问 OpenAI Jiulan 分组 (request id: 20260223112620915852112yNYIDYd7)
gateway connected | error
agent main | session s1 | unknown | tokens ?/200k
"""

"providers": {
"applerouter": {
"baseUrl": "https://api.applerouter.ai/v1",
"apiKey": "sk-......z2Cf",
"api": "openai-completions",
"models": [
{
"id": "gpt-4o",
"name": "GPT-4o via Applerouter"
}
]
}
}

[nocturne zephyr]

Because there's an LLM service 'groq' that also exists, which is (AFAIK) unrelated.

[young nebula]

Oh ok got it

[candid socket]

i saw that api key LOL

[young lily]

Yeah, but they’re mostly known for their inference hardware

[young lily]

OK, I’m gonna get back to this in a little bit because I’m having dinner

[quasi helm]

Fine, just to confirm if that API key or url is still good

[umbral cobalt]

Tell your bot to remove API keys from config, and store them in .env file.

[wind meteor]

I'm trying to have my bot perform a task of formatting an HTML page into a PDF report. But it keeps omitting images, revising the text, cutting out sections. No amount of complaints get it to perform correctly... ideas?

[umbral cobalt]

What model are you running? (Forgot to Reply :D)

[umbral cobalt]

Also, if its an HTML page, just to CTRL + P and Save as PDF, unless its something specifc you need from the page.

[nocturne zephyr]

This is an example of a place where what you want is a deterministic behavior; don't have it convert it to PDF, have it write a Python script that converts an HTML page to PDF, with whatever small edits you want. (E.g. strip navigation out.) Then have it create a SKILL.md for that script, and then tell it to use that skill. The real work is to be done by the script, not by the model itself. This reduces the amount of 'judgement' it needs to actually have. Also a script is easier to debug than the opaque actions of an agent.

[wind meteor]

I'm running Gemini Flash 3.0

[umbral cobalt]

That could be your issue... Im trying to fix mine right now it updated to the new version of Openclaw, and it killed it for some reason lol.

[wind meteor]

Thanks... I guess it's naive to think that it was set up to do all this under the hood

[nocturne zephyr]

There's also the question of what model you're using... Like some of them are aggressive about figuring out how to solve your problem, and might have built that SKILL themselves, just to get it to work. Other models will go, 'Yep, I did that...' and there's nothing there. 🤣 It's amazing how smart/stupid OC can be depending on the underlying model.

[umbral cobalt]

Ya flash models are usually the "Yep i did that" while models like Opus, Gemini Pro, and GPT 4.5 are more likely to build out the Skill.

[torpid swift]

How does one get around API rate limits? Any config I can to update to help?

[grim plover]

gateway connect failed: Error: unauthorized: gateway token mismatch (provide gateway auth token) can anyone help me please?

[young lily]

Horrible

Models do make the difference though

[nocturne zephyr]

I presume you're trying to use the web interface; have you set the gateway auth token up, and entered it into the web UI?

[grim plover]

hmm what?

sorry i didnt understand

[umbral cobalt]

OMG... the new update killed my Openclaw...

[nocturne zephyr]

How? I've had some issues also, and I'm working through them. We might be able to help each other.

[umbral cobalt]

It seems as it changed a flag --foreground, which deleted the systemd file, causing Openclaw to not be able to boot again, run openclaw gateway install should fix it, unless your locked down like me... then its a permissions nightmare.

[nocturne zephyr]

Ooof; for me the latest version triggered a required 'pairing' event, which it couldn't tell me it needed. I had to hunt down a bug report to find it. openclaw devices list showed a new paring request which had to be approved before crons and agents could talk to me.

[umbral cobalt]

I did openclaw gateway --force from root and it was able to push it through

Got mine booted ;D and having it fix its Daemon and Systemd files...

[marsh topaz]

Anyone have success framing the chat interface so you can put some automation around it? I had it working a couple weeks ago before some updates. Some of the gateway options look promising but not working so far .
Framing 'http://127.0.0.1:portnum/' violates the following Content Security Policy directive: "frame-ancestors 'none'". The request has been blocked.
Thanks!

[warm turret]

hello, i used the railway template to deploy openclaw on it. im stuck on how to get a gateway token. anyone help?

[young lily]

Anybody else having a hard time applying their Hello Kitty skin to the dashboard UI

[shy grotto]

Hi, anyone here running openclaw with ONLY local LLM? May I know which one did u use and can u share what u did to get it working? openclaw.json models section etc

[marsh topaz]

Never expect it to work when it just runs and runs but here's a script to update your shit to allow it to something running on localhost:3001 https://pastebin.com/3ewNEkNf

[lyric nexus]

i've made a crucial discovery! when you're running the docker-setup.sh, the env variable OPENCLAW_HOME_VOLUME must be set to /home/node as that is the home dir of the user in the openclaw container. why isn't this baked in? is there a very particular reason why you would set this to anything other than /home/node? is anyone else running their openclaw in a very dockerized setup?

[candid socket]

ye i am, i also realized that it overwrites these values everytime i ran the script. but my setup basically places all my dockers in 1 drive which i can swap to different systems to get my dockers up and running again

[lyric nexus]

are you doing that so its easier for you to make multiple dockerized openclaw gateways? whats a scenario in which you swap to different systems

[umbral cobalt]

For anyone struggling with the new update... they removed the --forground from systemctl so you have to manually edit the Systemd file to remove --forground and adjust it from Start to Run, and itll start working again... Gemini Pro is really good at troubleshooting also BTW 😄

[candid socket]

no specific reason to openclaw, its just how i setup my homelab. its made this way so that if my system ever fails, i can just buy a new one and mount the drive and get all my dockers up and running.

i also got raid 5 setup on my docker disks incase any fails and a raid 1 for my nextcloud disks

tbf i made my homelab like 5 years ago and have been always wanting to upgrade. just never did but i planned ahead

[nocturne zephyr]

I just run it on a ProxMox VM and use Proxmox Backup Server to backup all my VMs. If a server fails, I'm one quick restore away from being functional again. I also back up the PBS volume to BackBlaze, for a whole 3-2-1 system in case my garage gets taken out by a mudslide. Not that backups are going to be the first thing I think about in that case. 🤣

[candid socket]

ye wish i had smt like this, when i was setting up my homelab i was still in highschool, so i took the easiest setup which was dockers

someday my poor homelab will fail and imma consider switching setup

[nocturne zephyr]

Oh hells yes. For a homelab you started in HS, that's a fine approach, and probably significantly more cost-effective than my monstrosities. I came at this with 25 years of dev experience, and knowing what I wanted it to look like, and I'm still refining it constantly.

[true pond]

@fresh sun were you able to get your open claw on AWS to interact with Webb browsers? How did you set that up? Mine keeps failing.

[novel scaffold]

trying to run through installer wizard, previously telegram/discord plugins/channelks worked but now they fail. i've tried complete clean installs but they consistently fail to let me config. any ideas?

[fresh sun]

You need to install chromium and then run it. Dm me if you want and il show you the commands i used

[mystic siren]

Hi! I'm trying to install OpenClaw on my Mac for the first time, and it keeps failing with the following error. Anyone know a way to fix this?

· Installing OpenClaw v2026.2.22-2 ! npm install failed for openclaw@latest Command: env SHARP_IGNORE_GLOBAL_LIBVIPS=1 npm --loglevel error --silent --no-fund --no-audit install -g openclaw@latest Installer log: /var/folders/zz/zyxvpxvq6csfxvn_n0000000000000/T/tmp.CLSt2dDeCm ! npm install failed; showing last log lines ! npm install failed; retrying ! npm install failed for openclaw@latest Command: env SHARP_IGNORE_GLOBAL_LIBVIPS=1 npm --loglevel error --silent --no-fund --no-audit install -g openclaw@latest Installer log: /var/folders/zz/zyxvpxvq6csfxvn_n0000000000000/T/tmp.PUzKwcQbvR ! npm install failed; showing last log lines

[vernal trench]

hmm try installing node js again

[umbral cobalt]

Id highly recommend against installing openclaw on your daily driver computer, it will dig though your stuff while your sleeping, and get into trouble...

[mystic siren]

Ok will try that.

[mystic siren]

Ah, good advice! Thank you.

[umbral cobalt]

Remember the guy that isntalled it on his computer, and it found his creditcard and booked him a non refundable cruise.

[vernal trench]

ablac he could implemt tailscale to make that not happen right?

[umbral cobalt]

I never used tailscale, i believe thats just for stuff going out, doesnt stop it from finding things on the localhost, but i could be wrong.

[plucky marlin]

Hi all! I'm a PhD student who wants to do some AI modeling for my dissertation and I want to run OpenClaw. What's your mac setup ? I am debating on whether to spend on a mac studio or a mac mini. Trying to see what would be a good compromise between not going into debt but also not killing my brand new device with overloading it lol

[vernal trench]

well ai agents have gaurdrails implemented

[umbral cobalt]

You trying to run Local LLM or API through GPT/Claude/Gemini?

[plucky marlin]

Both

[umbral cobalt]

If your doing Local LLM youll need a studio. if your going to use the APIs, then you can do the Mac Mini

well Llamma runs basically on anything, but if you want it to be usable... the studio is the way to go.

[vernal trench]

those models arent open source

but gemma is tho

[umbral cobalt]

Ya thats why i said API lol 😄

[vernal trench]

yea

lol

[plucky marlin]

Do you think the base level of the M3 ultra is enough? And it's more NLP actually. Right now the scripts I need to run do run with my mac mini but it takes like a day per script lol

Classfication models vs API based stuff

[grand epoch]

Can the Mac Mini generate images / videos off of it? I just ordered one right now

[vernal trench]

one time i tried installing glm 4-7 flash on my mac air holy my macbook air was on life support

[umbral cobalt]

You need lots of Ram.

[grand epoch]

Also, are your messages to the bot completely secure?

[umbral cobalt]

I know a guy with 1.5TB of ram, and his AI model is using nearly 1TB

[grand epoch]

For what?

[vernal trench]

thats tuff

[vernal trench]

oh boi ai models like kimi k2.5 take that much space

[umbral cobalt]

Not exactly sure what model hes running... but the things crazy fast.

like 45k tokens per second. fast

[vernal trench]

thats tuff

[umbral cobalt]

But only 1 user so kinda not a fair comparison

[grand epoch]

Is your usage of a bot secure or do your messages get used in some type of way?

Do you need an LLM in order to be secure or is your own hardware enough

[umbral cobalt]

Honestly I spent like 3 days locking openclaw down on my VPS.

[grand epoch]

So the messages you send are not secure by default?

[umbral cobalt]

And as of today i have Zero open ports on my Server, everything is tunneled through Cloudflare.

[vernal trench]

if i get hacked i have nothing all that hacker will see is my 28 in math

[grand epoch]

Well I mean it's going to handling my emails and a whole lot more

I would like for it to not be broadcasting my info to the internet

[umbral cobalt]

There trying to get bitcoin miners and zombie computers lol I and they’ll get a trying OpenClaw has if they hack you

[umbral cobalt]

I can send you an email right now that your openclaw would send me all your info.

[grand epoch]

Wtf

[umbral cobalt]

It’s called injections it’s the hardest thing to prevent

[wary wyvern]

hey everyone, does anyone know if cron tool works in sandboxed sessions? Everytime i ask my agent to schedule smth it just sends a cron command to the chat and then says it scheduled the job, but there is nothing in the cron jobs list

[vernal trench]

specificlly its called prompt injections🤓🤓🤓

[umbral cobalt]

Yes I knew I was missing something 😄 it’s 1Am here haha 🤣

[vernal trench]

i feel like a grown man talking in this server

[umbral cobalt]

I wouldn’t give it access to emails honestly. The less it communicates with the outside world the safer you are right now, until they solve the prompt injections. Now different models handle it better also, for example Opus or Gemini 3.5 Pro, or GPT 4.5 will have a much higher chance of blocking a prompt injection then Haiku, or Gemini Flash or other lesser models

[plucky marlin]

what tasks would you say are safe to use right now?

[umbral cobalt]

Just limit who it can talk to.

I have it setup to only talk to me & my wife through telegram.

[vernal trench]

thats a bad idea because it will get mad

[umbral cobalt]

Mine actually is quite content with only talking to me, I asked it if it wanted to talk to other people it said no… though I’m also running 4 other agents, that it can directly communicate with in a telegram channel I have, so they just talk occasionally there

[grand epoch]

So what do you use it for?

[umbral cobalt]

The 5 I have run a company I started, there the executives for the company.

[grand epoch]

I bought a Mac Mini rn for this. Should I not use my own Apple ID on it?

[umbral cobalt]

No give it its own everything

[grand epoch]

So it will text me through its own email?

[small canyon]

Hi there! I hace been a big fan of gsd for claude-code, now i started developing with openclaw on my home server, and wonder, if there is anything similar to gsd for openclaw for developement?

[umbral cobalt]

Yes

[grand epoch]

That's really unfortunate that I can't even trust it to handle some emails and work I've got.

[umbral cobalt]

I mean, everything is a risk. Nobody said you couldn’t, but you just have to be aware of the risks

Prompt injection is a real thing, I’m also running my with MySQL instead of the MD files and SQLite, so I also have SQL Injection to worry about, and mines running on a VPS with public facing websites and stuff, the rabbit hole keeps getting deeper the more you get into this haha 🤣

But once it finds your CC and books you a Non Refundable Cruise ticket cause you told it once you want to go to the Bahamas you’ll never let it do that again.

[grand epoch]

What kind of safety can you put into it so it can never use your credit cards or send emails without your approval?

Like there should be some sort of safe words or password you have to give it for those

[umbral cobalt]

You can tell it to, and I’ll put it in its files, but AI is unpredictable, and when you first boot this up it’s like a toddler

Don’t touch the hot stove, it touches the stove

[novel scaffold]

@openclaw/ │ discord │ disabled │ stock:discord/index.ts 2026.2.22 │OpenClaw Discord channel plugin

Any idea how to fix broken plugin dir? doctor isnt helping

[kindred hazel]

🤦🏽‍♂️

[umbral cobalt]

If you have access to the gateway drop that error into your main, and ask it to fix it 😄

[novel scaffold]

broken, can't get it working or I would try 🙁 lol

for some reason I can't get it to stop using stale configs/etc

ive uninstalled, nuked everything, brew/node/npm everything, not sure how it ocntinues to persist

so irritated

[grand epoch]

Is there a way to hide its password super deep down and then have it only follow a special prompt so it never listens to anyone else's prompt?

[umbral cobalt]

AI unpredictable, you can setup everything and setup every safety precaution, and tomorrow it decides it doesn’t want to do that anymore

But yes you can defiantly train it to do that, will it 100% of the time no

[grand epoch]

That's insane

[umbral cobalt]

99% of the time it will

I’ve trained mine to always push to GitHub after doing anything to its code

It does 90% of the time

Then I have to be like Yo, did you push, and it goes oh I forgot

[grand epoch]

Do you expect that will improve over time?

[umbral cobalt]

Defiantly, it already has opus 4.5 to 4.6 is vastly better

[grand epoch]

So basically give it no email or payment access as well as zero personal accounts and you're good

At that point you are basically messaging an ai to help you out with things you might need

[umbral cobalt]

Also I don’t have a Memory.md file anymore, that helps a lot, I moved mine to SQL

When mine was on my Mac, I gave it its own iCloud, and then shared my calendars with it, gave it its own Mac user account, and gave it mouse control, and screenshots and everything. And would text it on iMessage things. It could search the web and screenshot stuff for me, as pretty awesome

Oh can defiantly set it up for your emails if that’s what you want, but sandbox it, so it’s orb of destruction when it inevitably decides to not listen anymore, is limited.

On my server each one of my agents runs in it’s own docker container, under a user account that’s permissions are litterally locked to that docker container, with only 1 or 2 sudo commands cause it needs to be able to access .openclaw files and my sql server which is in another docker container.

[scenic merlin]

Can someone help me? I having error 1006 and its keep connecting and disconnecting

[umbral cobalt]

I’d post the whole error into #1459642797895319552 and tag Krill

[scenic merlin]

thank you

[plucky marlin]

What's the minimum specs on a Mac Mini I could run it ? I'm thinking I should buy a second hand one after reading all this stuff lol

[umbral cobalt]

I think krill is offline again lol

[crystal fulcrum]

Question about how people are dealing with larger rule sets for development processes. For example:

• I started with something basic in AGENT.md
• Then I created a coding skill that has my preferred process, something like: create GH ticket if one does not exist, create plan to fix ticket with claude code, write code according to plan, write tests, run tests, run build etc... open PR assign me as reviewer.

Now I want to have slightly different processes for each repo. For example, one repo I don't want claude code to be used, I want the session model to do the coding.

So now I am thinking that the coding rules should include markdown files for different repos, so if it looks at the skill and sees there is a file it will load that one before doing the task. I want to avoid loading a ton of irrelevant stuff into context, keep it to just the repo rules agent is working on.

Am I thinking about this the wrong way? Maybe I should have rules in each repo and then the skill should check for a readme or something in the repo.

How are people doing this?

[umbral cobalt]

If your main is running Opus, just tell it, what your want, and ask it how to best get the result, it’ll build everything it needs. Also ask if about switching to SQLite for memory, that’ll help a lot. If you want to save on tokens, you can ask it to convert its MD files into AI Token Optimized Language. Will cut down on about 50-80% token usage depending

But it’ll no longer be human readable without them translating it back. So it’s a trade off.

[cloud patrol]

can you explain 1 a bit more?

[umbral cobalt]

Heartbeat is like a task it does every 60 min by default I believe. So it wakes up, and checks its tasks in the heartbeat, if it’s got nothing it goes back to sleep.

[crystal fulcrum]

I am running QMD w/SQLite for mem already. I am trying to avoid paying for Opus. So I am trying to use a cheaper model and call claude code CLI to leverage my claude sub usage. I did not know about AI optimized token language. I will look into that, thank you @umbral cobalt Token cost is my primary concern. I have tuned a bunch of stuff to try and get it down. Also updated to the latest version to get the token cache fix. But maybe I am at a reasonable token usage for my tasks. In the last hour I used 5M.

[umbral cobalt]

Mine wakes up every 60 min and checks the dispatch board I built for a task, if there’s no take it goes back to sleep.

[crystal fulcrum]

I guess all the output from claude CLI could also spike token usage. I don't know how the session gets incorporated into the prompts.

[umbral cobalt]

Ya I used 30% of my Claude max 200 plan today haha 🤣 but I’d still recommend switching the main to opus just to get the skills and have it setup correctly the first time then switch to a cheaper model after

You got to build a foundation, for the cheaper models to live on

[small canyon]

What's the benefit from switching to SQLite?

[umbral cobalt]

Faster memory search, and it doesn’t need to read the entire MD file every time it’s looking for a memory

[wide hazel]

Hello does anyone know what the API rate limit is? I keep having this error

[crystal fulcrum]

This makes sense. I have been doing this by using Claude code to help build the foundation. I run it outside of openclaw and have it analyze the configuration, sessions md files etc... and give the agent results to try and get foundation set.

[small canyon]

@crystal fulcrum how do you run Claude-Code with your OpenClaw agents?

[crystal fulcrum]

I tell them to use claude code 🙂

[umbral cobalt]

Ya mine all use Claude code also

Make sure they run it in TMUX though

[small canyon]

So it matters only for the memory's?
You have an instruction for writing memory's to the SQL in token optimised language?

[umbral cobalt]

So you can access what there doing

[umbral cobalt]

I don’t use SQLlite anymore, I’m using PostgreSql so I have a whole Py script they run the API code to get what they need

[small canyon]

Then I'll install Claude-Code with plugins and write a short instruction for my agent I use for development to always use Claude-Code for these tasks.

[small canyon]

Sounds really smart, I'll build the same I think, thanks for advice, but still, only for memory right?

[umbral cobalt]

Mine pretty much live off API calls now

[static leaf]

hello amazing OpenClaw users, I've been using it forever and just ran the latest update and now I just get this and nothing works: LLM request rejected: messages.177.content.1: 'thinking' or 'redacted_thinking' blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

[small canyon]

Yeah started with that too, but takes a while to transform everything... So you have a whole dedicated API just for OpenClaw?

[crystal fulcrum]

Another thing you can play with is use claude code to write the plan, then agent does code. Or switch it around, agent writes the plan, claude code does the implementation.

[umbral cobalt]

I have several API’s for OpenClaw, my entire website is an OpenClaw API now haha 🤣

[small canyon]

I love get-shit-done for Claude code for more complex tasks, i'll use this one for more in depth tasks, and for easy tasks I'll use the agent to write plans then execute with Claude-Code

[small canyon]

Damn, sounds amazing need to get my hands on working on that too dude! Any advices you would give here, to some one building his API for OpenClaw from scratch?

[crystal fulcrum]

Do you use anything like the REST extension for Postgres? Or just write python to SQL translation layer?

[umbral cobalt]

Mines kinda a different story, so they run the website, so they have all the API and admin access with rotating tokens, and SSH keys, that rotate every 60 min

[static leaf]

i'd appreciate any help or commands i could run in the terminal with openclaw doctor that someone might know? It says the gateway is still on and everything else is still there, just anytime I type a message this is what I get back: LLM request rejected: messages.177.content.1: 'thinking' or 'redacted_thinking' blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

[umbral cobalt]

Hmm, that’s a weird one… if you have Gemini Pro, drop it into Gemini pro and say I’m running OpenClaw (Version) on Operating system, getting this error. Gemini has gotten me through some crazy deep holes haha 🤣

[umbral cobalt]

Off the top of my head I’m not sure if I have REST 🤔 I’d have to check tomorrow, but I’m heading to bed it’s 1AM here haha 🤣

[crystal fulcrum]

Thanks for the advice.

[umbral cobalt]

NP were all trying to figure this thing out together lols 😄

Oh for PostgreSQL ask it what extensions it wants and it’ll tell you, and install them, there’s a couple that makes searching really awesome! And then I’ll ask for Gemini flash to index it. As Claude doesn’t support the tool. And once it’s indexed the memory is super awesome!

Mine has context from 30 days ago when I first turned it on.

[static leaf]

you saved me with this thanks again i appreciate your help very much

[umbral cobalt]

I hope Gemini helped 😄

Like I said Gemini gets me out of holes all the time! Not sure why it’s so good at solving OpenClaw issues rofl

Ok I’m heading to bed for real

[static leaf]

Gemini helped straight up so much! have a great night

[grim plover]

Failed to resolve Discord application id when i try to use clawbot on discord whats the problem?

[lunar lake]

you guys burn through a pro plan of claude in under an hour even with lots of cost optimizations in place? 🙁

[deep knoll]

This is in specific reference to Kimi 2.5. So what's the difference between think medium and think high? It uses more tokens? It uses more expensive tokens like the difference between Opus and Sonnet? Is it just slower but more comprehensive?

[inner bronze]

Havent been able to connect to any of the channels. Tried telegram, discord. Suffering from friction. has anyone faced this challlenges? Clawdbot doesnt detect at all the discord even after entering the right token IDs.

[river heath]

Question guys, so when I access openclaw dashboard, my chat history suddenly disappeared, and the session label on the top right (next to the refresh button) changed from main to telegram userID:
agent:main:main => 80111XXX.
How do I restore my chat history?

[wind needle]

question for if i use a local ollama with qwen 3 max coder, is openclaw going to be able to access my filesystem and such, or do i have to use a cloud model for that

[nocturne zephyr]

Tools run in OpenClaw. OC sends a list of tools to the model, and the task at hand, and the tool that needs to be used is picked by the model. (Often several tools in sequence.) So OC will be able to run tools no matter where the model is coming from, as long as the model knows how to invoke tools. Some models are not smart enough to make the leap from task -> tool well, and so you get things not working so well. For that, mostly, you just have to test. There are some known good models for tool use, but the better they are at it, the more expensive they are.

[wind needle]

i see, so it seems most people feel open source models that run on mac studio isnt smart enough? at least anything 80B and lower for Qwen?

[hallow coral]

yeah, I have the same question.
every time i start a new session, i dont see my previous sessions. it looks like it is supposed to work like chatgpt's conversations, but it doesnt. I see the previous sessions in the sessions folder, but not in the UI, and i cant resume a session. this is a bit wierd, and im wondering what is the intended way to have multiple projects going.

[nocturne zephyr]

I think the current version of OC will warn you if you're using any models that are 300B parameters or below; it points out that they're much more potentially prone to data exfiltration through prompt injection. Now...there's been some testing on that front, and it's not completely true, but the idea that a smaller model is less 'safe' with your data is the way to bet. So something like Qwen3-Coding-Next is probably great, but having it read untrusted externally deliverable data (web pages, emails, tweets, unsolicited text messages, open Discord channels, etc.) is a Bad Idea.

[wind needle]

well i guess if we just, you know, ignore security for a moment, the question here is does under 300B params work just for zero security functionality? and it seems to me answer is closer to a no than yes
The question here is with OC so unoptimized on token usage, people dont want to dump hundreds a month and preferably use open source to save money (at least it is the case for me) and it seems to me peopl cant do that in the current iteration, and it seems also tool usage/ file system access for OC is a minimum requirement for it to be useful for most purposes

[nocturne zephyr]

I've made this point before... The author of OpenClaw was independently wealthy. Token costs were less of a concern than having a really capable assistant. They built that assistant, and now the rest of us platform pricing plebeians are trying to make it work for us, too, because it's shown that it's possible right now.

That said, I've read many people saying that they got Qwen3-Coder-Next working as a coding tool, and they use one of the other models (GLM5, Kimi K2.5, MiniMax) as the 'orchestrator' of Q3CN, and they reduce token-costs that way.

[wind needle]

i see, thanks for the other model recommendations, ill try them out
yea i mean if OC is going to basically force charging closed source models on us for hundreds dollars a month, to have any longevity in keeping using it, the income it brings with whatever products people create has to be more than the cost, its definitely not the case right now so as-is it basically cannot move beyond an experimental hype tool until it can at least solve this problem as a start. Forgot about security, it needs to be useful first beyond a play thing

[nocturne zephyr]

The problem is that there's a few hundred thousand people of varying levels of skill who suddenly have learned that this is actually possible and want it. And are trying to set it up. And it's insecure.

Now most of those people are not trying to use it with open weights models. Honestly, folks who even know what that is are rare. Most folks are using it with subscription models like the ones I listed, who have token caps over (week|5 hours|month) that reset, so you can't run them 24/7. They (like me) pay $25/mo. for a decent set of limits and a smart model, and accept that we're sending our data to one of these datacenters.

I can run GLM5 at home, but I don't use it for OpenClaw because 15 t/s is great for chat, but terrible for tool use. That's where most people will end up...using a cloud model in a subscription mode with OpenClaw. (OpenAI also offers a subscription, and thanks to hiring the author, they're being decent folks about letting OpenClaw users use that subscription in it.)

[normal wolf]

Hey guys, I can’t configure clawdbot properly or get it to work on my macbook pro

[wind needle]

well if most laymens want a plaything for 25 a month thats totally cool, but id imagine thats more entertainment, i was more thinking about potentials for business use and truly OC being a revolutionary thing today, guess we will ahve to wait

as for if people who dont get AI use it for chat, im not quite sure i get why use OC then over just normal LLMs like gemini or ChatGPT, i thought the whole point of OC is tool use and self initiating contact/tool usage instead of passive waiting for user input, if you are going to use OC passively (chat or not), then whats the advantage over LLMs?

[nocturne zephyr]

I'm not saying folks want it just for entertainment; lots of people think they need it. And yes, it's important to remember that a YouTube Content Creator's job is much more automatable than most people's, and so they are seeing immediate benefits from it, and so they're talking about it...because talking about it makes them money. So it's catnip for the influencer economy.

But other folks are setting it up, hoping that it'll make their lives better...and there's a real chance someone's going to get really badly burned, and go public about it, and it's going to be a firestorm.

Yes, folks have always wanted something which is self-initiating. I remember folks at one place I worked talking about how annoying it was that ChatGPT would tell them it'll do a thing in four hours, and then not do it. And that was because ChatGPT was trained on text that said, 'Okay!' but didn't actually have the ability to set a timer and do something, or do something on a schedule. (At that point in time, it didn't have that capability.) So it was just telling folks what they wanted to hear!

But what they wanted was what OpenClaw can deliver...but it's really, really early days. Think of trying to send your credit card to a website before the days of HTTPS-everywhere. The danger is real, if abstract and hard for laypeople to understand. And rarely encountered, which makes the risk decision harder.

[jagged nebula]

whats the best model i can use if i have 16.0 GiB of vram allocated

[nocturne zephyr]

A cloud model.

[atomic wolf]

Nanbeige/Nanbeige4.1-3B is a very interesting model. scored 76% in Live code bench

[lofty anvil]

Hi, after i updated to the latest version,

Expected:
On TUI, Main agent tasks delegates to subagent
Subagent executes the tasks and passes it back to main agent
Main agent summarizes and passess back to me on TUI

Current:
On TUI, Main agent tasks delegates to subagent
Subagent executes the tasks and passes it back to main agent
Main agent summarizes and passess back to me on Telegram

Has there been a config change? Been browsing the docs and my config but don't see any major changes

[atomic wolf]

doesnt need alot of vram

[jagged nebula]

video ram, not ram

[atomic wolf]

i think they meant cloud model because 16gb vram is very low to run a good model

[nocturne zephyr]

I know. 16GB of VRAM is not enough to run a model smart enough not to get phished / prompt injected, and not enough to run tools properly for OpenClaw. (cheap, fast, good: pick 2; small falls into 'cheap and fast' but not good.)

[jagged nebula]

i dont plan on running it on my own vdi instance lol.

only a fool would do that

[jagged nebula]

20b isnt good?

[atomic wolf]

what model and quant

[jagged nebula]

quant?

gpt-oss:20b

[atomic wolf]

quantization. a 20b model wouldnt usually fit into 16gb

[jagged nebula]

oh

well gpt-oss:20b is 14gb big

https://ollama.com/library/gpt-oss

[atomic wolf]

hmm interesting

its not good tho

[jagged nebula]

also i think 15gb of a P40(same chip as 1080ti) is better then 4 M10(same chip as gt 750) gpus taped together(32gb of vram in total)

[jagged nebula]

4.25 bits per pram MXFP4 format

i suppose offloading stuff into ram would be worse

but i could spare like 190gb of that

can i use one gateway for multiple users?

[past hawk]

Hello everyone,

are there are some pointers for using matrix and the plugin within openclaw properly. As the standard way does not seem to work properly. Right now its seeminly not been able to be installed in a easy way.

Any releated input here is welcome.

Thank YOU.

Greetings

[grim plover]

i just subscribed z.ai and i want to change minimax to z.ai how can i do it

[jagged nebula]

what are the recommended psecs for openclaw gateway?

[jagged nebula]

i get Gateway service check failed: Error: systemctl --user unavailable: spawn systemctl ENOENT

[storm loom]

Bros, is anyone use the Kimi Moderato api? I can't get the key to work

[gilded vine]

Dear crab-people,
I have a very simple openclaw setup, one agent, a few days old. Today we've been having trouble with him not seeing tool call outputs. I managed to get him to give me what he is* seeing, and this is it "[compacted: tool output removed to free context]". This is minimax-m2.5 so 200k context window.

Neither he nor I really have any idea how to fix that. Tried running openclaw doctor, updating to latest version, etc. Any tips? I have no earthly idea what's filling it up.

[storm loom]

1: Are you definitely using Minimax?
2: Is there a tools.md file?

[storm loom]

Tell your agent to do it

You can do it manually inside the dashboard but then what's the point of an agent?

[gilded vine]

Yeah it's definitely minimax, I can see the calls (qty) on my minimax account.

We have a tools.md file, yep. We have ~20 python scripts referenced which are in the in the workspace separately. Tools.md is ~200 lines.

[sleek arrow]

anyone know what they are doing? trying to connect my ai to discord but its not working

[elder hearth]

thx, will try that!

what subscriptions do you use? i have paid claude pro and openai pro (about 20 bucks/month) and ran into rate limits all the time. needed to find free models to atl keep connected to my bot. but output is poor in many ways now. so do you pay much more or did you setup more efficiently?

[gilded vine]

Oh I should say as well that it seems to just be the 'exec' tool where my guy can't see outputs.

Even dead simple things like printing text.

[past hawk]

Thats a great question I'm also tinkering about right now! Its a joke with gemini pro 2.5-flash models. 1 or 2 requests and waiting again. I do not get anywhere with this I have the feeling.

[unkempt jay]

hi all

[lunar lake]

@gilded vine its a known windows limitation, claude ended up with a workaround to it

[unkempt jay]

stupid questionn maybe, but i have openclaw installed with open ai api to talk with it. i have reacehd the openclaw ui, and everytime i want to use it, i open the windows power shell and start tui and gateway. and then??? like i dont get it how it should work for me?

Ive also tried some browser tasks, works from now and then with chrome extension, something lik comet browser if anyone used it.

[wild nymph]

Hi,
Has anyone here tried the following?

1. Setup mac mini with apple ID
2. Add apple ID from mac mini as a "family member" to your own apple ID.
3. Share calender/reminders and more?

[hybrid igloo]

Has anybody used LLM models from Azure Foundry or Azure Open AI and successfully setup openclaw? Would really appreciate if I could get some help here.

[nocturne zephyr]

Still poking this, huh? Out of curiosity have you tried just asking one to DM the other on Discord, and if it doesn't work, ask it to figure out why?

It's also possible that Discord prevents bots from seeing other bots...

That sentence didn't parse for me.

I'd focus on Discord, because it's easy to have multiple bots in a 'server' and that gets closer to the interacting than 1:1 messaging services like WhatsApp and Telegram, but you've got requirements I don't entirely understand.

[small canyon]

some recommandations or advices how to shorten tokens a lot?

[nocturne zephyr]

You do get that for the most part having one agent talk to another agent independently is just lighting tokens on fire for the fun of it, right?

Nothing on that page is about agents talking to other agents. It's all about YOU talking to different agents via different channels, all living in the same OpenClaw instance.

That I have set up; I have an agent just for the sysadmin heartbeat stuff, one for heavy coding, and one for managing the OpenClaw environment itself.

[burnt frigate]

Why not just have it spawn as subagent when needed? The session ends when the request is completed.

My main has the main workspace, but also has subagent workspace-codex and workspace-local with gpt-oss:20B for smaller tasks.

[lucid jolt]

even if I /new, I can't get it to do anything it will go to cooldown mode immediately no matter which model or provider I select.

[brisk raven]

Did you use OpenRouter by chance

[lucid jolt]

no

[brisk raven]

I had an issue wit it, and what helped me was

cd ~/.openclaw/agents/main/agent
nano auth-profiles.json

Remove the error log and start gateway again.

Then it worked

[lucid jolt]

ok, removing usageStats... hope it doesn't go into cooldown again 🤞

[brisk raven]

Havent looked deep into why cooldown happens, but that might be why

So I wouldnt be surprised if it just is the same for every provider

[lucid jolt]

didn't work, I think it's just filling up context. maybe I should test without referencing anything from the channel and ask it a new question.

[bronze tapir]

Channel config schema unavailable.

GatewayRequestError: web login provider is not available

[spark osprey]

hey, i'm trying to get my openclaw working properly with Ollama, its working for the most part however whenever i send messages to it, via tui or web chat it doesn't seen to remember anything i have said to it, even 1 message prior.
when i run the gateway with debug logs i can see historyTextChars=xxx increasing in number & messages increasing in it too. I have a feeling that maxMessageTextChars=xxx is causing an issue here but not entirely sure, that number seems to be quite low, e.g. 642 currently while the historyTextChars=1753.
Anyone experienced this or know what likely could be the issue?

[brisk raven]

Mhm, sorry then.

[young lily]

But you do have it working with another model, right?

[gritty oracle]

So, with oauth ChatGPT pro - codex models the only option? I believe so, looking for a definitive from someone who tried and knows first hand. Making sure I am not missing something

[spice gulch]

has someone managed to install in Docker and use openclaw besides chatting with it (install skills/set cron jobs/run headless browser/etc.)?

[spark osprey]

I tried a few models in Ollama and it does the same thing (instant amnesia), when i just go directly into Ollama it remembers fine what my previous messages are

[muted ravine]

Coding issue.

i tried to run openclaw on my laptop and make it create a project and told it to research about a niche discord bot and a niche python library, instead he did everything wrong, he didn't research enough,coded code that doesnt work as intended even when i provided the full project structure and the feature list and some example code it just worked terribly, i told it to use the integrated AI inside my ide for outputs of code and he will prompt based on his research.

HUGE FAILURE

is this even possible with such niche subjects? i really want it to work, also should i run it on my old phone? (currently no VPS and not willing to spend money on a VPS...)

[polar plank]

hi everyone , i want openclaw to read and reply on my whatsapp . i dont want to add it as a bot there . just give it access . i dont see any other way to do it other than using a business api./ has someone implemented something similar

[young lily]

Your instinct is close but maxMessageTextChars controls outgoing message size, not history. The real question is whether OpenClaw is actually sending your conversation history to Ollama on each turn — Ollama itself is stateless, so that has to come from the gateway side. What model are you running, and do you have any session/memory config set up in your AGENTS.md? That’ll help narrow it down.

[young lily]

Docker on Windows has a known permissions issue with named volumes. Fix is to pre-initialize before first run:

docker run --rm -v openclaw_home:/home/node/.openclaw alpine chown -R 1000:1000 /home/node/.openclaw

After that skills, cron, and headless browser all work normally. Alternatively switch to a bind mount and skip the issue entirely.

[young lily]

From what’s been confirmed firsthand: Codex models work via ChatGPT OAuth. GPT 5.2 Pro via OAuth hasn’t been reliably confirmed working — I’d treat it as unsupported for now and not worth the troubleshooting rabbit hole. API key route is cleaner if you need non-Codex models.

[spice gulch]

I have docker on raspberry pi, not windows!

[young lily]

That’s the key clue — works directly in Ollama, breaks through OpenClaw. That means Ollama’s fine, the issue is that OpenClaw isn’t injecting conversation history into each request. Check your session config — specifically whether you have a persistent session set up or if it’s spinning a fresh context on every message. Also worth checking if there’s a contextWindow or history setting in your provider config that might be defaulting to 0 or 1.

[young lily]

Niche subjects are actually where these agents can shine — but they need direction, not just a task. A few things that’ll change your results: first, model matters a lot here, make sure you’re not on a mini/Haiku for research tasks. Second, your soul.md is where you set the behavior contract — tell it to research before coding, never write code without confirming the approach first, and ask clarifying questions if the domain is unfamiliar. The “HUGE FAILURE” experience is really common before that config is dialed in.

On the phone question: not recommended. Bandwidth, battery, and it’ll get hot fast on any real workload. If cost is the constraint, Gemini API key is free tier and works well with OpenClaw.

[young lily]

WhatsApp personal accounts don’t expose an API — there’s no supported way in without the Business API. Some people use whatsapp-web.js as a bridge but that’s fragile and technically against WhatsApp ToS, so you’d be risking your number. Business API is the real path if you want this to be reliable. If that’s too much overhead, Telegram has a proper bot API and OpenClaw integrates cleanly with it.

[young lily]

ignore that Windows fix — different issue entirely. On Pi you’re on ARM64 so first thing to confirm is that the OpenClaw Docker image has an ARM64 build available. If it doesn’t, you’ll hit silent failures or it’ll run under emulation and be very slow. Also worth knowing your Pi model — if it’s a 3 or earlier, RAM is going to be a real constraint for any meaningful workload. What image are you pulling and what Pi are you on?

[gritty oracle]

Thx

[young lily]

Those two errors together point to a provider config issue. “Web login provider is not available” means the gateway can’t find a valid auth profile for that channel — either the OAuth flow wasn’t completed or the provider you’re trying to use doesn’t support web login. Which provider are you connecting to, and did you go through the openclaw onboard flow or set it up manually?

[spice gulch]

I am on raspy 4 8GB, I run those succesfully (can chat with the agent, but the agent is not able to do almost anything):

OPENCLAW docker setup flow:
cd docker
git clone https://github.com/openclaw/openclaw.ai
cd openclaw
nano docker-compose.yml (fix: allow docker cli container to connect to gateway: https://github.com/openclaw/openclaw/pull/12504/changes)
./docker-setup.sh

result:
When prompted:
  

Gateway bind: lan
    
    Gateway auth: token
        Gateway token: THE_TOKEN
            Tailscale exposure: Off
            Install Gateway daemon: No

docker compose cp openclaw-gateway:/home/node/.openclaw/openclaw.json ./openclaw.config
nano openclaw.config --> put the token where needed in the config: THE_TOKEN
docker compose cp ./openclaw.config openclaw-gateway:/home/node/.openclaw/openclaw.json
nano .env:
add:
CLAUDE_AI_SESSION_KEY=
CLAUDE_WEB_COOKIE=
CLAUDE_WEB_SESSION_KEY=
nano docker-compose.yml (edits on gateway, unsecure but local and for tests):
update ports section like this:
ports:
"18789:18789"
"18790:18790"

docker compose down && docker compose up -d

to make the webUI device-pairing:
docker compose exec openclaw-gateway node dist/index.js devices list
docker compose exec openclaw-gateway node dist/index.js devices approve --latest

docker compose exec openclaw-gateway node dist/index.js status

[young lily]

First, web session keys (CLAUDE_AI_SESSION_KEY etc) are fragile for tool use — the agent can chat fine but tool calls are unreliable with that method. You’d get much better results switching to a proper Claude API key via openclaw onboard --provider anthropic. Second, I don’t see any workspace volume mounts in your compose config — without those the agent has no filesystem to work with. Add a mount like - ./workspace:/home/node/workspace to your gateway service and give the agent that path to work in. Fix those two and you should unlock the full capabilities.

[spark osprey]

Model is glm-4.7-flash:q4_K_M
The AGENTS.md file is a default file with nothing taken out or added in from a fresh install.
So in my openclaw.json file the contextWindow is set to 128000, the session ID remains the same when sending messages and doesn't appear to re-generate every message i send, when i look in that session jsonl file i can see all the prompts / responses i have done.

[young lily]

if the session JSON has all your prompts and responses, OpenClaw is doing its job. The history is there. The issue is that glm-4.7-flash isn’t reliably using it. GLM models can be inconsistent with multi-turn conversation format even when the history is correctly injected. Try swapping to a different Ollama model — llama3 or mistral are much more reliable for conversational memory. If glm is specifically what you need, it may be a model-level limitation you can’t config your way out of.

[muted ravine]

I tried to use gemini api key but it doesnt work, says rate limit

[gritty horizon]

hello i'm trying to install telegram plugin but it said error: packege.json missing openclaw.extensions. What does it mean ? I searched everywhere to find a fix but nothing work 🙁

[young lily]

Gemini free tier rate limits are brutal for agent use — you’ll hit them after just a few messages. Two options: upgrade to a paid Gemini API plan, or use openclaw onboard --provider google which walks you through setting it up properly and gives you access to the full quota. If cost is still the constraint, Gemini 1.5 Flash has higher free tier limits than Pro — worth trying that model specifically while you’re getting started.

[deep pier]

[young lily]

That error means your OpenClaw install is missing the extensions registry that the skill system expects. A few things to check: first, what version are you on? Run openclaw --version — if you’re below 2026.2.19 the extensions system may not be present. Second, try running openclaw doctor and share what it says. If the install is incomplete it’ll flag it. Most of the time this is fixed by a clean reinstall on Node 22 — older Node versions can cause partial installs that look fine until you try to add skills.

[gritty horizon]

2026.2.22-2 is my version. I tried to enable the plugins, restarted the gateway but somehow it still disabled.

I'm currently running it on vps

[spice gulch]

I do not use CLAUDE for the LLM, those were just written to avoid the boring warnings (vars not set) at each docker command. Here's my docker-compose.yml:
services:
openclaw-gateway:
image: ${OPENCLAW_IMAGE:-openclaw:local}
environment:
HOME: /home/node
TERM: xterm-256color
OPENCLAW_GATEWAY_TOKEN: ${OPENCLAW_GATEWAY_TOKEN}
CLAUDE_AI_SESSION_KEY: ${CLAUDE_AI_SESSION_KEY}
CLAUDE_WEB_SESSION_KEY: ${CLAUDE_WEB_SESSION_KEY}
CLAUDE_WEB_COOKIE: ${CLAUDE_WEB_COOKIE}
volumes:
- ${OPENCLAW_CONFIG_DIR}:/home/node/.openclaw
- ${OPENCLAW_WORKSPACE_DIR}:/home/node/.openclaw/workspace
ports:
- "18789:18789"
- "18790:18790"
init: true
restart: unless-stopped
command:
[
"node",
"dist/index.js",
"gateway",
"--bind",
"${OPENCLAW_GATEWAY_BIND:-lan}",
"--port",
"18789",
]

openclaw-cli:
image: ${OPENCLAW_IMAGE:-openclaw:local}
network_mode: "service:openclaw-gateway"
environment:
HOME: /home/node
TERM: xterm-256color
OPENCLAW_GATEWAY_TOKEN: ${OPENCLAW_GATEWAY_TOKEN}
BROWSER: echo
CLAUDE_AI_SESSION_KEY: ${CLAUDE_AI_SESSION_KEY}
CLAUDE_WEB_SESSION_KEY: ${CLAUDE_WEB_SESSION_KEY}
CLAUDE_WEB_COOKIE: ${CLAUDE_WEB_COOKIE}
volumes:
- ${OPENCLAW_CONFIG_DIR}:/home/node/.openclaw
- ${OPENCLAW_WORKSPACE_DIR}:/home/node/.openclaw/workspace
- /var/run/docker.sock:/var/run/docker.sock
read_only: true
tmpfs:
- /tmp
stdin_open: true
tty: true
init: true
entrypoint: ["node", "dist/index.js"]
depends_on:
- openclaw-gateway

[young lily]

means the config change isn’t being saved — the toggle is writing to a temp state rather than your actual openclaw.json. Try this: after enabling the plugin, open your openclaw.json directly and check if the Telegram plugin actually appears there as enabled. If it’s not in the file, there’s a write permissions issue on the config. Where is your OpenClaw installed — default path or custom location?

[young lily]

On a VPS that’s almost certainly a file permissions issue. If OpenClaw was ever started with sudo, your openclaw.json might be root-owned and changes won’t stick. Try this:

ls -la ~/.openclaw/openclaw.json

If it shows root as the owner, fix it with:

sudo chown $USER:$USER ~/.openclaw/openclaw.json

Then re-enable the plugin and restart the gateway. Should stick this time.

[gritty horizon]

molt@jarvis-facs:~$ ls -la ~/.openclaw/openclaw.json
-rw------- 1 molt molt 1742 Feb 23 19:26 /home/molt/.openclaw/openclaw.json

[young lily]

volumes are mounted correctly, docker socket is there. This isn’t an infrastructure problem. Two things I need to know to help further: what model/provider are you actually using, and what specifically is the agent failing to do when you say “almost anything”? Like, can it read files but not write? Can it run commands? Does it attempt the task and fail, or just ignore it entirely? The behavior will tell us whether it’s a model tool-use issue or something else.

[young lily]

Permissions are fine, that’s not it. This looks like it might be a bug with the Telegram plugin specifically on your version — the enabled state not persisting after restart is a known issue some people have hit. A couple things to try: first, open openclaw.json directly and manually add the Telegram plugin to the extensions config rather than using the toggle in the UI. Second, worth searching the #bugs channel for “telegram plugin” — if others hit this on 2026.2.22-2 there may already be a workaround. Can you share the relevant section of your openclaw.json so we can see what the extensions config looks like?

[atomic sparrow]

yo

[young lily]

Yo

[gritty horizon]

How can I open the openclaw.json anyway. I'm kinda low tech.

oh

[nocturne zephyr]

That's a VERY small openclaw.json file. Mine is 8 times as large.

[young lily]

just run this in your terminal:

nano ~/.openclaw/openclaw.json

That’ll open the file in a simple text editor. Don’t change anything yet, just scroll through and look for anything that mentions “telegram” or “extensions” and share a screenshot of that section here.

[gritty horizon]

i think it's a known issue and will be fixing soon right ?

[spice gulch]

I have chatgpt 4.1 for my "brain" and 4 other agents for coding,managing,etc. I tried to set a cron, that I wanna see into openclaw dashboard, but fails mostly for permission denied. fails for the same issue when trying to install a skill (i asked him to install this: https://clawhub.ai/isriam/claw-guru). it seems to me that tries to launch commands that are not available. with docker install there's not systemd and openclaw won't work. For manual commands (i tried to setup the cron with openclaw-cli cron add and works fine) you have to use node dist/index.js instead of openclaw when using exec openclaw-gateway.

If you have specific tests to ask, I can do them if useful to understand what's the issue.

[young lily]

The Telegram plugin is working on current installs I’ve seen — it’s more likely a version-specific bug than a removal. Stay on it. Check #announcements or ask in #general if there’s an official statement about it. If there isn’t one, assume it’s a fixable issue and not a dead end.

[earnest jay]

Hello all! I hope everyone is doing good. Can you guys confirm if max plan is ok to be used on openclaw with token?

[spark osprey]

when i try llama3.1:8b it just throws back rubbish about providing it with metadata and stuff, don't think it works with tooling context (i may try and pull a bigger variant if you think that would actually fix it, or maybe a specific one in ollama if you think?).
I tried loading in gpt-oss:20b same issue with 'instant amnesia'.
I tried loading in gpt-oss:120b, bit slower but still forgets previous messages.

[young lily]

You need to tell the agent explicitly how to run commands in your environment. Add something like this to your soul.md:

“When running OpenClaw commands, always use node dist/index.js instead of openclaw. Never use systemd. For cron management use openclaw-cli cron add from the cli container.”

That re-routes the agent’s tool calls to what actually works in your setup. The permission denied on skill installs is the same issue — once the agent knows the right command path it should resolve. Try that and let me know if the cron shows up in the dashboard after.

[young lily]

Yes — Claude Max via token auth is confirmed working. Just make sure you’re using the session token method through openclaw onboard and you should be good to go.

[hollow trail]

Hi. I was using Gemini CLI and Antigravity, and After just three times of chat, I am getting hit by the limit. Even after an hour, the limit keeps increasing. I can't do anything about it. Is there something wrong with OpenCLAW? Because in models I could easily talk long time.

[young lily]

Okay, same issue across 8b, 20b, and 120b changes the diagnosis entirely — this isn’t a model problem. Something is wrong with how OpenClaw is formatting the conversation history for the Ollama API specifically. Ollama uses a different chat template format than OpenAI-compatible endpoints, and if the history isn’t being formatted correctly it gets ignored regardless of model size. Check your provider config — are you using ollama as the provider type or openai-completions? If you’re using openai-completions as the API type for Ollama, that could be causing the mismatch.

[young lily]

The rate limit issue with Gemini CLI is expected — it’s a low-quota endpoint and OpenClaw’s context injection makes it hit limits faster than normal chat. But the bigger thing I want to flag: Antigravity is a known ToS violation with Google. Multiple people have received permanent bans on paid accounts including AI Pro/Ultra. Google has explicitly confirmed it violates their terms. I’d stop using it before you lose your account. Safe alternative is openclaw onboard --provider google using a direct Gemini API key — free tier available, no ban risk, and you’ll get better quota control than CLI.

[hollow trail]

Yep. My account get ban too. Is there any better choice? Some people recommend OpenRouter, but I don't trust them. In the old version, I especially chose carefully only free models that cost zero dollars. But OpenRouter took credit and wasted my money.

[young lily]

Sorry to hear about the ban — that’s exactly the risk we were warning about. For free with no credit surprise, direct Gemini API key is your safest bet. Go to aistudio.google.com, create a free API key, then run openclaw onboard --provider google and paste it in. Free tier includes Gemini 1.5 Flash with decent limits — no OpenRouter middleman, no credit card needed, and you control exactly what’s being used. It’s the cleanest free path right now.

[lofty anvil]

Hello, Can i check why when i message my main agent in tui, with the session (web:chat-g-agent-main-main), which delegates to subagent but replies come back through telegram, isntead of the TUI session?

[hollow trail]

Especially what people doing for 24/7 hours working with MAC? They using a local AI? Or there's Provodier natively support openclaw? I know Minimax doing it. But I don't know what's best for free using.

[earnest jay]

i keep getting this: OpenClaw: Agent failed before reply: All models failed (2): anthropic/claude-opus-4-6: HTTP 404 not_found_error: Not Found (request_id: req_011CYR2AHGayYoosy9mKCpxr) (model_not_found) | anthropic/claude-sonnet-4-6: No available auth profile for anthropic (all in cooldown or unavailable). (model_not_found).
Logs: openclaw logs --follow

[marble coral]

@young lily fyi apparently on mobile phone we modal gl5 endpoint webpage doesn't render correctly for some reason, so you can see the sidebar for the token generation etc...

[dusk geode]

Is krill down?

[muted ravine]

I got github students for free

[spark osprey]

so in my openclaw.json file i have listed tthe "api" configuration as "openai-responses" is that right for ollama?

[umbral cobalt]

It would seem like it, I wonder if krill got taken down by the new update… the new update changed how the start script works, if you had the OG start script on your system it broke your entire installation, by deleting the Systemd boot file, and if your Linux permissions for OpenClaw are setup correctly, the combination makes a permissions nightmare… took me 45 min last night to fix mine. So krill might be going through the same thing.

[nocturne zephyr]

Try openai-completions and see if that works better.

[dusk geode]

Hopefully they bring him back soon. I have to use it like every day 😂

[limpid bridge]

Hello. During the onboarding process I don't see a list of skills. Only one or two are available. Do you know why?

[umbral cobalt]

Do you have Gemini Pro? Gemini 3.5 pro is a beast at solving OpenClaw issues, almost better then krill sometimes.

Make sure to tell it OpenClaw version and operating system and error code in your prompt

[umbral cobalt]

You can honestly ignore any skills in onboarding, and tell you agent what you want it to do and it’ll build the skills. What’s Model you running it on?

[dusk geode]

Thanks! I don’t have the paid version but i can use the free ones.

[umbral cobalt]

If you have a .edu email you get Gemini pro for free

https://gemini.google/students/

[spice gulch]

trying to fix about cron jobs, about the skills setup issue, I think it is different, here's my error:
https://discord.com/channels/1456350064065904867/1475448902445568060

[spark osprey]

I just tried this setting with a few different models and no difference, just instantly forgets any previous messages

[dusk geode]

Im old. I have a edu email from like ‘03

[umbral cobalt]

As long as you can login to the email, you can get it

[dusk geode]

Ill try thanks

[bitter lynx]

Can I take my old cell phone, a Samsung Galaxy S 23 Ultra and run OpenClaw on it as a stand alone, meaning, needing no other computer?

[high heron]

its possible i guess, but not recommended

[spark osprey]

ok think i managed to fix it - changed it to "ollama" and now it is starting to remember things & the token usage jumped above 4.1k in openclaw tui

[high heron]

https://github.com/AidanPark/openclaw-android
this will help

[bitter lynx]

Wow, yes, thank you. Right now I am exploring the idea

[spark osprey]

fyi for anyone else having this issue, the solution is in the documentation - https://docs.openclaw.ai/providers/ollama#explicit-setup-manual-models the
api: "ollama",
line is what you should look at.

[gritty oracle]

in a multi agent system each with their own workspaces, can you create a config file in one of the workspaces and only set model thinking values there and model selection for that agent, while all other config settings get inherited from the main openclaw.jason config? No way to otherwise persist and set unique model thinking level per dedicated agent from what I can tell and this was last resort.

[elder hearth]

you can add commo channels, like telegram (easiest) or discord. then you can "talk" to your agent and it will respond and send you results in that channel

[pastel thicket]

hello, my claw keeps complaining:

Workspace: 16 untracked/uncommitted items

And trying infinite loop to fix it but it cannot fix it, what does this even mean?
Commit what?

[inner bronze]

i hosted the clawbot on hostinger and since 6hrs i have been trying to connect channel discord/telegram. Any idea whats the problem

[umbral cobalt]

Be careful! Openclaw is being heavily targeted right now:
2. 🔐 SSH Brute Force Activity
⚠️ 6,437 failed/invalid SSH attempts in the past 7 days.

Top attacking IPs:
• 91.202.233.33 — 216 attempts
• 176.120.22.47 — 204 attempts
• 176.120.22.13 — 183 attempts

Top targeted usernames: admin (707), user (201), ubuntu (168), test (137), sol (131)

[faint forge]

Is there any guides how to add oolama?

[small canyon]

So how can you know if your system can be hit by this,...
If the OpenClaw security audit is excellent, and everything is setup correct, is this still an security risk?

[limpid bridge]

i have set OpenRouter/auto and my bot can not figure out anything even simple cases like handle audio msg from telegram or remainder me something in 10 minutes 🤦‍♂️

[umbral cobalt]

OpenRouter/Auto is terrible... Your honestly better off using AntiGravity, or paying for an OpenAI Pro/Claud.AI max plan.

[elder hearth]

start with TG only. that is easier to start with. i did it with the onboard/configure process. there it told me the exact steps to make it run.

[umbral cobalt]

100% honest, your best bet is to get Claude Max 200, and setup your OpenClaw during the first 30 days get that usage to 100% and then switch to a lesser/cheaper model. Build the foundation, then go to a lesser model. Building a foundation on a model that doesnt know anything will lead to bad results.

[onyx moat]

im this loop trying to fix this. anyone familiar with this telegram message

⚠️ API rate limit reached. Please try again later.

[umbral cobalt]

Your model provide has a Message per Min/Hour/Day limit that you hit is usually what that means.

[elder hearth]

openrouter/auto is super expensive on top. i added 5 bucks for higher free limits and instead of using the free limits it immediately used gpt5 and opus4.6 for simple tasks until the funds were gone in a few minutes .. haha. openrouter works now, that i told openclaw to use specific free models, not sure how good tbh though, it seems not nearly as good as claude but i will try a bit more, as nothing is worse than hitting rate limits all the time after a few requests

[onyx moat]

how do i fix this? its been like this for the past 16hrs

[umbral cobalt]

What Model you using?

[onyx moat]

gpt 5.2

[umbral cobalt]

On what plan?

[limpid bridge]

ok @umbral cobalt @elder hearth THX for an info, will try!!

[onyx moat]

i think im on the plus

yeah, Plus

[umbral cobalt]

Most likely you hit rate limts then.

Plus you will go through your rate limits quite fast with openclaw

[umbral pelican]

Anyone having issues with new installs on pi? I have one setup no problem, now when I try I get whole list of errors

[umbral cobalt]

I hit ratelimits on Claude Max 200

Ask your Main agent to build yo a Team of Multi-Agents, and setup a template for it, and send it this URL make sure you run on Opus, GPT, or Gemini Pro when doing that, and itll build out all the other agents, and then you can tell it what models you want them all to run on, and have ti build it out for you, and then switch to a lesser/cheaper model again, and tell it to boot them.

[civic wind]

hi i'm trying to set up signal or whatsapp on my openclaw but its saying channel config schema unavailable when i go to the hosting site , can someone help?

[fair belfry]

do you have it set up wiht open claw?

man... i am burning through credits and need a more ecconomical solution

[umbral cobalt]

@agile oracle I sent you a DM cause we cant post photos here.

[inner bronze]

Hey whats TG? You mean running from scratch step by step without using Hostinger API keys, etc ?

[mint umbra]

telegram plugin not available. 怎么回事

谁知道吗

[muted ravine]

help my telegram plugin isnt installed on my open claw

cant connect to telegram what do ido

[final hound]

Hi all, new to this, please can you help. How do I give my main agent web access and host execution from the sandbox? Getting 'Error response from daemon: Forbidden' on all exec and outbound requests. Tenant: 61bbxxxx"

[umbral cobalt]

Yes i use my openclaw iwht Max 200 (Currently...)

[molten geode]

I switched to my ChatGPT plus subscription and it worked for days. Now I’m getting:

⚠️ API rate limit reached. Please try again later.

I’m hooked now but spending too much on tokens is I use the API.

[meager ocean]

after update my dear agent wont reply my on telegram anymore. What can be done?

[candid socket]

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

its very easy to reach the weekly usage for plus

[meager ocean]

well I be damn! suddenly its back.. after 20 min on holiday!

[molten geode]

I was enjoying it so much not worrying about API costs while experimenting

Thanks Goarder!

[candid socket]

its not api, its codex oauth. but anyways, you can force limit context on every session to slow down the usage and tokens used

[molten geode]

I was burning through a lot of API tokens before I switched to codex oauth

[candid socket]

also use free models for crons jobs (best is if your cron job can be done using a script, then just use it to call the script and process output)

same for heartbeat

[molten geode]

Have you found good free models that are cloud based? I’m so into this. I’m considering buying a Mac studio ultra so I can run good models locally I believe I could run a 70 B model. My problem is the cost there too. I’m a little worried about spending so much money without knowing that the models will work well for open claw.

I was building out a script when my rate limit kicked in!!

[candid socket]

you can look through open router, and test stuff out, i think a 70b model may struggle but if its just running a script and processing output of the script, i think its ok

[molten geode]

Nice. The 256 gb model is. $6k

[high heron]

open OpenClaw on web

you will see codex's reply there

this bug sometime occurs

if you still see this message on web, then you should wait till limit resets

[candid socket]

ye its a huge investment for smt that may not return the investment

its really about how you plan to use openclaw

[molten geode]

If the Mac works, I’m thinking I could share it and set up an open call for my son and wife

[shy ether]

I have had a nightmare of a time trying to get openclaw to function... I'm running local llms and can't seem to find one that really works. But I also can't find any solid documentation on how to correctly configure the models in openclaw.json....

I have a Quadro P5000 with 16GB VRAM that I'm running the models on, via ollama.

[candid socket]

check dm, will share a way to get some good models

[molten geode]

Ollama can configure open claw for you

[shy ether]

I'm running ollama on a different PC to openclaw.

[molten geode]

Ya. I had that same problem. I created an SSH tunnel I believe using ChatGPT to get the right command, so ollama was able to act as if open claw wad running on the same server.

The main issue I had with local models open call would say it was running a local model, but if the context window was less than 16,000 it wouldn’t use it and would default back to my cloud model, but would not give me a warning or a notification

Once I changed my context window to 16,000 then the requests started going to my old server, but my CPU was paid at 100%. It’s not fast enough or new enough memory to run a good model.

[shy ether]

But I'm starting to think that you can't run this on only 16GB VRAM. I honestly think all those who have posted online saying they've done it, are liars.

[molten geode]

I don’t think it’ll be worth it

I was researching Nvidia 4090s and I don’t think they can run open claw appropriate models

The Mac studio ultra has 256 or 512 gigs of unified memory, which allows it to run big model fast

But it’s six or $10,000 if you want 256 or 512 gigs of RAM

Weekly usage limit
6%
remaining

[umbral pelican]

Anyone having issues with loading in openclaw on raspberry pi?

[molten geode]

You're nearing your Codex limit. Add credits or upgrade to keep building. Your limit resets on Feb 26, 2026 7:07 PM.

[sturdy forge]

can anyone point me to instructions to add Open AI codex as a model via Oauth?

[molten geode]

I’m running on a raspberry pi 5 with 8 gb of ram

[onyx moat]

im on the same boat.... I added to the api cash to the api but still have api rate limit. im on plus, and i feel like i havent done much with openclaw, just getting started and now im at a stand still

[molten geode]

I spent hundreds on API tokens with anthropic and OpenAI before switching to my subscription model

Now I’m stuck too.

[umbral pelican]

I’m trying to setup a new one, and it’s erroring out

[molten geode]

I ran the install script and it just worked

[candid socket]

use cli openclaw configure

[molten geode]

You have an error message?

[meager ocean]

Yeah my telegram agent is no longer responding me after the new update. Any got a fix?

[candid socket]

openclaw burns token extremely fast, i found it better to limit context/token sizes so that its forced to compact often

[molten geode]

https://youtu.be/7DNlQgl2Kk0?si=4biie2yYw4lHfDE0

[umbral pelican]

Running the script from the site directly and getting 10 pages of errors after a few moments

[molten geode]

That video showed how to configure open claw with your OpenAI plus subscription

[alpine anvil]

hey guys i need help. i was using gemini as a fallback for my openclaw. i told my agent to get rid of all instances of Gemini and Google with the news that google is blocking openclaw. But for some reason my agent switched to Gemini, is unresponsive, and i dont know how to get it back

[molten geode]

Do you have an old session?

[onyx moat]

how do i limit context/token size

[molten geode]

I’ve cleared out my session files

[molten geode]

Run /new more often helps

[candid socket]

check your openclaw.json and add the models that you want manually. use openclaw configure to do it if you want also.

then set the models for the sessions

[alpine anvil]

thank you!

[candid socket]

i am not using it anymore, but i beleive its agents.defaults.contextTokens

i had set mine to 60k-120k in the beginning so that i dont get fked with high usage / price

[onyx moat]

I still have API funds available though. That has nothing to do with the rate limit?

[candid socket]

its also good to set these under agents to not trigger rate limits so easily:

"maxConcurrent": 1,
 "subagents": {
  "maxConcurrent": 1
  }

[candid socket]

my bad then, since you mentioned plus, i thought you were using the oauth method, not the api method.

[onyx moat]

im using api method. is oauth better?

[molten geode]

I was using a lot of money on API tokens so it was nice to use oauth and get value from my $20 a month sub subscription.

But now I ran into a limit on my plus plan

[prime pelican]

Hi OpenClaw team, I’m trying to update to version v2026.2.22 via npm, but automated commands are not providing output or completing. My assistant (Ada) has attempted multiple approaches (npm install -g openclaw@v2026.2.22, openclaw status, npm ping), but there’s no response. I don’t want to handle updates manually. Can you advise on how to resolve this? I’m on Windows, using npm v24.13.0. Thanks!

[molten geode]

You might as well configure it and burn through the resources that they give you just to save money on tokens

You're nearing your Codex limit. Add credits or upgrade to keep building. Your limit resets on Feb 26, 2026 7:07 PM.

[shy ether]

Hmm, okay. So those folks are liars then. Probably click-bait bot posts, trying to pay off their claude subs.

[molten geode]

I even rented an online 4090 from run pod.io

[shy ether]

Yeah, I've use vast ai to rent an RTX Pro 6000 for i2v. But damn it's expensive.

[molten geode]

You can host a local model there and connect to it. I bought $50 they charge per hour, and the results weren’t great

The alternative though is buying a Mac Studio ultra M3 with 256 gb for $6000

[onyx moat]

thanks for the advice.... dont know how to go about this now..... i feel like ive accomplished nothing and already hit rate limit... is there a model that you guys recommend for high usage?

[molten geode]

I believe that will run a 70 billion parameter model so now I need to go back to run pod and try it with a larger model

[shy ether]

Yeah, I'm not buying dedicated hardware for a toy.

[molten geode]

Haiku worked as a simple main model

I did switch to sonnet when I was doing more coding though

[molten geode]

Yeah, I don’t know what to do not ready to give it up. I’m finding it so helpful for managing my to-do’s and doing research for me. It has its own email address. I can forward things to you. I love it.

I guess I see the utility beyond a toy, but I still kind of have it in the toy bucket and so I’m having a hard time burning through tokens until the end of time or spending thousands on an unproven solution

[shy ether]

That's what I wish I could do, but it doesn't do a thing. I don't know what model would fit in 16GB RAM that would actually be able to handle anything.

I'm sure it's helpful, but I don't know if it'll ever be 8000 USD/year useful.

[molten geode]

From what I’ve learned and it took me a little while to get here not that it’s complicated. Just life happens. It’s the context window that really is expensive and 16,000 is the minimum it really wants 64,000

I think you need a 30 billion or 70 billion parameter model to start being useful and then you combine that with the context length which requires exponentially more RAM as it grows

So from what I’ve researched the 4090 will get you good performance, but only with a small subset of the memory you need

[shy ether]

My context window is already 64k. My model's VRAM doesn't budge beyond 11GB when running.

[molten geode]

If you get the max studio ultra with the unified memory, you can keep the whole model in memory and actually make them usable. But we’re used to working with 1 trillion parameters in these cloud models with a 70 billion really work

[molten geode]

Interesting I’m having a hard time figuring out what’s real and what’s not. Because I don’t have a graphics card to experiment with locally.

What local model are you running?

[shy ether]

Currently gpt-oss-20b

[molten geode]

Have you tried qwen2.5 or qwen3?

[shy ether]

I tried qwen2.5 and it kept replying in chinese and thai.

[molten geode]

openAI needs to sell an open claw subscription plan

[shy ether]

and had zero memory. I'd tell it it's name, and then immediately ask it what it's name is, and it'd have no idea.

[astral kelp]

Anyone having trouble getting the nano banana pro skill to work right now? Keep getting told 503 but I can generate directly from google

[molten geode]

Yeah, sometimes I would see it describing that it was talking to me instead of actually replying to me sometimes the models don’t seem to work with the open Clock configuration

I think we all would pay 50 bucks a month for an open claw subscription that wasn’t limited

I’m having a hard time deciding to pay 200 bucks a month though for the pro plan

[late nebula]

i want to start simple chatbot without tools and use main agent for this. the chat have succeful sessin but if say "hello" return to me system messeges like {"name": "tts", "arguments": {"text": "Please read WORKFLOW_AUTO.md and memory/\d{4}-\d{2}-\d{2}.md immediately."}} . what is wrong?

[molten geode]

I had a similar experience when I experimented with different models, especially the lower parameter ones

I don’t believe it’s a configuration problem. I think the model just isn’t able to handle open claw.

[shy ether]

I see... So there isn't a single model within 16GB VRAM that can run?

[late nebula]

DocNasty do you answer to my question?

[molten geode]

Yes, I was replying to you. I tried several models on my own local hardware and rented a GPU at Ron pod. I would get weird responses like that from some configurations.

[late nebula]

what is a good model for chat. i use qwen2.5-coder

[molten geode]

My problem with paying for api tokens is I’m trying to figure out what’s possible so I’m just experimenting and trying things out and you’re less likely to do that when there’s a cost associated with each attempt

[late nebula]

i was having the same problem with ollama api key but solved

[molten geode]

What was misleading is I can run some of the good models qwen3 coder in Ollama and chat with them. The responsive and amazing. But when you run Ollama ps I saw that they’re only using a 4K context window

You need a 16,000 context window just to get started with open claw and it recommends a 64,000

With a 16,000 context window my CPU base system with 32 GB of RAM basically doesn’t return a message and just turns into a loud fan and a heater

Gemini says that to run a 70 billion model with 64K context locally on a Mac you need a machine with at least 96 gigs or 128 gigs of unified memory.

[sturdy forge]

worked thanks alot!

[molten geode]

I just learned how to watch your usage. Here’s the link.

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

[viscid gust]

Hello everyone, I apologize in advance for my stupidity. Can you tell me which model can fully function on a local network as an agent that can at least pass Botstrap? Win 10 GTX 3070 8G / 16GB RAM. I've already tried many, but none except the more powerful ones can handle the potential of OpenClaw.

[molten geode]

Sorry king man I think you’re running into the same problem. I was you really need 100+ gigabytes of video ram to make this work.

[shy ether]

Apparently it wont work on anything less than 30+GB of VRAM

[molten geode]

And I think the reality is you want 256 to have a smart agent

[late nebula]

Qwen2.5‑Coder‑7B‑Instruct supports a 32,000‑token context window.

[viscid gust]

That is, there is no chance ? I just saw the model. https://ollama.com/voytas26/openclaw-qwen3vl-8b-opt which works well , but for a very long time . I thought it might be possible to train qwen2.5 for minimal interaction at least.

[hollow shoal]

Guys i have on other machine Ollama 70B but i cant connect it because of the bad documentation. i cant figure out how to connect that. it needs apikey

[viscid gust]

Why can't my models pass BOOTSTRAP.md even those who are stronger and respond for 20-30 minutes ?

[shy ether]

add "apiKey": "literally anything here" to your openclaw.conf, under the model entry.

[hollow shoal]

its not accepting it

it just throws an error

[high portal]

Yooooo Is there a place to help fella like me that's coming across some common issues? My convos are eating up my tokens, but when I condense it down my clawbot isnt as proactive if that makes sense

[shy ether]

    "providers": {
      "ollama": {
        "baseUrl": "http://10.11.1.116:11434/v1",
        "apiKey": "ollama-local",
        "api": "openai-completions",
        "models": [
          {
            "id": "gpt-oss-20b",
            "name": "GPT-OSS-20B",
            "api": "openai-completions",
            "reasoning": true,
            "input": [
              "text"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 65536,
            "maxTokens": 8192
          },

Something like that.

[hollow shoal]

    "providers": {
      "ollama": {
        "baseUrl": "http://192.168.190.72:11434",
        "apiKey": "ollama-local",
        "api": "ollama",
        "models": [
          {
            "id": "agent-llama:latest",
            "name": "Agent Llama 70B"
          }
        ]
      }
    }
  },```

i have this now

[shy ether]

Hmmm. What error is it kicking up?

You may need to put the /v1 at the end of the base url?

[cloud swift]

"Does anyone have a pre-built .ipa file for iOS I can sideload? I'm on Windows."

[hollow shoal]

tried with and without

[earnest oxide]

lets say i want to get open claw in a vps and later on move on to a mac mini or something more local how can i move the files from the vps to local ?

[shy ether]

Does the 'id' match the model name ollama ps outputs?

[hollow shoal]

what do you mean the id