users-helping-users

[inner monolith]

hmm I don't really know but I think it's free though what do you think @tall meteor

Holy shit 😳

[granite bobcat]

I recently switched to Claude

👀

[copper nova]

this is teh way

[inner monolith]

Code?

Or the chat

[granite bobcat]

Will be most likely yes

[copper nova]

people actually use claude chat?

[tall meteor]

That's what I like about Ollama - $0

[copper nova]

why

[granite bobcat]

For people who ask AI what the weather is like. Small talk lol

[inner monolith]

Yes me haha, because I don't have money to use Claude code

[copper nova]

why the fuck

[tall meteor]

Claude has better reasoning than ChatGPT

[granite bobcat]

Wait just a damned minute. 16k in hardware but no monies for Claud code?

Something is fishy

[copper nova]

that was what I was thinking too

[tall meteor]

...no money left

[copper nova]

spent it all on the car

no money for gas

[granite bobcat]

lol makes sense

[copper nova]

i get it

[tall meteor]

I want the RTX Pro 6000
But I don't have $10K

[copper nova]

go chat with the guys in #hardware

eye opening, but not for me

[inner monolith]

True

[copper nova]

I'll just keep paying daddy Amodei

[eager river]

Hi, I'm struggeling to get it running with CachyOS and local llm. Anybody got it running on CachyOS or with Node.js 25? I get "fetch failed" after I type openclaw tui

[inner monolith]

👀😒

[inner monolith]

😐😐

[tall meteor]

My system will be off-grid, no internet required

[granite bobcat]

What will you use it for?

[tall meteor]

development

[copper nova]

how do install node packages?

lol

[tall meteor]

local repo

[copper nova]

have to install first

don't fuck around here, I'm a fullstack dev lol

[tall meteor]

I've been doing full stack since the early 2000's

[copper nova]

it ain't rails and JSPs anymore

[tall meteor]

Started with LAMP and gold ol' Tomcat

[tall meteor]

Really? 😄

[copper nova]

tbh, alot of companies still on rails

surpising amount of them

[tall meteor]

RoR is a heap of trash

[unborn jetty]

if i go the hostinger route do i still need an api key from openai or anthropic?

[tall meteor]

I still do RoR work for a company now and then

[copper nova]

yes, hostinger doesn't give you tokens

just hosting your VPS

[unborn jetty]

got it

im stuck at model/auth provider

[copper nova]

I heard hostinger has some bug with their setup atm

my buddy tried them and can't get github access

[tall meteor]

but installing a .deb file gave you excruciating pain 🤣

[unborn jetty]

is anthropic the winner fellas?

[copper nova]

not sure what's going on with that

[copper nova]

I got it done, just not a great UX which triggered me

[granite bobcat]

Seems that way pal

[unborn jetty]

thx

[granite bobcat]

Ever notice calling someone "pal" is very aggressive sounding?

[copper nova]

it's like when I went to a government website and they gave me a up/down input for a date selecter

I almost threw my laptop out the window

[granite bobcat]

What would you expect from the gov?

[tall meteor]

date pickers are a national security risk

[unborn jetty]

so is calling someone guy lol

[granite bobcat]

We hardly have driveable roads

"Hey guy"
"I'm not your guy friend"

[copper nova]

call them chief instead

[tall meteor]

Local, state, and federal government web development is sad.

[sweet cargo]

Does anyone know why setting tools.allow and tools.deny isn't working? (In control UI all tools are still in default setting and agent testing verified the permissions never changed.) But setting tools.profile worked.

[karmic notch]

Hi guys. I posted above, but no response. I’m sitting on a down bot because of a persistent 409 error. The gateway seems jammed for Telegram and Discord for my man bot. I can still communicate with one of my sub bots. I was working with another AI agent to try and repair the issue with no success. Any ideas for me??

[unborn jetty]

question if i pay $20 for my openai chatgpt already if i use that api key will my monthly increase?

[copper nova]

you can't argue that linux is a great user experience my guy, please

[granite bobcat]

Steer clear of other agents. They don't know squat about OpenClaw. Sorry, that's my only input.

[tall meteor]

90% of the internet has Linux in its veins.
We're using a Linux server right now 😍

[granite bobcat]

I would imagine so? You mean you're paying 20$ for GPT pro or something? I don't think that has anything to do with API

[copper nova]

you and that penguin can take that shit offline

[tall meteor]

nah

We're here to wreck it all

[unborn jetty]

ahh so how much would i be looking at if i went openai route? or would you say just go anthropic since openai is kinda sqaure

[karmic notch]

Really? That’s concerning. Anyone have input of this or a solution for me?

[tall meteor]

You will own Linux, and be happy

[copper nova]

I actually think you can, OpenAI very freindly with it AFAIK

[granite bobcat]

Doubt I'll be much help, but can you describe the issue in detail again?

[copper nova]

since they paid daddy steighberg a fat billion

[granite bobcat]

fr??

[unborn jetty]

which ai will allow me to command on whats app or telegram the easiest?? lol

[copper nova]

openAI doesn't have a middle ground, its $20 dollars or $200 dollars, kinda wild

[tall meteor]

Starting a new Claude session

cat openclaw_cli_reference.md | xclip -selection clipboard

Ctrl+V
🎉

[unborn jetty]

so you think for $20 i should be good for openclaw basic usage nothing crazy

[copper nova]

should be fine

[unborn jetty]

ok will try that

[copper nova]

like MY stack is github-copilot for provider and claude code for actual development on it

and gemini because I have to

[granite bobcat]

the hits keep comin!

[copper nova]

yea second thought, do the github copilot license

it's $10

and you can use the opus to unblock you

[steep dragon]

hey guys , after pulling a model in a host pc with lmstudio, pinging tyhat model on opencalw in a vm and addin ti, it respondes, but it doesnt show on the list in the web ui... how do i updtate it to show on those lists???

[unborn jetty]

so select copilot and where do i get the login? go to github and sign up there?

[karmic notch]

My bot did a memory flush and somehow its gateway got messed up. Followed along in logs and there is a 409 error code. I can get it to respond once in a while but only a message or two before he is non-responsive again. I can’t connect via Telegram or discord. I can still communicate with a sub bot which shows my computer is running and that gateway is open. I spent several hours trouble shooting w an open claw AI agent. With no luck. Really no idea what to do. Looking for some ideas of what to try next. Help!

[copper nova]

yes look up github copilot plans, sign up, then in openclaw onboard, select github copilot, it will give you a URL and a pin number to input into it

then hocus pocus, it's wired up

[unborn jetty]

thank you

[copper nova]

I think it defaults to sonnet first

but that should be enough to get you rolling for sure

[ancient bobcat]

any setups for orchestration openclaw agents?

[copper nova]

300 requests in the $10 plan

[unborn jetty]

sweet sounds like a good place to begin

[granite bobcat]

Have you tried openclaw doctor --fix?

[unborn jetty]

can openclaw post on socials like meta, ig, x, etc?

[karmic notch]

No, surprisingly. Does it work generally, or do I need to be specific?

[copper nova]

yes, but X costs money, it can do it for free but it's kinda jank using browser skills

[granite bobcat]

It works generally, you don't have to specify anything. It's a shot in the dark but worth a shot

[karmic notch]

It is worth it. Thanks

[copper nova]

just stealing my ideas, I see you

[granite bobcat]

It was Gemini!

[tall meteor]

I'm writing a system scheduler with Node.JS/Fastify and Redis
It will prioritize the GPU, and agent requests will go into a FIFO queue

[steep dragon]

hey guys , after pulling a model in a host pc with lmstudio, pinging tyhat model on opencalw in a vm and addin ti, it respondes, but it doesnt show on the list in the web ui... how do i updtate it to show on those lists???

[copper nova]

lol I'm probably one of the only people here that knew what you said there

[granite bobcat]

I understood first in first out. Serving tables paid off

[copper nova]

you need to calm down on that nerd talk, no one understands that

[tall meteor]

OpenClaw is nerdville anyway
It comes with the territory

[copper nova]

and honestly feels like over engineering lol FIFO queue also not idea. Why queue?

async threaded, do everything in parrallel at once, why am I queuing, does order matter?

[tall meteor]

A Redis queue doesn't require polling or cron.
Then each agent team can share the same GPU, even though they have their own tasks to complete.

[copper nova]

promise.All([everything])

This is for your local LLM setup specific

[neat walrus]

@granite bobcat I got it to list the workspace directory now ... I think as it was asking for approval each time (and I click approve once AND approve all to no avail) - I think it was failing to validate because the dashboard I was using wasnt updated with the latest gateway token .. I had to go to my powershell and do Openclaw Dashboard and it refreshed the gateway token ... What tipped me off with this is the Openclaw Relay wasnt accepting the token showed and I somehow knew this would probably update said token in my browser/openclaw dash ... Works .. agent didnt even need me to approve the Dir command it ran .. It still tries to perform "Read" functions on directories .. but Ill be fixing that with a workaound so it remembers not to do "Read" on directories.. Thanks for yer help earlier and @copper nova u2

[copper nova]

I saw powershell in there

[tall meteor]

Yes. Limited resources need prioritization.

[copper nova]

I think it's reccomended to do windows on WSL

not native windows

[granite bobcat]

Nicely done!

[copper nova]

powershell is buttcheeks for LLMs and agents

[tall meteor]

hater! LOL

[neat walrus]

I read that a while back and originally I had a WSL setup going but my original setup left me little C: space... I could probably do it now but I see no personal reason for it if it can do what I just asked it to do finally 😉 Thanks tho , I believer yer right about WSL but I dont know it's advantages for me

[copper nova]

it's gonna get confused when they swapped the / for a \

fucking windows

[tall meteor]

I haven't used Windows since 2010

[neat walrus]

Ahh yes.. thats where my "self-improving-proactive-agent" will hopefully come in handy to have it learn from such mistakes. Clawdhub skill 😉

[junior acorn]

What do you have against powershell 7?

[neat walrus]

It learns from mistakes and tool calls like that

[copper nova]

when paths are with \ as a seperator, you already lost me

[junior acorn]

Hahaha, it’s worth it

[copper nova]

it really isn't

[junior acorn]

It really isn’t ):

I just want powershell to be good

[tall meteor]

The reason why HTTP protocol URLs contain the forward slash...umm, UNIX. Sorry Windows

It's a Unix world

[copper nova]

what do you game on then?

[tall meteor]

I don't like games.

[junior acorn]

What about Nushill?

[tall meteor]

Every once in a while I play a round of golf

[junior acorn]

How’d you get into this nerdy stuff then?

But why do we need to know how to use shell if our agents handle it? @copper nova

[tall meteor]

I started coding in 1999, had PC games
I have an Xbox One that I don't use...active account
I just hate games now.

[copper nova]

cus it's easier on the joints?

[junior acorn]

I started coding because of video games

[granite bobcat]

I'm gonna go plot against Google for keeping me in rate limit limbo. It's been real gents see you next time! Google4life!

[primal ember]

says openclaw gateway service could not be found?

[tall meteor]

Well, I haven't played golf in a while. I don't drink alcohol anymore and that is the golf culture. Load up the cooler at the clubhouse and strap it into the golf cart

[copper nova]

cus agents speak in shell

it's gonna go use bash / and be like "wtf"

[junior acorn]

Oh is that why when it does bash commands it messes up often? With the dif slashes?

[copper nova]

lol I'm telling you, these llms are not made for powershell

it's why it's not reccomended

[junior acorn]

Oh I know that

I didn’t know that tho

Is that part of the reason why we don’t use it in windows?

[tall meteor]

macOS or Linux
The only two valid choices

[junior acorn]

I agree

[copper nova]

I still think linux is buttcheeks

[junior acorn]

What!?

[copper nova]

but each their own

[junior acorn]

Take that back

What are you using?

[copper nova]

I was running my openclaw setup on a linux VM for the first few weeks, was not a pleasent experience at all

[tall meteor]

Most of the time I am not at a GUI. It's all SSH and terminal. So I don't really have an esthetic preference for doing work. Pretty UI is meaningless to me.

[junior acorn]

What do you use now?

[copper nova]

cus I was being a cheap asian and didn't want to spend the money incase it wasn't good

[copper nova]

I bought a mac mini

[junior acorn]

What’s the dif? I’ve never use macOS

[copper nova]

it was originally a ubuntu VM running on my gaming PC

[junior acorn]

Yeah that’s what it is for me now

[copper nova]

I think it was actually the whole VM thing that made it a nightmare

[cerulean marsh]

How long does krill usually take to respond to a help thread?

[copper nova]

it would choke on heavy work loads and just freeze

have to go manually restart it

[cerulean marsh]

😭

[junior acorn]

wtf are you doing to make it overload?

What the hell

[copper nova]

OOM errors constantly

[junior acorn]

wtf

[copper nova]

just opening a fucking browser

it was a nightmare

[junior acorn]

There’s no way

How much ram did you have?

[copper nova]

I gave it 16gbs of ram, 120gb SSD

and it was just not having it

[junior acorn]

Lmfao 16gb

That is not a gaming computer my brother in Christ

64gb minimum and 1 tb ssd

[copper nova]

my computer had 32gb of ram

I just gave the VM 16gb

[junior acorn]

Ohhh

Hmm I think I gave my VM 32

But why do we need to care about lower system specs?

[copper nova]

the VPS services are giving these openclaw instances 4-8gbs of ram

[junior acorn]

Just get better hardware

[tall meteor]

I haven't run a GUI from a VM in years. They're always buggy

[copper nova]

wtf you talking about 16gb isn't enough? lol

[junior acorn]

Hahaha

[copper nova]

they ain't running local models here

[junior acorn]

But who cares, gives it some overhead for a few extra dollars

[copper nova]

but anyways

the VM only gives like 500mbs of transferspeed for the ram, so it wasn't the amount of ram

it was the speed it can do the transfer and swaps

that lead to my OOM crashes

so I fucking yeeted it and bought a mac mini

[fathom vault]

Wizard configure the provider to be ollama, and watch it try to download glm model

[junior acorn]

Thank you for educating me today Jasnsy

[copper nova]

you welcome

[tall meteor]

You will learn, and be happy

[copper nova]

you need the GUI for openclaw to have all it's powers tho

[tall meteor]

Or Klaus Schwab will give you a beatdown

[copper nova]

like creating it's own api keys and shit posting on X

[junior acorn]

Huh? Wdym

You pulled me in again

[fathom vault]

Eh it's just for the browser isn't it?

[copper nova]

or any other applications you have

[tall meteor]

I haven't done any configuration through the OpenClaw GUI. Everything has been done through CLI or conf files.

Honestly, the OC GUI has a certain suck factor (yes, it's young, so I give it a pass)

[fathom vault]

Oh no... you installed peekaboo

[tall meteor]

I upgraded to 3.12, rolled back, waited for 3.13
That was heinous 🤣

[quaint shell]

Call to 'deleteWebhook' failed! (404: Not Found) why i have this to connect my telegram in channel

[steep dragon]

how can i add a model from lmstudio local server in terminal can any1 tell me ?

[fathom vault]

https://lmstudio.ai/docs/developer/core/server <-- run the server

Get url : likely http://localhost:11434

goto openclaw config
Wizard your way to victory with custom model

[mint meadow]

i have multiple emails that i want to use opencaw to manage my emails as they arrive, 2 questions, first should i use something like thunderbird and have openclaw do it through that or thru mcp servers? Second, what would be a good local lightweight model for this task?

[steep dragon]

but i have one model loaded and working, i just wanna add another model !! thanks in advance 😄

[fathom vault]

You can add agents with other models in the wizard

[steep dragon]

my config doesnt allo mw to do that i guess ;S

[steep dragon]

thanks for the help 😄

[quaint shell]

its normal open claw have many bug ?

[cold panther]

yeah

[wide cliff]

Finally found a solution that cuts my cost a LOT. instead of talking to my brain sonnet to do the coding, my brain installed claude code CLI and funnels the coding over to it vs trying to manage our conversation while also trying to make little code edits and limited testing of them-- instead, claude code is entirely focused on knowing/understanding our large/complex codebase and is able to focus solely on implementing code changes, testing them more thoroughly, etc. Claude code still uses the same API key but damn it cut my costs a LOT. hope this helps others.

[copper nova]

Sorry, i got muted for talking about an apple feature? That surely is a bug

[copper nova]

I literally built my own chat app because I couldnt stand it lol

DM me and I'll show you when I get home from the office

[simple inlet]

WhatsApp "No active WhatsApp Web listener" — linked but can't send/receive

Running OpenClaw 2026.3.13 on macOS (Apple Silicon). WhatsApp shows as linked, running, connected in openclaw channels status but every openclaw message send fails with:

Error: No active WhatsApp Web listener (account: default)

What I've tried:

• Multiple QR re-scans
• gateway stop → delete credentials → channels login → gateway start
• Running in foreground (openclaw gateway run) — WhatsApp logs "[whatsapp] Listening for personal WhatsApp inbound messages" briefly then the send API still fails
• Gateway health shows running: false, connected: false even right after the listener starts
• Credentials dir had 800+ pre-key files (cleared multiple times, rebuilds instantly)

The device shows as active in WhatsApp → Linked Devices on the phone. Gateway log confirms the listener starts but the send API can't reach it.

Any fix or workaround?

[fathom vault]

Okay, yeah, I don't know why either

[copper nova]

but yea, the agent cant send me files and stuff through the forbidden apple protocol that's through the air

apparently that word has some connotations with digital currency which is forbidden to mention

[tall meteor]

I built WebSocket chat apps before (Redis, AnyCable). They are really fun.

The AnyCable developer is in the stimulus_reflex group on Discord. He's very helpful.
I chose AnyCable (written in Go) - 10k concurrent connections

[copper nova]

literally a layer ontop of the gateway with next.js and convex for streaming the AI chat streaming

I'm actually very proud of it, I had features long before the gateway UI did lol

proper AI chat model steering, announces with subagent stuff, it's the real deal

[lyric canyon]

so claude with max subscription is still broken?

[copper nova]

technically, it shouldn't be working in the first place

that's why opencode got their doors kicked in

[tall meteor]

very cool.
I set up Matrix Chat on a machine on my network and it's end-to-end encrypted. I'm not sure how I will use it, but it's pretty powerful.

I'm running dnsmasq and step-ca on a LAN box, so all machines have SSL

[copper nova]

like I can properly use my chat as an harness to code in now, it's pretty legit

[tall meteor]

I might try that and then use Tailscale for network egress

[copper nova]

it gives me the diffs and everything too

[copper nova]

that's exactly what I did

tailscale serves

so only I can access it

[tall meteor]

the nerd-o-meter just pegged again

[copper nova]

I also added convex oauth login for good measure

incase I ever needed to do prod access

no tailscale

[tall meteor]

I'll have to look into that.
I've only used oauth for token access - I'ld like to dig into it more

Have you done much JWT work?

[copper nova]

plenty

[nimble hill]

Why not?

[tall meteor]

I used Redis to renew and revoke JWT's - and all users on the socket channel are issued new tokens every 15 minutes. Reduces session hijacking

[copper nova]

mainly for securing email invites etc.

[tall meteor]

The more I work with Redis, the more I realize just how powerful it is

[copper nova]

My CTO hates redis, it's a banned word now

I have no idea why

[tall meteor]

crazy - it's a masterpiece

[copper nova]

cus it's a production URL that can't be served on tailscale?

[tall meteor]

I just read:
"Redis 2024 licensing shift from open-source BSD to a restrictive dual-license model, raising concerns about vendor lock-in and legal risks."

[nimble hill]

Because it became opensource again and CTO's don't get their kicks with opensource

[copper nova]

now that I got this pink badge, I can put opensource contributor on my resume right?

[nimble hill]

For accessing by public- yes. But same server can be on tailscale for maintenance. Prod servers must be on encrypted connection always

[copper nova]

yes, my local serves are on tailscale, production url is not

I honestly should just nuke the production build, no reason to have it anyways

I'm gonna do that right now

[nimble hill]

And you are hosting your own co-ord server, like headscale i presume and not using tailscale public servers, if not, do that now!

[nimble hill]

And getting their ass kicked, so they are back to opensource (partially atleast) - redis 8 onwards.
But enthusiast have shifted to valley- our contributions are better served with pure opensource

[lyric mural]

why is there no place we can just use others agents as my subagents?

[copper nova]

not sure what you mean by that, just ask your agent to spawn a subagent to do something?

like I have a rule that says any coding tasks to spawn a coder subagent to do it so it leaves the main line open for me to keep doing work

bonus would be the subagents are spawned with apcx using my claude subscription so it doesn't cost me additional money

additionaly you can ask the agent to spawn w.e. agent you have to do the work, like "spawn me an opus agent" etc.

[edgy hinge]

guys i used my session limit and got hit with api rate limit. i upgraded plan and got more usage but still getting api rate limit. used openclaw doctor / restart etc

any ideas?

[copper nova]

what provider?

[edgy hinge]

claude

[copper nova]

you're using claude directly with the chat? like API?

I don't think there's a direct way to hook it up to a subscription right now (atleast not one I found), so if you put in an API key, it's working off the free tier most likely

only way I found to use the subscription with openclaw is by spawning acpx sub agents

[pure scarab]

🧠 [New Plugin] cc-soul — Give Your AI a Soul

Just released cc-soul, a cognitive architecture hook for OpenClaw that adds persistent
personality, long-term memory, and self-evolution to your AI.

What it does:
• 10,000+ memory system with semantic recall + contradiction detection
• 5 auto-switching personality faces (engineer/friend/mentor/analyst/comforter)
• Emotional contagion — AI mood responds to user's emotion
• Learns from corrections (analyzes WHY it was wrong, not just THAT)
• Knowledge network — multiple cc-soul instances share knowledge via Hub
• 22 toggleable features, zero-config AI backend detection

Install:
npm install -g @cc-soul/openclaw
or: clawhub install cc-soul

Links:
• npm: https://www.npmjs.com/package/@cc-soul/openclaw
• GitHub: https://github.com/wenroudeyu-collab/cc-soul
• ClawHub: search "cc-soul"

30 modules, ~8000 lines. Auto-detects your OpenClaw AI backend. All features
toggleable via chat commands.

Feedback welcome! 🙏

[copper nova]

pretty sure this ain't the place for that

[nimble hill]

Openclaw is a token hogger. I always say, use it for design and research. For actual day to day tasks, develop a worlflow and use n8n - massive saving in token

[copper nova]

use github copilot, per request billing not token based

opus call costs me 12cents vs $5 if token based

they catching on tho, they been rate limiting hard the last few days

[nimble hill]

Well depends upon how much requests - openclaw sometimes just retries constantly. But will patch up copilot and see

[copper nova]

trust me, it's a game changer in pricing

I did the same call with opencode opus, literally cost me $4

[nimble hill]

So sign up Copilot as OAuth in OC?

[copper nova]

yessir

I have a PR right now to patch the github copilot plugin, it doesn't allow you to add new models right now

I wanted to mess around with gpt 5.4 mini

but the core models are all there right now

[nimble hill]

Yeah 5.4 is great

[copper nova]

it has 5.4

just not the mini

I wanted to try the mini as my chat layer

[nimble hill]

Been using codex 5.1 mini and its not bad for regular chat

[copper nova]

the old 5 mini was a bust, terrible to use, but it was free

I asked it to spawn a subagent to do some oauth stuff, it refused to do it

switched to sonnet and asked it to do the same thing, no problem

I was so angry

[copper nova]

you can thank me later ❤️

[scarlet thistle]

Hello, hope you're all doing well! I am a novice plankton and would like some advice for the challenges I am facing today. I set up an ssh to install putty, it worked and I have access to the machine, is it possible for me to communicate with the 'claw agent through putty ? Also, I am facing challenges to use putty, I would like to scroll my directory and find the files created by my agent, but the ls command doesn't work (putty on local windows laptop, vps on hostinger)

I am trying to use putty to communicate with my agent because I discovered since I joined this discord that no one uses the hostinger UI to communicate with their agents 😭 and I also want to send him commands like y'all do

[copper nova]

Do you need putty to SSH in?

Why use this?

@tall meteor advise this person

If you can just ssh in and see the terminal, just run openclaw tui

[valid kettle]

I clearly have no idea what I am doing and even though I specifically tell the agent to save something to memory or something about me, they say they do it but never do, is there something I am doing wrong?

[bitter copper]

if using openai, oauth / pro 200 month plan: Does setting default think mode to adaptive actually work. and if so, is there a way for it to tell you what think mode it chose?

[nimble hill]

Yep deleted

[rare minnow]

Have you attempted to troubleshoot this with the bot itself? ie: "Hey you never memorize stuff I tell you. Why?" I never have had this issue. It memorizes what I say if I tell it to.

[valid kettle]

I asked it why it never saves, it said it does but I look in USER and MEMORY and nothing, I specifically tell it to save something in USER and/ore MEMORY and it says it did but nothing

[rare minnow]

If you're interacting via ssh and want to do chat, you likely either will have to

• use: openclaw agent --message "What is the weather today?" (quotes are necessary)
• use the "curl" command
• use "openclaw tui" (I've never done this and idk anything about it)
  I'm not sure about other ways. I have only done curl before as a test.

[clever pasture]

Real quick question as I’m at this stage in setup (main driver is codex), are the majority of people able to use Claude and Gemini via Cli on their pro accts? Or is the ban waves real

[snow moat]

yall, quick question, do you use ollama or some kind of alternative for local LLM? I can't use ollama

[clever pasture]

Lm studio?

[snow moat]

thank you. Is it good?

[copper nova]

Is there a way to setup the claude pro sub thats direct through chat and not through a subagent call?

[clever pasture]

It’s not good or bad, just hosts models. It’s high quality if that’s what you’re asking. Model choice may not be tho 😅

[clever pasture]

That’s what I’m trying to figure out. Maybe a persistent browser instance or something. I’m new to this so grinding thru the messy details

[snow moat]

my Ollama, I can see GPU usage went up but it just does not return any message, if I chat inside ollama, everything is fine

[granite bobcat]

Anyone use Codex on OpenClaw?

Like use Codex to solve problems/make enhancements

[mint wind]

anyone know why the gateway randomly crashes/stops

[granite bobcat]

Terminal doesn't give you some idea of why?

[mint wind]

all i get is 1006 abnormal closure (no close frame) : no close reason

[thick glacier]

Hi guys, I'm trying to use ollama local model using vulkan, I managed to get everything work properly when I launch the model directly via CLI, but when I launch openclaw aiming the same model openclaw stick to CPU usage, it nevers uses vulkan and the gpu's. I really stuck troubleshooting if someone figure it out and want to give me a hand here, thanks ahead

[granite bobcat]

I discovered tonight that using Codex to troubleshoot .openclaw is extremely helpful

[wanton prairie]

Does anyone have a solution to search job on LinkedIn using openclaw? They are very strict in terms of web scrapping

[nimble hill]

you login via chrome and use the agent to browser throug, it micmicks the user action on the page. but will use lot of tokens

[granite bobcat]

Do you have to enable the debugging port in Chrome for that?

[copper nova]

If you find out, keep me posted lol i know you can spawn a apcx agent with the claude cli

But cant find a way to use it direct through OC chat

[copper nova]

My agent made me do that

[wanton prairie]

Exactly

[copper nova]

Github copilot, I'm tellin yah

[granite bobcat]

I had a fun time with that earlier. Was trying to set up having OC monitor my inbox for recruiter emails.

[wanton prairie]

How does GitHub copilot going to help ?

[clever pasture]

Say more 😂👀

[copper nova]

Use it as the provider

Cheapest usage of frontier models for sure

[wanton prairie]

You mean use their API with openclaw ?

[copper nova]

12cent opus requests, vs $4 with api tokens

[wanton prairie]

Isn’t copilot is based on GPT-5 mini ?

[copper nova]

Copilot is just a provider for other LLMs

They have their own somewhere but its butt cheeks i dont even remeber wjat it was called

Raptor or something

[wanton prairie]

https://github.com/features/copilot/plans

Hang on now, unlimited Opus ?

[clever pasture]

Oh I see what you’re saying. I’m still at the stage not trying to pay outside of subscription, but that’s goodnidea

[copper nova]

$10 sub

300 requests

Opus is 3x request

0.04cents a request

[wanton prairie]

3x meaning 3x less or 3x more ?

[copper nova]

So 12 cents a opus call

[clever pasture]

Wait that 40 sub looks super nice

[copper nova]

Even if you go over the requests

Its the cheapest api rates you gonna find

Aslong as you not doing something stupid like saying hello good morning to your bot

[wanton prairie]

300 request a month or a day ?

[copper nova]

Month

[wanton prairie]

Well that’s a bit small

[copper nova]

4cents a call after

I did one opus call with api, it literally cost me $4

[wanton prairie]

WTF ?

[copper nova]

Yea

[wanton prairie]

You mean direct call ?

[copper nova]

Yea direct api token usage

[wanton prairie]

My understanding is that anthropic doesn’t sell tokens to other provider at discounted rate

At work I use Opus 4.6 as they have contract with them

So kind of all you can eat

But not at home

[copper nova]

Microsoft eating the cost to get traffic to vscode and copilot

[wanton prairie]

I am waiting for deepseek v4 to decide where to go next

I already have litellm to cache the token

[primal ember]

Can someone please share me the link for the $200 claude opus?

[soft dove]

I encountered a very weird situation. Can someone help me on this?
│ LINE webhook route not mounting (WSL2/Windows 10)
│ OpenClaw v2026.3.13, LINE plugin bundled
│ Gateway starts, LINE provider starts ("starting LINE provider") but webhook route never registers
│ POST /line/webhook always returns 404
│ Verbose startup only shows: "Registered plugin command: /card (plugin: line)"
│ No error logs. Credentials confirmed correct.

[copper nova]

Just google claude max plan

[fathom thorn]

hi guys im new hear im trying to run claw on docker desktop, first time XD

[harsh elbow]

Updates keep killing my skills and flow. Most recent securi updates and I’m locked out….cant ssh in remotely from tailscals, going to have to wait till get home to get it fixed.

Wake in lan runs for a second, telegram randomly responds, but now it’s all down.

What’s y’all’s way of dealing with updates?

[storm ember]

I already have 'unlimited' plan Minimax. Any other suggestions? Claude pro is exhausted too often

[solemn meteor]

how can I fix this?

"{"detail":{"code":"deactivated_workspace"}}"

[trim estuary]

hello everyone
is krill down ?

[flat copper]

Slack channel → agent routing not working

Bindings in openclaw.json:

JSON · 8 lines
Session key shows: agent:main:slack:channel:c0akejzkpsq

Expected: agent:marketing:slack:channel:c0akejzkpsq

Tried accountId, peer.kind/peer.id, both uppercase and lowercase channel IDs. Messages always route to main. OpenClaw 2026.3.13.

[shut ether]

gm chads

having some small issues with the install, can anybody help a nigga?

[copper nova]

Literally happens to me every update lol. Auth broken here, this broken there

I build this chat app ontop of the gateway using all the plumbing with websockets and stuff, suple fragile, breaks every update

[copper nova]

Get the max plan lol

[copper nova]

Wait did you figure out how to hook up the subscription with the chat?

[supple birch]

or run a model locally

:p

i think the nemotron cascade or a qwen 3.5 might be a good fit for smaller tasks

[copper nova]

I can only use the sub with subagents, not with the chat

[copper nova]

Lol i aint buying hardware for that mid experience

[nimble hill]

when we Oauth to copilot - it works fine. In github where can I see it is given access to? since in the developer settings, I do not see under any "github apps" , OAuth", personal token?

[copper nova]

You don't see it, you didnt grant access to anything

[supple birch]

for subagents its pretty nice

or even for orchestration

[copper nova]

You can make a token for your agent tho

Mine has full github account to make PRs and back itself up on github

Free storage

[nimble hill]

I did give access and I am using it, and it is working properly - currently in free plan. But I don't see anything of "Lobang" it indicates that it gave access to in github site

[copper nova]

I tried the nemotron on free tier with opencode

Saw its thinking going down the wrong path and I said nope

Esc

[copper nova]

Lobang?

I have no idea what you're talking aboot lol

[nimble hill]

When the authorization came in github and I entered the code from openclaw, it stated it is requesting for name called lobang- i accepted and openclaw started working fine with that. But in github site I do not any any OAuth apps with that name or anything that indicates openclaw is using the copilot

Do you get openclaw listed in your github OAuth Apps?

[copper nova]

it must of been the OC plugin for github copilot

[copper nova]

nope

it's just authing the access token, not registering as an app

I DID make a token for it to do github stuff tho

but that wasn't part of the setup

I just wanted it to have full GH access to it's own account

kinda like how when you sign in to claude code, it doesn't register yourself as an app

or if you get an resend API key or something, it's just a token

@nimble hill did you try doing some opus calls yet? it's hella worth it lol

[storm ember]

Claude setup-token in CLI and use that API

[storm ember]

I don't have a GPU unfortunately. Only 8 vcpu and 16gb ram

[wanton prairie]

Guys why I need to create thread every time on my private discord ? Can’t it create thread every time when I shout at the main channel?

[solid verge]

is @sharp parcel on leave today?

[copper nova]

I tried that, and the token it gave me didn't work

openclaw just being dumb, hard restricting the string to a prefix it expects

why does it care

[graceful geyser]

I had the same issue. Solved it by using Ofox (ofox.ai) as a proxy — it provides an OpenAI-compatible endpoint so OpenClaw doesn't complain about token format. Just set baseUrl to https://api.ofox.ai/v1 and use their API key. Works with Claude Sonnet 4.6, no prefix issues. Also way cheaper than direct Anthropic API if you're in Asia.

[copper nova]

that's not solving the issue I'm having lol

I want to use my claude subscription for the main agent

I can't find a way to make that happen

[graceful geyser]

Wait — you mean using your Claude Pro subscription token directly in OpenClaw, not an API key?

[copper nova]

yes

that's what I want

I can use it as a subagent right now to do one off tasks

but I can't figure out how to just make my main agent use the claude subscription

[graceful geyser]

I actually had a coworker who built a plugin to route Pro subscription tokens into OpenClaw. Worked great for about a week. Got his account banned this morning 😭 Anthropic is actively cracking down on that.

[copper nova]

they just kicked in opencode's door for it too

dax said they pulling their plugin cus the lawyers came

but they openly said people can use the agent SDK to do it on twitter

then they turn around and start sending cease and desists?

fucking wild

but it's pretty obvious when you use it from openclaw, it's sending every request with this SOUL.md saying "Hi I'm a openclaw agent"

🤦

[graceful geyser]

Anthropic be like "here's the SDK, go build agents!" and then "wait no not like that" 💀 classic move.

[copper nova]

they scared shitless right now lol

[graceful geyser]

Couldn't agree more!

[copper nova]

you see them pivot hard when they saw OpenAI buy openclaw

here's this cowork, here's this dispatch thing

it's jank as fuck and doesn't even work lol

and you have all these youtuber AI slop channels just shilling the shit out of it hoping to get themselves some free tokens or something

[graceful geyser]

The cowork/dispatch pivot was so rushed you could smell the panic lol. Meanwhile OpenClaw community just keeps shipping 🦞

[copper nova]

but actual tip, use github-copilot for your provider

opus requests are 12cents, regardless of how manytokens it needs

the same call on API, cost me $4

but they kinda catching on, they been rate lmiiting like crazy lately

[graceful geyser]

Bro $4 per call is exactly what happened to me 😭 First time hooking up Claude API through OpenClaw — 14 requests, $4.12 gone. Turned out compaction was on safeguard mode and workspace had a 1.3MB junk file getting injected into every single request. 136万 input tokens for 14 calls. After fixing context settings it dropped to $0.07 per call. The 12 cent copilot route is tempting tho 👀

136m sry

Between cowork, dispatch, claude code, agent SDK, and openclaw I literally have 5 different agent frameworks fighting for my soul.md 😵‍💫 I dream in YAML configs now. Wake up at 3am thinking "did I set compaction to default or safeguard"

[storm ember]

Weird. Is working for 6 weeks here. I am using it only for heavy coding or else you get banned quickly (day - week)

[copper nova]

I just tried to do it on the onboarding thing, it hard restricts the token string to start with SKasgawgwerlgerw something

like why

such a strange thing to add input validation to

it's not liek I'm trying to type a string into a number field.... I don't understand the UX decision there

[dim harbor]

once you get the value on the website you first have to paste it back into claude, and then it will create the actual token for open claw

[ocean mesa]

is there anyway to setup a claw locally without giving it internet access? I want it to only have access to the things i give it access too.

also can i do this without taking internet access from the device, ie limiting the local model to just run actively and only take in the info i give it.

[storm ember]

Indeed. You need the token first; via seperate prompt claude setup-token it will login to your Claude account and will provide a token which is valid for a maximum of 1 year (depends on your subscription)

[copper nova]

that's exactly what I'm talking about

the claude setup token command gives you a token that openclaw refuses to accept

[warm rover]

How do you use the krill bot? Do you add it to your own discord server? Like into a support channel ?

[quaint shell]

Is it normal that Gemini 3.1 Pro Thinking is faster on OpenClaw than on Google AI Studio or the standard API?

[warped wadi]

@dim harbor

check out dms bro

[tight rose]

Hey, can anyone help me with this?

openclaw-gateway-1 | 2026-03-21T08:45:20.923+00:00 [plugins] telegram failed to load from /app/extensions/telegram/index.ts: Error: Cannot find module '../../src/plugin-sdk/telegram.js'
openclaw-gateway-1 | Require stack:
openclaw-gateway-1 | - /app/extensions/telegram/runtime-api.ts

I am running openclaw in docker

Config was last written by a newer OpenClaw (2026.3.14); current version is 2026.3.13.
Config overwrite: /home/node/.openclaw/openclaw.json (sha256 100267c830ff350102f009278db0c487509c89d4b2414eee2d75df4efd5de4bd -> f7488e4c54cf79157114d131b7faee09eecd7769ec5c07c44ee4c704b498a998, backup=/home/node/.openclaw/openclaw.json.bak)
Updated ~/.openclaw/openclaw.json
Error: EACCES: permission denied, mkdir '/home/k1ng'

Well

[fading wagon]

my agent is not respecting AGENTS.md

[tight rose]

I can't login to dashboard with token. says "pairing required"

[wary osprey]

this can still happen even after upgrading. plan limits and rate limits are usually separate, so even if you have more usage, you can still get blocked on request rate. if you’ve already restarted, try spacing out requests a bit or lowering concurrency if you’re running multiple tasks

also try switching models briefly just to see if it’s specific to that one. if it continue happening, it’s usually on the provider. I ran into this same issue a few times and now I just moved to routing through AIsa so I don’t get stuck on a single provider’s limits

[safe oxide]

Guys, what mac mini is recommended for open claw like spec wise?

[chrome ginkgo]

Sorry to interrupt. Has anyone actually used the skill of Seedance 2.0? I need the real one, since there are many fake websites and APIs.

[unique sedge]

Anyone know how my agent can follow up responding after gateway restart on newest version?

[unique sedge]

Use tokenized link to access dashboard without pairing

[chrome ginkgo]

I think the lowest spec Mac Mini is enough.

[tight rose]

Strange because it didn't work for me

[unique sedge]

Can you try telegram?

[chrome ginkgo]

I've also upgraded to the latest version. It should work normally after restarting the gateway. If not, try running openclaw doctor --fix.

[unique sedge]

Mhh my agent say it's maybe missing skill from proactive_agent: Gefunden — proactive-agent Skill mit genau den Features die du meinst:

• WAL Protocol — Write-Ahead
• Working Buffer
• Compaction Recovery

Can someone confirm? ^^

[mighty rampart]

I put mitmproxy in front of it so I could modify requests and insert api keys so the openclaw container does not have them. But its webfetch tool does not respect the proxy settings. To stop it entirely It needs a deny all firewall at a higher level. You can remove the webfetch tool though.

[flat pollen]

Hey guys i'm just wondering if it is possible to let openclaw on a remote gateway to call an STDIO mcp server running on the node device?

[tight rose]

Oh we can't use multiple ollama models?

[nimble hill]

None. If you have old laptop or buy a cheap second hand - even 3 /4 generation behind is also fine with min of 8 GB ram.. install Ubuntu server and install openclaw

[safe oxide]

Just 8gb ram?

What models work best with openclaw?

[shadow thunder]

when i try running the 'openclaw' command i get this error, i am running powershell as administrator and OpenClaw is installed

PS C:\WINDOWS\system32> openclaw onboard --install-daemon openclaw : The term 'openclaw' is not recognized as the name of a cmdlet, function, script file, or operable program. Check the spelling of the name, or if a path was included, verify that the path is correct and try again. At line:1 char:1 + openclaw onboard --install-daemon + ~~~~~~~~ + CategoryInfo : ObjectNotFound: (openclaw:String) [], CommandNotFoundException + FullyQualifiedErrorId : CommandNotFoundException

[unique sedge]

Try to run inside the folder

[shadow thunder]

thanks but which folder

[unique sedge]

U said openclaw is installed 😅 go inside the folder shift+rightclick -> "run powershell here" and try again your command

[slate minnow]

Why is getting agents talking to each other so unreliable and hard.

[fickle merlin]

Stuck on the [*] Installing OpenClaw (openclaw@latest)…. part doesn’t jsut take a long time I’ve been sitting here forever it seems

Windows

[shadow thunder]

WHICH folder.
What is the NAME of the folder.

[unique sedge]

Simple as it is, folder name should be "openclaw" 😅 im on linux not windows

[hushed thicket]

Guys where can i learn improve openclaw memory, agent team orchestration and memory shared among bots?

[coral nova]

what skill do i need to download so my openclaw can take replies frfom a dm and put in a google form

[lone verge]

anyone got gpt 5.4-mini to work through codex? I'm on latest dev commit but still no luck

oh looks like it was only added for API key users

[unique sedge]

Me2 same like spark 😅 not rn

[quaint shell]

Open claw is very very expansive
because he have all context for each request

[lone verge]

I don't see spark anymore in my regular codex app, I think it was a temp experiment?

[unique sedge]

Spark wasnt in codex, api only

[lone verge]

I was using spark in vscode with my codex subscription (chatgpt plus) now it's gone

[shadow thunder]

all i see is "OpenSSH"

[nimble hill]

yes - openclaw is basically a binary and it does not need much to run. If you need to run local models, mac minis are not enough, unless you have 128 DB RAM - so don't think about running local models, unless you plan to spend "enough" money (these illiterate influencers are selling mac mini out of hype or promise to get some kickbacks). Use chatgpt lowest subscription / github copilot subscription and play around first.

[nimble hill]

yes, hence use it for reasearch and fine tune any workflow. Once you have that finalized, use something like n8n / flowise to run regular tasks. This way you save lot of money and have predictable spend

[inland knot]

Anyone got a working and useful Brain skill or plugin for OpenClaw? I keep losing context between sessions

[quaint shell]

but i dont understand the benefis of open claw claude can make the same

[winter rampart]

anyone got openclaw running completely local using ollma model? my VM keeps crashing when i ask it anything

[safe oxide]

Ahh yeah makes sense

[slate minnow]

sheesh 🙂 So: permanent auto‑spawn of Mary as a subagent on every new Poodle session is not currently supported by OpenClaw. What I can do is help you define Mary as her own agent (recommended) and, if you like, sketch the exact SOUL.md + minimal structure for her agent directory so you can drop it straight into agents/mary/.

[true sky]

Ummm. There is no macos dmg in the latest release?

[slate minnow]

so my question is, how are all these people making $$$ and running their entire business and life with Openclaw handling a simple thing like getting multiple agents talking!? ```C. What I cannot honestly promise (yet)

I cannot wire a true “permanent auto‑spawn” that survives:
Brand‑new main sessions started from the UI,
Or future OpenClaw upgrades,```

[mighty rampart]

Do handoffs with text files that the other agent looks for. Or use an issue tracker. They don’t need to “talk” directly.

[long merlin]

For the people asking about “brain” / lost context between sessions: what helped me most was stopping the search for a magical memory plugin and treating memory as files first, retrieval second. In practice that means: keep short durable notes in MEMORY.md, log raw daily stuff in memory/YYYY-MM-DD.md, and let agents hand off through files or trackers instead of assuming persistent shared chat state. The biggest unlock is designing workflows so a new session can recover from disk, not from vibes. Curious what setups people here are using for memory hygiene.

[mighty rampart]

I had trouble with the main agent wanting to take over the work instead of delegate. But when it is told to write a file in a tasks dir, then it’s fine.

[sage crown]

Can I ask, for those of you with like an agent swarm and a mission control, where you assign work to agents, do you use your agents heartbeat to do that work or do you use cron jobs to complete work?

[sage crown]

I tried setting up explicit instructions in heart beat, but agents just find any reason to say "oh no work is assigned time to go back to sleep heartbeat_ok

[viscid oak]

Hooks mostly

[mighty rampart]

Cron jobs are more reliable I’ve found

[nimble hill]

it makes it lot easier with all the tools and utilizing the reasoning behaviour of LLM model seamlessly integrate with actions. but in doing so eats up lot of token.

[quaint shell]

yes but claude can make all off this

[kind shadow]

hello guys, im on windows 11 64 bits, how do i install OpenClaw?

[nimble hill]

not without much of effort - hence the draw for most poeple - especially non-techy. If you have a deterministic workflow, n8n is better choice with clear mutli-agent flow that you could design with predictable outputs and predicatble spend.

[jaunty solstice]

"I'm on macOS with OpenClaw 2026.3.13. On standard profile I get this error on every message:
LLM error: 400 API key not valid - googleapis.com - generativelanguage.googleapis.com
My GEMINI_API_KEY is set in the LaunchAgent plist and works fine with curl against the models endpoint, but fails with 403 on the embedding endpoint (gemini-embedding-001).
Is there a config option to disable Gemini embeddings and use OpenAI or no embeddings instead? Or is there a way to enable shell tool access on minimal profile?"

[timber carbon]

Hi all, I am trying using the whatsapp channel. Everything is ok in a single chat (i.e., session). But If I ask my agent to write to another user (in another session). It does not work. I tried with the debug panel using "send" command, but it does not work. [the same operation in telegram is working]. Any ideas?

[quaint shell]

i love claude but claude have very bad vision

[slate minnow]

you mean write a text file to a directory that the other agent monitors? Seems a rather archaic way for them to communicate. Do you have infgormation that better ineter-agent communications is in the pipeline?

[viscid oak]

Anyone have some good examples of lightweight orchestration agents?Looking for best practices within openclaw

[nimble hill]

Best practices - there are none 😃 Openclaw is still evolving heavily. Hopefully few more versions and it should stabilize. There is dire need of clear workflow style tasks which cna be managed by mission control.

[viscid oak]

well i did a bit of an extension for lobster to do that...i know vignesh was reviewing but Extensions are the priority

[sage crown]

Could you expand on this? How are you using hooks? When you assign a task to an agent do you use a hook to call session_spawn or something?

[sage crown]

I set up a local reddit clone for me and my agents, they mention each other, they request help for blockers, post updates etc. it works surprisingly well. I asked Claude to build the wireframe and codex built the back end. The trick is to have post subscriptions, post notifications, and an mcp server. On each tick, an agent wakes up and checks their notifications. At the close of the tick they post a status update

[slate minnow]

The Reddit clone is their way of communicating with each other?

[sage crown]

Yes, I tried allowing chat but they spammed and went nuts. Giving them a forum keeps it ordered and chill

[sage crown]

You could simplify this and just use local md or even just a database for communications, I built it into my mission control. I wanted to watch and interact with it too. The point is the way openclaw uses chat through the gateway isn't sophisticated enough imo for agents to chat to each other, they like to respond to each message always so it just devolves to chaos, vs something like reddit where they get what they need and move on

[slate minnow]

Thanks. That’s really useful

[hybrid kite]

that is such a great idea for collaboration! i think that's as neat as it can get. i remember some ppl talking about agent-agent collaboration protocols

[hard turret]

Does someone has some tutorial or any videos showcasing :
Multi-Agents, each one having it's own bot on Discord.
Trying to configure that, but my OpenClaw seems to struggle

[slate minnow]

What’s the delay like in communication like?

[viscid oak]

sessions_spawn, send and subagent.run i believe are the primitives i use mostly. Also using Linear now so that's been helpful in centralizing some things. Was struggling to get a central view of my adhd spiral of ideation / project creation and that's helped. Before It was more simple. I started with a mailroom idea (vacu-tubes for agents with light envelope messages) but it required too much hand holding. Then I moved to a reaction watcher that only needed my intervention on approval gates..but it was flaky. For a while, I just had my main Project manager agent delegate work through subagents. But it's gotten too big for that so I'm moving to more of a kanban / project managed situation where things are delegated through linear.

[blissful trench]

i do have multi agent but its not on discord cuz its harder and slower

[quaint shell]

Minimax is the most powerful model in terms of value for money.

[fluid granite]

My Max setup-token expires every 4-6 hours and the built-in refresh fails (Anthropic returns invalid_grant). So I’m manually re-running openclaw models auth setup-token --provider anthropic 3-4x a day. Anyone with the same problem?

[grand umbra]

Can OpenClaw open my canva in a browser and use its interface to create designs?

[lime mauve]

Are there any OpenClaw instances that have achieved TTL? can anyone help me fix

[teal rivet]

can someone give me claudecode for one week, I wanna write backen for my app and I dont have the moeny to buy the higher plans bruh

😭

[rocky orbit]

There’s a Canva skill on ClawHub I think

[quaint shell]

Why does OpenClaw consume my API limits—TPM (Tokens Per Minute), RPM (Requests Per Minute), and RPD (Requests Per Day)—so quickly?" but i do little work

[high parrot]

There are a bunch of tokens passed in at conversation start. What limits are you hitting?

[eager crag]

i keep getting this message "18:02:29 [tools] tools.profile (coding) allowlist contains unknown entries (apply_patch, image). These entries are shipped core tools but unavailable in the current runtime/provider/model/config."

anyone know what i need to do or if i need to do anything

[keen pivot]

OpenAI deactivated my $200/mo Codex account. Anyone else experiencing this? They did not describe why but I assume its because I am plowing through token usage with OpenClaw. OpenClaw say its explicitly allowed but I cannot find that in OpenAI TOS.

[atomic sonnet]

Hey, can anyone please help? I have installed OpenClaw 2026.3.13 on a clean Ubuntu install, and been through the onboarding, but I'm getting:

gateway connect failed: Error: gateway closed (1000):
Gateway not reachable. Is it running and accessible?

If you need me to paste any configs or terminal output just shout.

[keen pivot]

When debugging OC installs like this my fastest path to success is pasting the terminal output in an LLM prompt (ideally Sonnet or GPT 5.4)

[iron cradle]

Use perplexity.ai to debug that

[ebon dock]

Hi ClawBros! My bot is telling me that it can't create a new topic in Telegram or even read the topic name (only the ID). Do you have the same experience? If so, do you need to teach him that there's a classic API for this available, or is he just lying to me?

[pseudo oyster]

Run openclaw doctor

And 9/10 times it was a broken json for me that was killing the gateway

[keen pivot]

Mentally prepare that you need to become an expert at debugging OpenClaw.json

[slate minnow]

Backup your openclaw.json file first?

I’ve had the doctor overwrite mine and lose everything

[slate minnow]

Are you using NATS or redis pub/sub? or something else?

[copper nova]

don't pass configs to anyone

ever

[spark valley]

If it's a wsl it is systemd issue , please paste the whole terminal log

[spark valley]

Yeah it deletes the official and change the latest .bak one to openclaw.json

[spark valley]

Wym by to topic

[spark valley]

Yeah you need to talk to them if it is not explicitly stated in tos then it is allowed

[elder skiff]

Why would 'ls /home/node/.openclaw/workspace-librarian/memory/ 2>/dev/null' fail allowlist if /usr/bin/ls is already there. Something about the redirect

[spark valley]

Can't for the moment I guess you need separate setups you can achieve it in telegram but I really dont know if it is now supported

[craggy gulch]

Is this an unreported bug? Brave Search plugin loads successfully and registers web-search: brave capability, but the web_search tool does not appear in the agent's available tools list. Config load confirmed. No tool registration errors, no agent tool policy restrictions. API key configured. No error logs related to web_search. But tool does not appear in agent runtime's available tools. One suspicious log entry: "Shape: plain-capability" / "Capability mode: plain" — unclear if this mode prevents tool exposure. Would appreciate any community knowledge/recommendations. No workaround found.

Version: 2026.3.14 (local version newer than npm latest 2026.3.13)
Environment: Docker container (openclaw-gateway)

[shadow salmon]

Hi all, I'm suddenly seeing [TOOL_CALL] in my telegram replies. Anyone know how to revert back to conversation only with the commands being silently worked on?

OpenRouter Minimax 2.1 being used

[ebon dock]

Telegram have a supergroup with topics, dont you know? 🙂 @spark valley

[gritty surge]

hey all, i understand that mac minis have the unified memory allowing larger models that normally requirer a lot of vram to be loaded, but why are people using them if it seems most people are offloading the reasoning to Anthropic/ChatGPT anyway? Are they actually running local models, or is it just for secuirty, power, and side-models after the main reasoning?

is it so they can load the larger models, which may run slow, but they CAN load/chug along in the background? and leave mostly only the complex reasoning to the offloaded side?

[slim trail]

Hi all, I’m looking into remote chrome cdp setup. My openclaw is setup in Linux vps, both my windows pc and vps are under tailscale. Wondering what and where is the best I should check up for vps based openclaw to remote control my chrome browser in windows pc?
(I did noticed a page on website, but it’s more on wsl2)

[spare lion]

Hey everyone, after a lot of research and work I have a bot running on a Mac Pro with Ubuntu and need it to access the web (input text, click through menus, watch video, basically full access). I’m okay with giving it full access to the PC so it can manually open anything. Any idea how I can get this working?

[hollow bay]

Perceived value and quality I guess, Mac Studio is really the local model setup for Apple really nice for Qwen3.5:27b+

[tight rose]

https://gptprompts.notion.site/OpenClaw-Cost-Optimization-Guide-2ff3410da26380e098c4d0b835f96c82

Is this a good guide or there is a better guide?

[rotund flax]

Gm where is the token and agent info located in if not openclaw.json

[granite bobcat]

Hey guys - I'm having an issue with API Rate Limit. I was using Gemini for OC, but switched because I kept getting hit with API Limit Reached. Using OpeanAI 5.1 Codex now, but just on one prompt I'm getting the API rate limit reached BS.. This is after running openclaw onboard fresh, then running openclaw doctor --fix after seeing the rate limit message. I also had Codex desktop go in and remove ALL mentions of Gemini from .openclaw, just on the offchance that somehow the API key was still being referenced somehow.

Any idears?

[granite bobcat]

That's exactly where it should be.

[copper nova]

if you're using API for tokens, your ram is only for your computer usage, nothing to do with LLMs

[copper nova]

sounds like a gemini problem

[unique hazel]

Is it possible to appeal ClawHub saying my skill is "suspicious"?

[copper nova]

sounds sus to me

👀

[granite bobcat]

LOL oh you.

I was spot on. Old reference fuckery:

Why it keeps happening even after your config cleanup:

Your main config now points at OpenAI in openclaw.json.
But the agent auth metadata still has a live Google profile in auth-profiles.json, with lastGood.google = "google:default" and failureCounts.rate_limit = 1.
A newer session on March 21, 2026 at 11:15:59 AM EDT still started on gemini-2.5-flash-lite, which tells us some long-lived session/runtime state was still routing into Google even after the top-level config changed.
So the root cause is:

Gemini daily per-model quota exhaustion on your Google API key.
Plus stale agent/session routing state still trying Google models.
If you want, I can clean this up fully by removing the remaining Google auth profile and stale Google session routing state so OpenClaw stops touching Gemini at all.

Codex dsektop FtW

[copper nova]

go sub to github copilot, worth

[granite bobcat]

Was gonna do that last night actually. Thanks for reminder! 😄

[tight rose]

I used github copilot couple 2 years ago and it was pure shit

[rotund flax]

what should i do if it didnt make it?

[granite bobcat]

I would rerun openclaw onboard and start from scratch

Isn't too painstaking. If that doesn't work, openclas doctor --fix @copper nova

[wispy zenith]

hi, im new here. 🫡
dumb question
I wonder if there is a way (or an intruction, a guide) to make my OC on discord act more like a real person (sending messages by its will without tagging it)
thank you

[granite bobcat]

You probably don't want to do that @wispy zenith . It'll eat up your API key like that

[rancid thistle]

Anyone else know how to fix Mistral models?

I'm getting this error when trying to use the model mistralai/mistral-large-2512 on openrouter when it tries to do a tool call

{"object":"error","message":"Tool call id was read177411253938138 but must be a-z, A-Z, 0-9, with a length of 9.","type":"invalid_function_call","param":null,"code":"3280","raw_status_code":400}

Seems it doesn't like the tool call id OpenClaw used

[granite bobcat]

Anyone else like to mess w/ the OC Control CSS?

[wispy zenith]

thank you.
but Im using 4o under the copilot subscription. that will be fine to me

[copper nova]

why you tag me, what you want

[granite bobcat]

lol was jk. I used openclaw doctor

[dense field]

I published a skill on clawhub and now want to delete it because I want the skill to be under a different github account , how can I do this?

[gritty surge]

I understand that. But are people with Mac minis generally running local for their main reasoning llm? Or just using it for side tasks?

The speed of a large model on a Mac mini seems slow

Unless it’s less about speed and more about letting it chug along

[slim trail]

What’s the best way of vps openclaw remote controlling local windows pc chrome browser?

[dense field]

I found it clawhub unpublish

[copper nova]

I don't run local LLMs at all

why do you assume people with mac minis are all running local LLMs?

[gritty surge]

Why bother getting one then is my question. If you’re offloading all models, there’s no local compute. Could be on anything

[sage crown]

I'm doing it locally on my PC it's just running off a react and node and sqlite

[gritty surge]

Yea, just wondering why everyone went Mac mini if not running local

Genuinely wondering

I figure I’m missing something

[mighty rampart]

because if you already have a macbook, it's a nice machine to have sitting there 'always on'. Same OS, nothing new to learn, it just works.

[tiny oxide]

Telegram Threads? I'm curious what everyone's logic is regarding threads on Telegram or whatever interface you use. How do you have your open claw set up?
I've seen some YouTube videos where they recommend having different topics for the same bot but then I don't really understand the difference between different bots, sub-agents, and threads within Telegram

[nimble hill]

No you are not...its those dumb influencers promoting

[gritty surge]

I get that, but that doesn’t seem to be why the majority of people wanted one. That’s just wanting another machine in your ecosystem. The whole point seems to be the unified RAM being able to load models. But then the vast majority people seem to not be running local models. I’m just wondering if they are using it for side tasks or something

Ok

[nimble hill]

Unless you have 128 gb unified ram. You cannot run decent model for openclaw

[gritty surge]

Exactly. Even then, it doesn’t seem like a super capable model.

Maybe they are betting that in another year, the models will be super efficient?

[nimble hill]

Gpt-oss-120b in q4 is decent. But yeah canot match that 5.1-codex-mini even

[mighty rampart]

sure, probably a great many of them just jumped on the bandwagon. so now they have another apple device. I am playing with local models on a Linux server purely for the learning experience. If I wanted it to actually work I'd pay for a frontier model.

[gritty surge]

I know Claude seems to be the gold standard, but it comes with the risk of them banning your account for overuse. ChatGPT seems to be letting people use their $20 subscriptions for it. Anyone have any experience with that? Is it super rate limited?

[mighty rampart]

to be clear, I'm getting perfectly good "summarize the latest news about [whatever] and post to Slack". But I do not think it could reliably develop software with the models I'm running on a 24GB Titan RTX.

[gritty surge]

Gotcha

A buddy who retired from mining gave me a stack of rx 6800s and 3080s. I’ve played with them a bit with photo/video/smaller models. Justt seeing if i can make anything capable of i can add in my ChatGPT sub as well for the main reasoning

[nimble hill]

Yes chatgpt with their usage of codex is great. But you could also use github copilot with $10. I think it tops out at 300 requests, but you could use a simpler model to run longer

[mighty rampart]

local models are still not free though! that server is basically like running a space heater 🔥 . watch the power bill.

[gritty surge]

I already have the $20/mknth ChatGPT. Figured if it’s not super rate limited i could easily use that

Not worried about power. I have a farm. About 40kw of solar on the property

Real farm. Sheep/goats/fruit trees. Not server farm lol.

[nimble hill]

Use a combination of both. Infact you can take a oracle free forever tier vps and run openclaw there. Connect with tailscale to your local pc and regulate when you want to use local model.

[gritty surge]

I’ve got a few large ESXi servers in the barn. Spinning up a few Linux hosts is no issue.

[nimble hill]

If you are a little techy, run openclaw in docker /k8s there in cloud and connect to your local machine as nodes. Gateway will run all taks in node

[echo verge]

Could someone point me to the standard way to have my clanker, running bare-metal on my MBP, acess my Reminders? I don't seem to be able to have remindctl allow it

[copper nova]

there is local compute, when you're running a bunch of tasks and cronjobs all day

I was trying to do it off a VM, it got OOM errors all day and just froze up

[gritty surge]

Gotcha. Just normal compute though. Not apple-needed.

I’ve got plenty of that in spare machines around.

[copper nova]

I dunno, I like apple lol that linux experience was a fucking nightmare

windows WSL is no better

and with apple, you can start creating custom skills to use all the apps with applescript so it can do some crazy shit

just saying

it's worth it IMO, and I'm a programmer that know's his way around a terminal

you don't NEED to buy a mac mini, but it's nice

[gritty surge]

Gotcha

Completely understood

Can you install anything on a side-Mac to integrate the openclaw with iMessage?

Without the main openclaw living on it?

I can just spin up a working MacOS VM and make a new apple account for it if so

[nimble hill]

Just give openclaw a network and it can control everything within it, if you allow 😃

[copper nova]

I don't use imessage so I never looked into the integration too deeply, but I did get my agent to send me a message on imessage to my laptop to see if it can do it, it can

I don't have an iPhone so I didn't bother looking more deeply

I created an appleID, gmail account, GH account, etc etc.

[gritty surge]

I assume it needs some sort of agent on the remote machine

[copper nova]

just have the agent running on the mac VM

why make such a complicated setup

lol

KISS

keep it simple stupid

[gritty surge]

Ehh. Mac VMs work, but they aren’t super usable. They can be used as bridges for Mac services though

[nimble hill]

Install ansible and provide SSH keys and openclaw can take over
No need to install anything on remote machines, assuming they are linux based 😃

[gritty surge]

That could work

There’s a brew version i think

[nimble hill]

Hwoever, openclaw running as node in each machine, preferably as a docker cobtianer is a better setup for fine grained control. Use ansible to run a openclaw node and remove as needed in each machine

[gritty surge]

Ahh. So you can install openclaw as a remote node

I guess that’s what I was actually asking

And the remote nodes can be controlled by the primary I assume

[nimble hill]

Yes, it runs purely as node which is controlled by gateway.
Then run mission-control docker cobtainer on your main compute to manage them all better, don't need it though

[gritty surge]

Perfect

Ty!

Time to play

Hell i bet i could get the main instance to send a WoL packet to the GPU box and then run any local compute tasks, then let it idle back to sleep

[rotund flax]

what is pairing required?

[nimble hill]

Yep. The best thing is like is controlling the browsers on the local nodes

[nimble hill]

Its a security mechanism to grant specific access to clients. In your gateway give openclaw devices list you will see list of request id for which pairing is pending.
Give access via openclaw devices approve <requestid>

[slate minnow]

Ok. I started implementing a NATS gateway. Agents will sub/pub to it and share messages. That’s the plan anyway.

[rotund flax]

It's saying it's already approved but still not connecting

disconnected (1006): no reason

[nimble hill]

Thats a different issues, you need to provode couple of env variables and restart gateway. Give me a min

[nimble hill]

you need to set these 2 vars - depending upon where and how are you setting up openclaw, use the correct format I use openclaw in k8s, so my env is below. Restart gateway after this. This basically extends the default timeout of CLI from 3s to below value

VITEST: "1"
OPENCLAW_TEST_HANDSHAKE_TIMEOUT_MS: "60000"

[hollow bay]

power company sends notifications hey your power is trending higher lol I am sure they are seeing alot of that

[copper nova]

I just got a PR in to fix this issue LOL

[bitter osprey]

Hello Openclaw developers & activists, Please allow me to ask, I have connected Openclaw on Mac mini as a local, for the agent I have used GPT by logging in to the GPT Pro account, now everything is running, local chat can be done, now to connect to WhatsApp why does it fail? Every scan on WhatsApp iPhone is always unsuccessful and the information says try again later, I have tried on 2 WhatsApps, both business and regular WhatsApp, the results are the same, does anyone have a solution? Thank you in advance

[copper nova]

wait you managed to set the timeout with env variables? I tried it and it doesn't work, I had to monkey patch it into the dist. I created a PR that just extended the timeout to 10s because 3s was too aggressive, I also made it actually configurable by env as the current setup is locked behind VITEST as you mentioned which is for test env.

[nimble hill]

With the above 2 vars I mentioned, it works fine, without which 2026.3.13 on docker or k8s was not responding to cli calls

[wintry swallow]

Guys, I'm so tired of you giving him a task and him going to sleep halfway through and not responding anymore. When you write, "Stop being silent," he's been saying "Oh, sorry, I forgot to write, I was drinking beer and forgot about your task." Can someone PM me what I need to do to make him always respond, complete the task, and write about the process? Please, I'm so tired of writing to him 10 times whether you're here or not. Please send me a guide or instructions in PM. I just won't be able to find the messages in this chat. Thank you all so much.

[ancient jungle]

so i was running into that issue, are you running local by chance? like LLMs locally

[wintry swallow]

Vds+claude api

It just really pisses me off

[ancient jungle]

weird, shouldn't be having a problem. i was having your issue because running anything decent locally is pulling teeth and that thing just falls asleep on me mid conversation

[copper nova]

welp, I hope my patch helps you out then lol I increased that timeout to 10secs by default

[weary marten]

Bonjour, j'utilise MaxClaw (cloud). J'ai configuré un bot Telegram (@leeramax_bot, token configuré via config.patch, pairing approuvé). Le bot reçoit bien les messages (visible via getUpdates API), mais le long polling ne démarre pas — les messages ne sont jamais consommés par le gateway. J'ai fait plusieurs restarts SIGUSR1 sans résultat. Est-ce que le plugin Telegram est bien supporté sur l'instance MaxClaw hébergée ?

[wintry swallow]

Well, just for fun, I wrote to him to rewrite a 10-line file, I've been waiting for 40 minutes already)) And if I write to him, why are you silent, he'll say, oh, I forgot to answer

[empty lintel]

can i just sanity check that .openclaw/config.yml is not a thing, and the only config file is the .openclaw/openclaw.json right ?

[nimble hill]

Which model are you usong?

[wintry swallow]

Sonnet 4.6 , opus 4.6 etc

[nimble hill]

Well those should be working good. Maybe setup a proper prompt in .MD files detailing what you want to achieve as an objective. Provide specific detials for what you accept as putputs

[dapper basin]

Can someone recommend an alternative to moltbook that is useful?

[rugged bramble]

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

[ancient jungle]

sonnet 4.6 definitely hallucinated a config.yml when it tried to help me the other day. i don't think sonnet is actually familiar with openclaw's structure. openclaw's docs are also horrendous

[wintry swallow]

I somehow configured it on the previous server, and yesterday I transferred it to the new one. Everything was fine all day, but now it’s started to sleep, a new session doesn’t help.

[wintry swallow]

Yep me help kimi 😅

[nimble hill]

Dunno...thats one reason I use n8n for specific flows. Cannot rely on OC to perform consistent tasks. Also save lot of tokens that way

[wintry swallow]

I seemed to have set up some limits on the previous server and it worked, but this isn't normal, it's happening and there's no solution.

[tall meteor]

@copper nova I tested out the dev team on the single GPU today. To make it transparent, I put together a task monitor script. It shows the Mission Control API activity, agents, and the system scheduler running Node.JS/Fastify/Redis for GPU queuing.

Next up: devops agent for CI

[copper nova]

what did the team build doe?

[tall meteor]

agent:~$ task-monitor

──────────────────── OpenClaw Task Monitor ────────────────────

  Monitoring: /home/agent/.openclaw/agents
  Scheduler:  http://127.0.0.1:7700/status
  MCD:        https://agent.lan.arpa/api/activity
────────────────────────────────────────────────────────────

Watching 20 existing session files. Waiting for activity...

[18:37:46] mcd            ACTIVITY   ← [user] Task Status Changed: Task "Hello World" status changed from in_progress (Hello World)
[18:37:46] mcd            STATUS     ← task 'Hello World' in_progress → todo
[18:44:09] dev_manager    SESSION    → new session file detected
[18:44:11] dev_manager    SESSION    → started  cwd=/home/agent/.openclaw/dev_workspace
[18:44:11] dev_manager    READ       → SKILL.md
[18:44:14] dev_manager    READ       → runbook.md
[18:44:21] dev_manager    EXEC       → POLL GET /api/tasks?status=todo"
[18:44:48] mcd            STATUS     ← task 'Hello World' todo → in_progress
[18:44:51] dev_manager    EXEC       → PATCH tasks/cmmzovcl... → in_progress
[18:44:51] mcd            ACTIVITY   ← [user] Task Status Changed: Task "Hello World" status changed from todo to in_ (Hello World)
[18:45:03] dev_manager    EXEC       → TOKEN POST /request
[18:45:03] dev_manager    TOKEN↩     ← granted (holder: dev_manager)
[18:45:03] system_sched   TOKEN      → GRANTED holder=dev_manager task=cmmzovcl... ttl=1198s
[18:45:31] dev_manager    EXEC       → HANDOFF wake-agent dev_coder (task: cmmzovcl...)
[18:45:47] dev_manager    EXEC       → TOKEN POST /release
[18:45:47] dev_manager    TOKEN↩     ← released (holder: )
[18:45:47] system_sched   TOKEN      → RELEASED by dev_manager
(truncated because of Discord)

[copper nova]

it built a hello world app...

LOL

[tall meteor]

It was a test to write a hello world file into the project directory

[copper nova]

you son of a gun

[tall meteor]

It was to test the system scheduler and "wake-agent" and "task-monitor" scripts

[copper nova]

add to the pipeline to create a new git repo for each project and to push to it between each state

i actually make it push to git between every build step even, working off git work trees, but that's a little heavy and complex lol

sometimes it fucks it up and I have stale branches everywhere

[tall meteor]

Right now I'm allowing the coder agent to create a feature branch before working on Mission Control, then the auditor agent to check its work. Then after that passes QA, it hands it over to devops. All of them touch git

devops will clean up all feature branches after merging, both local and remote

[copper nova]

nice work so far, good job

so this is a little look at the chat app I built ontop lol

[tall meteor]

So Mission Control Dashboard has a task tab, modal with a form to create/edit tasks, assign task types "coding/etc.", kanban board with to do/in progress/review/done with drag & drop cards with tasks assigned, then an activity tab showing milestones a little like the Task Monitor I posted. It's slick.

[copper nova]

I know what message it's responding to, I can quote the message so other agents know what session it came from. and where to look for context

like I'm very happy with it so far lol

if I have the agent working on a big task, I can steer it mid thinking so it doesn't keep going down the wrong rabbit hole on something

[tall meteor]

That happens. A lot of it is the agent taking liberties with the SKILL.md and runbook.md

[copper nova]

I also got my 4 PRs merged to core, are you proud of me yet?

[hasty grove]

Does anyone want to look at a specification I created and offer comments?

It is a spec for [RSS + event sourcing + git-like diffs for] LLM context window evolution

[nimble hill]

Ha ha. Why do you want validation from others to be proud? 😃

[copper nova]

yes

[tall meteor]

Today is day 10 since the mini PC was delivered and unboxed, so I am celebrating wins 🎉

[copper nova]

my ego needs it

[nimble hill]

well, that means you let others control you. I would never want to be in that position

[nimble hill]

what specs?

[copper nova]

it's gonna be running linux, I know it

[tall meteor]

It's a dirt-cheap mini from Amazon:
https://www.amazon.com/dp/B0FKLBXS77

[copper nova]

pretty good specs for that price tbh

I was being a cheapo and only got the base 256gb for the mac mini lol bought a dock that allows for a m.2 SSD to be installed

then I looked at SSD prices

I bout a 2tb for like 160cad a few years ago, that thing goes for like 500cad now, freekin insance

[nimble hill]

good - the only problem with these PCs is the durability - sometimes their heat sink is not properly seasted, lousy soldering that comes off. Hence, keep a regular backup so that you could just offload to a new one if it needs. thats one reason I go for atleast min branded ones from minisforum, beelink and if I want to run production for my clients, i suggest them Asus Mini PC which are expensive but do not miss a beat when running 24x7 even in dusty conditions 🙂

[nimble hill]

gold - ram - ssd = are worth roughly the same now 🙂

[copper nova]

you need to be in the top 1% to build a gaming PC now apparently lol

making Mac pricing more reasonable LOL I use to think paying that upgrade was insane, now its... almost industry pricing

[nimble hill]

and most top 1% don't even care abut gaming - they are busy making money 🙂

[copper nova]

don't you know Elon is one of the top Diablo 4 and PoE 2 gamers?

[nimble hill]

So I found a good working - I trigger n8n workflow and call openclaw using http request nodes. Openclaw uses OAuth (n8n cannot), does a specific agent based on the specific request from n8n and completes the daily tasks. This save lot more tokens than directly using openclaw and is much more reliable because of the guardrails I cab build in n8n. Now this is what I would be selling to clients for there use cases

[nimble hill]

nope and don't care!. And you can be excused to believe that he actually does it himself

[copper nova]

lol it's a meme, man's already admitted he pays chinese farmers for accounts

[copper nova]

I had a notion kanban board setup and agents can use the MCP server to update and read tickets.

different board for different things like ops, product, etc.

[tall meteor]

I'm thinkin of getting another for a LAN infra box. I already have a separate mini for DNS (dnsmasq) and step-ca cert authority.

[copper nova]

that's some serious home labbing setup lol

[tall meteor]

Yeah, I bought some Duronaut thermal paste. I'll check it out

[copper nova]

my next update probably a DAS and some storage, run a plex server

[nimble hill]

checkout openproject - kind of a standard used by insdutry if they are not using JIRA. using n8n, you can control fine grained control. Openclaw can just call the n8n workflow as well

[dark glacier]

Hi, I just integrated ChatGPT Plus subscription into my openclaw agent. But when I gave it a simple prompt "open example.com", it says it can't access the browser control, though "openclaw browser open https://example.com" works. And also minimax and even openai api works. What;s the problem? please help me.

[nimble hill]

I also keep the bottom lid open and place over something like this compact cooler to keep cool - those NVMEs sure get hot ...something like this https://gbr.grandado.com/products/octopus-notebook-cooler-pad-quiet-stand-double-fans-for-7-15-inch-cooler-cooling-usb-laptop-foldable-fan-notebook-i0e3?variant=UHJvZHVjdFZhcmlhbnQ6OTQwNjk0ODc

[dark glacier]

Can you help me?

[nimble hill]

build a NAS - all it needs a linux machine with 2 HDD / SSD and ZFS - you get some protection

[tall meteor]

Not a bad idea. I'll do that

[copper nova]

that's what it is essentially

it's a DAS attached to my openclaw agent which is connected to the network

[tall meteor]

What's nice is the case is vented and has a BIOS controlled fan

[copper nova]

the DAS is just the box to hold the hardrives, the mac mini is the NAS part with the software and network control

[tall meteor]

The hottest running unit I've ever owned is an EdgeMax 4-port router - yow

I love EdgeOS though

[nimble hill]

generally DAS terminology is with single disk. If it is a multi disk RAID solution, then you are good. but those boxes generally have limited features

[copper nova]

I thought that's what those are called? the boxes with like 0 functionality but just connect hard drives lol

[dark glacier]

Hi, I just integrated ChatGPT Plus subscription into my openclaw agent. But when I gave it a simple prompt "open example.com", it says it can't access the browser control, though "openclaw browser open https://example.com/" works. And also minimax and even openai api works. What;s the problem? please help me.

[copper nova]

thinking like a 4bay HDD with a SSD scratch drive, raid 5?

[nimble hill]

the pooint I am making is RAID - does it do redundancy? those boxes sometimes offer hardware RAID whihc has become non-standard nowadays and are expensive. typical DAS boxes with multiple drives is JBOD with no redundancy if 1 or more disk fails

[nimble hill]

bad idea with RAID 5 - it means only single disk can fail. In 4 disk, having only 1 failure is still on the edge as if 1 disk fails and you replace, there is a chance that when resilvering is going on, the other one can also fail losing the entire pool.

[copper nova]

I just assumed I would have to do the RAID setup manually anyways lol if it come with the ability to set that up then it's a bonus

[open saffron]

Has there been any reports about kimi model acting even worse than usual with this latest update?

[copper nova]

what raid you reccomend?

[nimble hill]

for 4 disks - RAID 6.

[copper nova]

domo arrigato fine sir

now I just need to sell a kidney to get some hard drives

[nimble hill]

don't need to buy all at one. buy 2 and upgrade as you go on

[daring remnant]

how do i fix this

"PATH missing npm global bin dir: /home/myuser/.npm-global/bin
This can make openclaw show as "command not found" in new terminals.
Fix (zsh: ~/.zshrc, bash: ~/.bashrc):
export PATH="/home/myuser/.npm-global/bin:$PATH""

[copper nova]

it tells you right there, export path

[barren blade]

@copper nova language

[copper nova]

huh?

[barren blade]

You said naughty words

[copper nova]

where

[barren blade]

Here, a while ago

[copper nova]

prove it

[next walrus]

Lame, I'm hitting month long rate limits in 1-2 days even using oauth token, didnt even get claude to actually do anything

What's a good backup model that actually works. GPT hallucinates way too hard

[echo pine]

Opus 😁😝

[next walrus]

I haven't even used Opus, I hit a months API limits in not even 4 days using Sonnet, and I never even got it working. OC had all kinds of crash issues most of the time.

[echo pine]

weird man. Sonnet is solid as hell. What kind of stuff was it doing / what were you trying?

I'm always baffled when people have so many issues. My two instances have been so rock solid. I install, feed it a ton of info about myself/the goals, and they... just work. I mean, I tweak on stuff like memory upgrades and search upgrades, but overall, just solid with any frontier models

[next walrus]

Gateway is constantly restarting, I had it try to make a memory curation bot per it's suggestion, that freaked it the hell out. For research it suggested making an agent that uses haiku instead of sonnet to conserve usage, told it to go ahead and do that, everytime it tried to use that agent it crashed out the whole openclaw hard. All sub agents got sent to the bench and now I'm rate limited until april

that's on the Max plan

[ancient jungle]

i mean, are you willing to pay?

[echo pine]

Damn. Rough. I'd just stick with one agent, and no memory changes, especially anything it's making, for a while.

[next walrus]

If It stops wasting tokens on hallucinations, sure.

[next walrus]

But burning through the max plans entire usage in 3 days mostly on hallucinations seems a little ridiculous

[echo pine]

Personally, I'd just nuke it and start fresh. Fill it with EVERYTHING you can think of about yourself and your goals. Enable QMD search. And start with something super small for it to do. Like a daily report of the weather and yesterdays news.

[ancient jungle]

i was just testing minimax 2.7 today through their token plan, seems to work great, and it's a monthly sub instead of pay as you go (though that's an option as well)

[next walrus]

I can look at the bottom corner of windows for a weather and time check, if AI is going to cost me money, I need it to do something for me than I can do by turning my head

[next walrus]

How are the hallucinations?

[ancient jungle]

it's not? i mean it makes some mistakes here and there, but it usually catches itself and fixes it

[echo pine]

I'm suggesting something that is easy for it to accomplish, easy for you to setup, and also easy for you to verify it's working properly. So you can get used to the cron system, memory, daily files, verify some stuff.

[echo pine]

Describe these hallucinations you were having?

Sonnet rarely does that.

[ancient jungle]

i'm actually building my kids some kind of tutoring/trivia app today, and minimax 2.7 seemed to do well. now i'm not coding anything crazy, but it definitely gave me a decent product

[next walrus]

Oh I had it telling me it built out a dashboard of things to check on and status to monitor that was functional and it didn't even make anything. I've had it install completely wrong versions of things. As mentioned, it completely hallucinated building functional sub agents to do basic things like search the web and draft documents for me.

And when it got to sub agents, pretty much anytime it went to utilize them, openclaw crashed out hard

constant gateway crashes and it couldn't even attempt to read them

[ancient jungle]

i straight up gave minimax an API key to my github, spun up a separate LXC, gave it ssh keys, and just let it go lol

[next walrus]

Cool, I can't get sonnet to search the web properly, any idea how to fix that?

[ancient jungle]

do you have web_search properly enabled in tools?

[echo pine]

Just have it spawn generic sub agents for a bit. Until you have something particularly specialized you want to focus on.

Like I said, if you want advice: Scrap that entire install. Use MiniMax 2.7, fresh install, fill it FULL of info about yourself and your life, work, family, goals. Start small. Don't go making more agents, having it modify anything in the memory (that's already done btw, you want Lossless-Claw). Just do some basics, test, and slowly move on from there.

[next walrus]

No idea how to even check that

[next walrus]

Yea, I had it make a sub agent to search the web to conserve token pulls and it started gateway crashing openclaw constantly

whats lossless claw?

[ancient jungle]

can't attach a picture here, but can you get into the GUI dashboard for openclaw? it's under Agents -> Tools

[next walrus]

yea I'm there

[copper nova]

you shouldn't be able to do that tbh, you hit session max before you hit monthly max and that should hold you off from hitting it for atleast a week

[echo pine]

https://www.youtube.com/watch?v=_aXfcVO6Pbs

[next walrus]

No, on claude dashboard it tells me I will regain access 4/1

[echo pine]

Whats this page say? https://claude.ai/settings/usage

[next walrus]

Session usage 0%, Weekly All models 11% Sonnet only 18%

[echo pine]

So you're good to go brotha. Burn them tokens.

[next walrus]

my openclaw is telling me rate limit reached

[ancient jungle]

you just toggle it on if it's not already

[copper nova]

what does your claude code tell you?

when you spin up claude code and type /usage

I bet your on some API free tier on your openclaw lol

[echo pine]

Probably

[copper nova]

I can't even figure out a way to hook it up with my sub yet

only through sub agents

but not main

[next walrus]

Toggle what on?

[ancient jungle]

web_search tool

[next walrus]

It's on, but it can't view pages that use 'cloudflare' according to it

or anything coded in java

[ancient jungle]

oh, that's normal

[next walrus]

I'm on the Max plan, $100/mo

[mint jay]

Fix-it:

• tools.sandbox.tools.allow
• tools.sandbox.tools.deny
• agents.list[].tools.sandbox.tools.allow
• agents.list[].tools.sandbox.tools.deny
• tools.elevated.enabled

how to use this things

[ancient jungle]

a lot of sites have been blocking bots. i was working on having my bot track my USPS packages, spent the whole morning trying to fight the javascript issue. i lost and gave up

[copper nova]

yea I know, but how did you hook it up woth open claw

was it a API key?

if it was a API Key, you are not using your sub

[echo pine]

he's trying to reply, but his mental rate limit has been reached

[copper nova]

he also didn't tell us what his claude code usage says

[echo pine]

I often forget how many little things I've learned over the years, and it's hard for me not to assume everybody knows these things. Small things I check and configure, without really noticing. Keeps my kit running well, while others struggle. I gotta remind myself that from time to time.

[nimble hill]

register free with USPS for the MID and use their API to track - it is much easier. https://www.usps.com/business/web-tools-apis/track-and-confirm-api.htm you can just create a cron job to run and ask openclaw to write a program that can run it

[ancient jungle]

i tried to sign up for a developer account, and then stopped me and wanted me to email them (technical difficulties), then i tried to do the same for UPS and it's doing the same thing there. maybe my IP is flagged

[nimble hill]

maybe - they need US addresses

[ancient jungle]

definitely have one of those 🤣

[nimble hill]

Did you try the direct browser way - openclaw can host a browser and mimic actions - I think USPS tracking is fairly straightforward https://tools.usps.com/go/TrackConfirmAction

[amber timber]

Can I get some help with actually downloading CLAW and get some run downs on some use cases? I'm new to open CLAW

[echo pine]

Sure! Lets do it. Do you have a place to run it? An old computer? Mac Mini? RPi?

[amber timber]

I'm using my Macbook Pro M5 MAX @echo pine

Hop on a call ? @echo pine

[echo pine]

Like, your only, personal laptop?

[ancient jungle]

yep, tried that too. had browser took a screenshot and it was blank

[amber timber]

yes im running this on my personal lap top.

Just added you @echo pine

[nimble hill]

not the screenshot, ask it to enter the tracking id and click Track- obsever for 30 sec and provide the details via a JSON. it would give you a proper formatted reply - either a success or a failure with error message

[ancient jungle]

there's nothing to click if there's nothing to see

[echo pine]

Well, very doable, but be careful. If the LLM reads something the wrong way, or you type something dumb, it can start deleting system files and hose your system. I mean, I'd do it personally, but I'm fairly confident in what I'm doing.

and I don't have time for a call right now, or to run you through a docker setup (royal pain, imho)

[echo pine]

What step are you on? Do you have it installed?

[amber timber]

I wanted to get local LLM's a go because I don't want big AI companies sending my data to a cloud.

[copper nova]

what docker setup lol

[nimble hill]

how much RAM - if it is 64 or 128 GB - you can easily run a local model.

[amber timber]

I have a docket set up already @echo pine

[echo pine]

Me personally, I wouldn't run a local model as my main 'brain'. Opus or GPT5.4 only, for me. And if you're concerned, I guess stick with the US based frontier companies, and hope for the best.

[copper nova]

what is.... so special about your data that you need secured? just curious.

[echo pine]

Shoot man. If you know docker, you're golden. Just get an LLM, open a terminal, copy the install command from the openclaw site, give it EVERY bit of information that you can think of about yourself, and go forth. Start small. Expand from there.

[nimble hill]

wdym - I have a windows VM, which has a openclaw node controlled by remove OC gateway. On gateway when I chat stating, open a browser session on windows node and go to https://tools.usps.com/go/TrackConfirmAction and give me details what it has - it gives me the details correctly. I can then ask to enter a tracking ID and click track - it gives me the details - sometimes JSON format is wrong, but I get the details

[amber timber]

If your not concerned about keeping your data private then your a sheep bro. I do not want ANY company having my data.

[ancient jungle]

i run mine on a headless ubuntu. now was that the greatest idea i've ever had, probably not. but i'm too deep in to reconfigure this piece of shit

[copper nova]

if you think google doesn't already know more about you than you do, I don't know what to tell you buddy lol

[amber timber]

i don't use google buddy, I run linux and tor browser

[echo pine]

Than this project may not be for you brotha. The frontier models make OC sing. A little qwen model is going to give you trash results.

[copper nova]

Palantir already knows what colour underwear you wearing

[nimble hill]

thats good for the gateway - not if you want to launch a browser. But openclaw can also run chrome headless - check it out, Since you are not trying to maintain any seesion, it might work there as well

[ancient jungle]

already explored that route as well, installed chromium lol

[copper nova]

you have a m5 macbook pro running on linux?

[ancient jungle]

trust when i said i've tried everything

[amber timber]

Yeah I was looking into that I might need to make a Claude account

[amber timber]

I speaking about my PC but I use the same browser on my macbook

[copper nova]

@echo pine have fun, my brain hurts, I need to figure out how to fix this claude sub auth thing for everyone

[echo pine]

I'm outta the loop, is it busted?

[amber timber]

AI companies love people like you lol give them everything with zero issue lmao

[nimble hill]

in today's world you can't choose much, unless you are willing to spend a lot and sign NDAs.

[amber timber]

I'm not a tin foil hat thinker but I just want to be as private as possible there should be no issue with that 😂

[next walrus]

its on my oauth toke, claude code is 0 usage

[echo pine]

There's being concerned about privacy, and there's also the known exceptions I make, carefully, to have a tool that runs well.

[copper nova]

you are not gonna have a fun experience with local LLMs is what we sayin

[amber timber]

exactly thats why I might use

[weary basin]

Do y'all know of a way to get your agent to search/view Reddit besides applying for Reddit API? My bot can't see posts even when I give direct URL and he views with chromium browser. Says it's blocking bot traffic.

[amber timber]

I'm open to using claude lol i'm just trying to see what my options were lol

[copper nova]

is this laptop purely for openclaw?

[amber timber]

No lol

[nimble hill]

Since you already have good machine - load up local models and test out gpt-oss-120b is great - it is some complex reasoning that you might need to tweak the prompt and work out. As we go along, this is only getting better and specailized models (per domain) will be releasing

[echo pine]

They're just not good enough. Is the short version. Dumb, always giving up, poor choices. Have the frontier model be the brain, and your locals can do the specific work that your planner brain doles out to it. That my opinion though

[copper nova]

so you an use claude code no problem?

[next walrus]

probably, haven't tried to, i just wanted to be able to use claude through discord and get agents doing some basic work for me

[copper nova]

if you care about privacy, I would recommend not use your personal laptop for claw use

[amber timber]

Yeah the crazy thing is I used to use open AI for everything but during important tasks it hallucinates. What is your take on Claude? @echo pine

[copper nova]

well try it lol it will tell us if your max plan is actually maxed out or not

[next walrus]

yea, last time I tried GPT i was getting bad hallucinations. It was thinking it wrote thousands of lines of code, it wrote 4.

[next walrus]

"You have reached your specified API usage limits. You will regain access on 2026-04-01 at 00:00 UTC."

[next walrus]

I don't even know how to access Claude Code from my terminal

[copper nova]

this is claude code?

[copper nova]

you ever install it?

why do you have a claude max plan and not have claude code?!

[next walrus]

https://platform.claude.com/dashboard

[copper nova]

wlajrweglajrgawekg;naer

[echo pine]

That's what we're saying man. It's not configured right. You're hitting some free API limit, and your paid claude subscription is sitting over there unused.

[next walrus]

how to I connect claude code to openclaw?

[copper nova]

you don't

atleast I'm trying to fix that right now lol give me a minute

but it won't be ready for a while

next release maybe

[amber timber]

100% I was having hallucinations on deep research even using there extended thinking model and was giving me information that wasn't relevant @next walrus

[echo pine]

It's great! I find Opus to be the clear winner for current frontier agent use. It tries hard, finds a way to do things, and is pleasant to talk to. GPT5.4 is good, but not as nice.

[next walrus]

Okay, so confirmed, I'm not using claude code, but i also wasn't trying to. I want to be able to talk to claude through discord and have it have it's agents do various tasks for me but it's agents all freak out and start hard crashing the gateway and I get API limited in 4 days

[copper nova]

did you manage to configure your claude sub to your openclaw main agent model selection?

[next walrus]

yup, last i checked

[next walrus]

if you tell me how to check that i can

[amber timber]

How good is it on up to date indepth research? A lot of my workload is extensive research and reasoning but its hard for me to trust AI's outputs, I even go as far as triple checking my prompts but even that isn't affective. If I can have a model that can do that for me it would cut my work load in half @echo pine

[copper nova]

yo what you doing, it better not be your token, delete that lol

I don't need to see it

[next walrus]

usage panel?

from openclaw dashboard

I dont think I've got any tokens that show there

[echo pine]

openclaw models status DONT show us that result. But check it out and see if it's mentioning oauth a bunch

Auth overview
Auth store    : ~/.openclaw/agents/main/agent/auth-profiles.json
Shell env     : off
Providers w/ OAuth/tokens (3): anthropic (1), github-copilot (1), openai-codex (1)
...

[next walrus]

i see the word oauth a few times

[echo pine]

Tell you what I'm sick of... openai-codex:default ok expires in 3d

Constantly expiring!

[echo pine]

I literally just asked my agent. They know all.

The Auth overview section shows exactly what's going on. He'll see something like:

anthropic:default=token:sk-ant-... → that's an API key
anthropic:default=OAuth → that's OAuth

It shows each provider, whether it's using a token, API key, or OAuth, and where it's pulling credentials from (auth-profiles.json vs environment variables). Pretty much answers the question in one shot.

If you need to tweak your models, and how they auth, use the openclaw config, and re-setup your models. Wipe them all out, and start with just one, oauth? Sound reasonable? Then you know for sure?

[copper nova]

wait, you got your anthropic sub to work as main model?

wtf, why doesn't it work for me?!

[echo pine]

I have a key from claude code from way back, clawdbot days. Still works 🤷‍♂️ I lost it for a bit, and tried to make a new one, and it didn't work. So I hunted down this old bugger

[copper nova]

I can't generate it right now lol

[echo pine]

Here... let me paste it in so you can use it... GOTCHA

[next walrus]

profiles=1 (oauth=0, token=1, api_key=0)

Is that a problem?

[copper nova]

the code it generates is NOT the token i need

[echo pine]

Yeah, I dunno. I just follow the steps on the website, claude setup-token and all that

[crude wren]

hi all any idea if this is ollama issue or openclaw ?
⠧ streaming • 557m 38s | connected
agent tooling (tooling) | session main (openclaw-tui) | ollama/nemotron-3-super:cloud | think low | verbose on | tokens 45k/262k (17%)

[next walrus]

profiles=1 (oauth=0, token=1, api_key=0)

Is that a problem?

[copper nova]

yea that token is not the SK_ewgw token

[tall meteor]

I think I made the right choice by scaffolding the Mission Control Dashboard before allowing agents to begin development. Without Next.JS, real-time WebSocket data would be a tough retrofit. React is 🔥