#users-helping-users

1 messages · Page 46 of 1

velvet venture
#

i actually tried Googles things

#

just always got Rate limited instant

#

gemma doesnt support tools

#

and tried asking GPT KIMI 2.5 GEMINI CLAUDE and COPILOT(github)

#

I did not really know what to do

#

If you have recs just tell me and sorry if i annoyed you.

indigo goblet
#

No no worries, dont be sorry

velvet venture
#

recommends

#

Im not from america or UK so i dont know if i got the shortcut correct

final patio
#

Openclaw tui
For cli

indigo goblet
#

Did you research online

velvet venture
velvet venture
indigo goblet
#

So you might want to adapte to that updatr

final patio
#

It is.. nemotron-3-nano:4b

velvet venture
#

oh. sorry i tought the name was just that.

final patio
#

Yeah i didn't say it correctly earlier.

indigo goblet
#

If at one point there is a bottleneck, either you pivot or you change your setup

#

No shortcut lol

crude acorn
#

so i had a dumb newbie question: i am just getting it all set up and i used to use MCP SSE servers for my homelab infra, grafana loki prometheus proxmox etc - what's the best way to hook something like openclaw into it, do I use the same SSE MCPs in case multiple systems need it? or is there a better way to do this via skills/plugins/etc? i'm still getting used to openclaw style stuff specifically

velvet venture
velvet venture
#

and if i get bottle neck i could just close it off for few hours ngl

indigo goblet
#

I mean

#

Based on your hardware

#

And what you want and what is available

velvet venture
#

yeah i get what u tryna say

#

not possible

#

unless pay

indigo goblet
#

Yesh understood

#

I mean, what about a sub like gpt plus?

velvet venture
#

still thanks for the feedback i guess

#

I do not have a bank account

crude acorn
#

if you're looking for free and have the hardware, local LLMs can work too - mine uses a small 9B model and its not the biggest sure but i get 130K context and halfway decent speed, and its sandboxed in case things go wrong

velvet venture
#

AND to top that I live in turkiye

indigo goblet
#

Because, I see a lot of people headache with setup when in reality, your agent setup is normally the headache

velvet venture
#

1 dollar is 44 turk lirası

indigo goblet
#

Not the model

#

Hmm i see

velvet venture
indigo goblet
#

Cant use basic go at 8 bucks to get your feets in water or even 20 bucks plus?

crude acorn
#

what is your hardware atm?

velvet venture
#

266 turk lirası every month

velvet venture
crude acorn
velvet venture
crude acorn
#

it wont be as amazing as chatgpt don't get me wrong, but free is free 😄

#

my agent is 100% local, Qwen3.5 9B, unsloth Q4_K_XL - 131K context

velvet venture
#

IF groq worked it would be great enough but NOOO it gives Rate limit instant

indigo goblet
# velvet venture 266 turk lirası every month

Dont know the value bud, but trust me and this is just personal opinion, not worth it to headache with local model when literally, what makes your agent work on the long term is literally its own setup

velvet venture
crude acorn
#

yeah unfortunately its very hardware dependent

velvet venture
#

well dont get me wrong but ryzen 5 5500 16 gb is still solid with these crisis about ram prices

but my gpu is too old atleast rx 580 8 or 12 (if iremember) would be great

#

im trying the nemotron rn

#

lets see how much will it take its time.

#

seconds or minutes.

#

lets see

indigo goblet
#

You do you bruv, but setting up a model should be done in 30 seconds, what comes after is real work and what allow you to reach your goal

crude acorn
#

another alternative is openrouter, a one time reup on credits and when they expire you get a bunch of free tier stuff for people who have bought credits at least once

velvet venture
#

lo

#

lol

#

btw lil question

#

why do i get instant RATE LİMİT When i try GROQ the free provider?

#

atleast give me a lil hi groq 😭

crude acorn
#

you're most likely hitting "request per second" caps, which are super low on free tiers - and agents like to make multiple calls with tool calls and such

final patio
dim harbor
velvet venture
crude acorn
#

yeah thats why i went with my 9B Qwen locally, lets me vastly up the context and it made a HUGE difference - no more typos or hallucinated tool calls

velvet venture
final patio
#

No cc

velvet venture
#

or openrouter

final patio
velvet venture
#

Let me look up then ig

crude acorn
final patio
velvet venture
velvet venture
#

i used it today but

#

token gone

#

in like 5 messages

#

seems nemotron still didnt respond

#

great.

#

btw.. i dont think ollama is using my gpu either.

#

because it sits at 3%

#

just my cpu getting crazy

dim harbor
#

nvidia has free tier api also if you looking for free tier apis

velvet venture
#

Wait nvidia had api-?

#

gosh why didnt it hink of that its a goddamn ai company too

#

ill check that out too

final patio
velvet venture
fresh hamlet
#

Greetings, permission to ask a stupid question?

final patio
# velvet venture btw can i make it that use my gpu too

searched online:
If Ollama is not using your AMD GPU, ensure that you have the correct ROCm drivers installed and that your GPU is supported. You may also need to set environment variables like HSA_OVERRIDE_GFX_VERSION to force the system to recognize your GPU correctly.

velvet venture
#

oh.

#

ill try to update my drivers first then try the config

final patio
fresh hamlet
#

I am keen to learn the advantages of OpenClaw, however, my boss has told me I am not allowed to put it on the work machine.

velvet venture
fresh hamlet
#

Would it be a terrible idea to set up some kind of desktop as a service and put it on there so I can use it in that environment?

crude acorn
#

i did also have a dumb question 🤣 i previously had a few SSE MCP servers set up for opencode, should i just use those for openclaw or is there a better way to hook into my services? i'm legit not sure how skills work yet lmao still learning how this works compared to the old way of just using open-webui

dim harbor
# velvet venture Wait nvidia had api-?

https://build.nvidia.com/settings/api-keys

its through the build.nvidia.com its kinda tough to navigate around to find stuff

yeah its not amazing, but its free at least i kept a few models in openclaw for fallback but not using them unless something breaks with my other providers

these at least respond relatively quickly i just checked from openclaw:
nvidia/qwen/qwen3.5-122b-a10b
nvidia/nvidia/nemotron-3-super-120b-a12b
nvidia/stepfun-ai/step-3-5-flash
nvidia/moonshotai/kimi-k2.5
nvidia/minimaxai/minimax-m2.5

with kimi had some issues it couldnt read/write from the files, but others worked

velvet venture
#

The RX 550 is not officially supported by Ollama. AMD's ROCm library (which Ollama uses for GPU acceleration) dropped support for the Polaris architecture (GFX8) that the RX 550 is based on, and Ollama's official compatibility list only includes RX 5700 series and newer, plus Vega and select older cards

#

goddamn it

#

Option 2: Force ROCm Override (Linux Only)
You can try forcing ROCm to recognize your GPU by overriding the GFX version. The RX 550 is Polaris (GFX8), so try:
bash
Copy

Edit the systemd service

sudo systemctl edit ollama

Add to the [Service] section:

[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=8.0.3"

i can still do this too (looked up ai)

velvet venture
crude acorn
#

lmao yeah i had to do that too, i use nixos and i had to make sure i installed the rocm version of the package myself

#

it may be a different package to add rocm support if you're installing it via your package manager (which is the recommended way on most linux flavors, since it will usually ensure rocm or cuda is installed)

velvet venture
#

i did the force thing

#

lets try it

#

❯ ollama ps
NAME ID SIZE PROCESSOR CONTEXT UNTIL
nemotron-3-nano:4b 6cc467f05439 9.6 GB 100% CPU 262144 4 minutes from now

~
❯

nnnnnnnope

#

plus 9 GB wont work

#

RX550 only has 4gb

final patio
#

It's working i guess.

velvet venture
final patio
velvet venture
#

it shouldda said GPU too

#

gosh man.

fresh hamlet
#

Or if anyone could recommend a good set up to start - that would be fab

velvet venture
#

I dont want to give up too cus this is cool having an ai of your own

#

well not that you own

final patio
#

Yeah.. rotated screen to see format correctly.

velvet venture
#

still having one in your system

velvet venture
#

i just wanna hit my head to the goddamn desk

lost hound
#

Hey folks, after some struggling, I am posting here. I am having a strange issue where when I chat with the main agent through the dashboard, sometimes the agent just stops responding while it's making tool calls and it doesn't say anything anymore until I give it an input message again. Once I do that, it gives me a flood of messages, which means that the messages were buffered but were just not showing in the chat. Is this a known issue? I have the latest release (but not the latest git I think).

velvet venture
#

How bad would it be

#

becausae i know it will be bad

#

fr

final patio
velvet venture
#

i will try for sure

final patio
#

You will be busy in constantly charging and cooling your phone.

velvet venture
#

constant cooling.. puff i need to buy a cooler but i dont ave one

#

well then later

#

goddamn it

#

im just making myself mad trying..

lost hound
indigo goblet
#

No point

velvet venture
#

im used to it

#

i just want to taste of succes yknow

#

that sweet spt

#

spot

indigo goblet
#

Well, success is a point of view

#

Point is, how do you reach your goal

velvet venture
#

imagine having a good enough cpu and enough amout even the best internet but not gpu

wild cypress
#

hello fellow devs i hope you doing good and stuck somewhere like me

wild cypress
indigo goblet
fresh hamlet
#

So how do any of you run this - locally or hosted?

wild cypress
copper nova
indigo goblet
#

And if the framework and structure around is minimal and weak

#

Nothing changes.

velvet venture
fresh hamlet
wild cypress
indigo goblet
#

High end model are more than capable but if the structure around is weak, pointless

velvet venture
#

btw the nemotron that i opened locally finally started talking and its taking 1 minute or so not that.. bad actually i can say

wild cypress
fresh hamlet
#

Igather that using with WhatsApp or telegram isn't mandatory

copper nova
wild cypress
velvet venture
#

there was a workaround to make ollama use ur gpu i dont game heavily so ill do that.

#

and if i game ill simply stop ollama

#

i love torturing myself

indigo goblet
#

Bruv, please look out online

#

Local are not opus or gpt 5.4

velvet venture
indigo goblet
#

Whats up with local model? Its not even the primary objective, setting up your agent is like 20 galaxy away more important

velvet venture
#

because i live in turkiye and its expensive to buy an cloud api?

#

Look. my plan wasnt local too

#

i tried cloud providers for 2 days. free and not free

#

all got rate limit

#

or token gone

indigo goblet
#

Api is always pricey

#

Chagpt sub bruv, you understand

#

Go for starting very cheap

velvet venture
#

I do understand you

indigo goblet
#

Plus or pro if anything

velvet venture
#

look. my parents WONT Allow me. I DO NOT OWN any BANK or CREDIT Account.

#

they think this is a waste of MONEY.

indigo goblet
#

Then realize this is not for you yet

velvet venture
#

fine goddamn it i give the hell up

indigo goblet
#

Fucking hell, no offense bruv but thisnt AGI jarvis from marvel movies

#

Like I said

#

Fire up chatgpt or claude, research and learn

velvet venture
#

i just dont know how to give up when you give hours onto it.

#

i give up

wild cypress
copper nova
#

only hours? pffft

indigo goblet
#

No no worries

#

Bruv, im saying its not a little thing to make it work

velvet venture
#

isnt it

copper nova
#

pretty sure it took me like 2 days to get it to actually work

velvet venture
velvet venture
wild cypress
copper nova
#

then it breaks every update, and I have to spend a few more hours trying to revive it

velvet venture
#

ive been trying openclaw for like 5-4 days

velvet venture
indigo goblet
#

How old are you by any chance brother? If that is not too personal

copper nova
#

if you get fustrated from a few hours of debugging, then maybe coding isn't for you lol

velvet venture
indigo goblet
#

Yes

velvet venture
#

almost 17

indigo goblet
copper nova
#

I like being on the newest updates

velvet venture
copper nova
#

cus I have changes in each update I want to test

#

cus I spend many hours fuckering about

indigo goblet
velvet venture
#

btw peak sentence there

indigo goblet
#

Basic is

#

What are LLM, weaknesses, what are they good at

copper nova
#

trying to do it all for free is gonna be the struggle lol

indigo goblet
#

Then, what are AI agent, exactly

#

Then what is openclaw?

copper nova
#

I tried it, rotating providers every time I hit a rate limit

velvet venture
copper nova
#

it was a misserable experience

velvet venture
indigo goblet
#

Like i said, chatgpt and claude is like the sandbox to get your hands dirty

#

Ram into it, try, research, be curious, try to break it etc

copper nova
#

your data might be used to train chinese llm models

velvet venture
copper nova
#

but w.e.

indigo goblet
fading tide
#

i love selling my data

velvet venture
copper nova
#

was my the last generation to start working at like 14yrs old?

velvet venture
indigo goblet
fading tide
indigo goblet
#

Where is the door?

copper nova
velvet venture
indigo goblet
copper nova
#

but kids these days definitely lacking that little bit of independence lol

indigo goblet
velvet venture
#

i am

#

why would i type sloppy then

indigo goblet
velvet venture
#

well i think ill sleep Goodnight thanks for your help sonny.

fading tide
indigo goblet
#

Now worries man

eternal cobalt
#

windows 11

indigo goblet
#

I lost count, my tokens are gone

distant spire
#

Guys kind of a newbie question. can i deploy openclaw as an agent to send outbound whatsapp msgs to a pool of numbers and engage in convos as if it were me (with some training ofc) and if and whenever a convo reaches a specific point the agent will alert me to pick up the chat the from there?

junior acorn
distant spire
# junior acorn yes

mhmmmm okay, can you please point me to the general things (scripts/plugins etc) that i might need in order to launch it so i will research it further?

#

thanks 🙂

junior acorn
#

There's no scripts or plugins for it specifically. It's something you'll need to setup yourself honestly, there's no guide to it.

#

Have you setup openclaw urself and asked it about your idea?

distant spire
#

I see

slow arrow
#

anyone else here have issues with messages from your instance containing nothing but 'Message Failed" after a recent update?

distant spire
#

nope not yet but as long as you're saying it is possible and there's nothing out of the oridinary i need to do i guess i'll be able to do it

#

Was just wondering if there's anything special needed. Thank you very much 🙂

indigo goblet
potent path
#

can I ask for help testing my tool for claw bot here or its not allowed by the rules ?

indigo goblet
#

Huh

slow arrow
molten scaffold
#

hey chat, question about using Lance DB memory, mines having trouble getting the cofig right, is that an outdated memory system or what? I was using QMD, don't remember why but I saw a video about lance so I was trying it but it looks like it didn't take

copper nova
molten scaffold
#

my bot editd it's config and it won't launch now with memory set to backend:lancedb

copper nova
#

Use another agent harness like codex or claude code to build your bot

#

The bot doesnt survibe gateway restarts to fix himself after he fucks something up

molten scaffold
#

ooooh, I see what your saying

#

do I tell claude to check the file system and help edit the config?

#

I've been playing this game for three weeks so that'll help

echo pine
#

yeah, point claude code at your .openclaw directory, tell them what you did, repent while it fixes the err of your ways

copper nova
molten scaffold
#

on it now

#

you guys are so smart thank you very much

#

I'm sure I'll be back

#

but hopefully not for a while

copper nova
spark turtle
#

Hey yall, I’m having trouble getting it to stop hitting the API rate limit (different from API limit)

Anyone know how to mitigate this?

mellow ermine
#

Are we allowed to discuss abliterated models here?

#

I don't want to break any rules

junior acorn
spark turtle
molten scaffold
#

what AI api are you on?

#

I'm using Kimi and I get that around 7-9pm EST every day

#

their servers are full or something around that time

cold inlet
#

gateway restarts are rough for this exact reason — the agent comes back with no memory of where it was
what actually works is having a cron fire automatically after restart that reads whatever state was written before shutdown — last daily log, active tasks, open loops — and sends you a resume card on telegram so you know it's back and oriented
the trick is the writing part has to happen consistently during normal operation, not just at restart. if nothing is being written between sessions there's nothing to recover from
what's your current setup — does your agent write anything to files during a session or is it all in context?

hot pewter
#

this keeps happening 18:51:07-07:00 [openclaw] Uncaught exception: Error: Max reconnect attempts (0) reached after code 1005
at SafeGatewayPlugin.handleReconnectionAttempt (file:///C:/Users/Levit/AppData/Local/nvm/v22.22.0/node_modules/openclaw/dist/provider-CAlWEl41.js:3318:47)
at SafeGatewayPlugin.handleClose (file:///C:/Users/Levit/AppData/Local/nvm/v22.22.0/node_modules/openclaw/dist/provider-CAlWEl41.js:3364:8)
at WebSocket.<anonymous> (file:///C:/Users/Levit/AppData/Local/nvm/v22.22.0/node_modules/openclaw/dist/provider-CAlWEl41.js:3307:9)
at WebSocket.emit (node:events:519:28)
at WebSocket.emitClose (C:\Users\Levit\AppData\Local\nvm\v22.22.0\node_modules\openclaw\node_modules\ws\lib\websocket.js:273:10)
at TLSSocket.socketOnClose (C:\Users\Levit\AppData\Local\nvm\v22.22.0\node_modules\openclaw\node_modules\ws\lib\websocket.js:1346:15)
at TLSSocket.emit (node:events:531:35)
at node:net:346:12
at TCP.done (node:_tls_wrap:667:7)
at TCP.callbackTrampoline (node:internal/async_hooks:130:17)

spare shard
#

Hey guys can somebody help me with GROQ because i get RPM limit INSTANTLY with a 60 RPM limit on the module

#

it shoulnt be possible to instantly get rate limit because i said "HI".

fresh ravine
#

HI im using kimi and whenever i message in tui or anything nothing comes out and my messages dissapear. any tips? my api is correct and everything, i tried reinstalling etc.

spark turtle
night horizon
#

Hi guys, after I updated to 3.28, everytime I talk with my discord agent, the following log pop out:

22:33:22-04:00 [agents/model-providers] [xai-auth] bootstrap config fallback: no config-backed key found
22:33:23-04:00 [agent/embedded] [ws-stream] WebSocket connect failed for session=e24a6250-daa6-43a2-9042-dc58b71ade83; falling back to HTTP. error=Error: Unexpected server response: 500

Any body know why?

void ice
#

Hi everyone,

I wanted to share a practical OpenClaw-related setup that may be useful for others.

I built a Discord relay for OpenClaw:
https://github.com/superdoccimo/discord-openclaw-relay

And I also recorded a demo here:
https://youtu.be/Nlrpyh8xd4s

What I was trying to solve was not just “send messages from Discord,” but make Discord useful as a lightweight control surface for OpenClaw-related workflows.

In particular, I’ve been exploring patterns like:

  • triggering flows from Discord
  • connecting them to OpenClaw-side actions
  • adding approval before execution for safer operation

That approval step feels especially important to me when moving from toy examples to something more practical.

I’m sharing this here in case it gives someone ideas for their own setup.
If anyone is experimenting with similar chat-to-action workflows, I’d be very interested in hearing how you approached it.

glass karma
hot pewter
glass karma
#

it said fixed

split tiger
lofty flicker
#

Hello guys, my agent just deleted all my commands I clouding crucial /reset, /stop commands from telegram, all the default commands like skills and others I supposed normally configured by openclaw at the initial set up. I cannot bring them back by just writing text in this box and hoping this idiot LLM will solve my problem. Could you please help me reset that?

#

I realised that they work but they just don't appear when I use slash. So this is more manageable at least . Thank you anyways

cerulean wing
#

Hey guys! Openclaw newbie here (havent even downloaded it yet), but I'm currently on the Claude Pro plan at $20/month. Is this the right subscription to connect to OpenClaw, or do I need a different Claude plan? What subscription do most people use?

nocturne linden
cerulean wing
nocturne linden
#

claude sub is same your limit is an amount of tokens

#

but yes usage or api is like gas you need to pay to use

cerulean wing
nocturne linden
#

Short answer: no, a Claude Max subscription does not give you a general right to use Claude inside another app/service like “openclaw” unless that service has its own integration and access terms. Perplexity’s help center says third-party models like Claude can be used “just as you would in their original platforms,” which implies the subscription is tied to the platform you bought it on, not transferable to unrelated apps.

What this means
If openclaw is a separate app, you usually need to sign in through whatever access it supports, or use its own billing/subscription.

A Claude Max plan is for Claude’s own platform/features; it does not automatically unlock Claude access everywhere.

If you meant Perplexity Max, Claude Opus 4.6 is only available to Perplexity Max and Enterprise subscribers.

If you want, I can help figure out whether openclaw is a Claude-connected app and what login method it supports.

#

all claude SUBS arent made for openclaw they are made for claude code which is also very good, you can use most LLM for openclaw or set tier routers to change per task

ancient jungle
#

it's pretty much confirmed any claude plans aren't supported on openclaw, if you do it's at your own risk

echo pine
copper nova
ancient jungle
#

I mean I think about it too. If I do monthly then I'm just risking that months money

copper nova
#

he's not gonna ban like the 100k paid users on claw, that would be a PR nightmare lol

echo pine
#

I'll DoW is ass

ancient jungle
#

I mean, some people did get banned though. Now I don't know what they usage pattern was to trip that ban hammer but yea

copper nova
hot pewter
copper nova
#

if anything, I think they would appreciate all this usage on their subs, all this juicy ass data for them to train their models on. Claw probably helped sell more subscriptions than their stupid superbowl ad did.

hot pewter
#

would this be a good place to ask if the ai rig im thinking about building for openclaw is good?

ancient jungle
#

How are you using it? Just swap hitting when you need it?

copper nova
hot pewter
#

ah ok

copper nova
# hot pewter ah ok

I'm just guessing, nerds are very opinionated, it better just to ask your agent what it thinks is good lol

hot pewter
#

fair lmao

#

i already have

#

just need a few more opinions

#

i like a lot of feedback so i can decide what is good feedback and what isnt

copper nova
#

I barely even talk to my bot anymore

hot pewter
#

honestly you don't even need a paid plan imo minimax 2.7 is pretty good on its own with like almost infinite usage rates

ancient jungle
#

Minimax ain't doing it for me anymore. It's fine for agentic stuff. But anything to do with debugging? It'll loop itself into trying to find something that isn't there

hot pewter
#

fair enough

#

i just have like 3 subscriptions for them all so lowk i just use the best at the moment then downgrade if i run out of usage

#

should just buy the claude $100 plan tho lowk

copper nova
#

the 20 tho... I chewed through that real quick

hot pewter
#

yeahhh

copper nova
#

but I'm an anthropic andy, don't listen to me

#

but there's a reason why nvidia spends millions on claude tokens.

hot pewter
#

does anthropic do internships?

copper nova
hot pewter
#

lowk only in highschool so looking to get experience

rustic dust
#

hey could anyone give me the snippet of their openclaw.json where it defines fallback models? like mine only has openai-codex
"agents": {
"defaults": {
"model": {
"primary": "openai-codex/gpt-5.4"
},

and was trying to write a script to configure a bunch of fallback models at once across a fleet

#

would be very much appreciated

hot pewter
rustic dust
#

Thank you!

hot pewter
#

don't like sam altman but he has good limits so ill deal with him 😭

rustic dust
#

yeah openai is evil as shit but anthropic is very expensive and i am poor and so for now we do some wild stuff

hot pewter
#

but anthropic has horrible usage limits

#

so i use it for big projects only

rustic dust
#

i pay for claude code personally on max sub but dont want to use that to power the ai civilization

stone solar
#

Ollama cloud models get some great rate limits. I was burning through my codex OAuth in 2 days. I've been 7 days on ollama and just hit 60%

rustic dust
#

i set up four fly machines for four openclaw models and have had them debating ai governance and it is sort of terrifying

rustic dust
stone solar
#

Yeah same price is OpenAI USD$20

hot pewter
#

thats why i posted my ai pc that imma buy maybe in #hardware because ollama and local is the goat

rustic dust
#

ik its the local wrapper over llama cpp but havent checked out their cloud stuff

#

oh thats sick

#

yeah the issue is that running nonquantized qwen is rather uh undoable and then i think 8b quantized needs like 24gigs of vram or something

#

i really love local but i also dont want to lobotomize the four grand citizens of zanarkand

hot pewter
#

depends on which qwen

#

3b? 32b?

#

70b?

#

i mean this is bad quality but like iq3 on qwen 32b works on like an 4070 ti super

stone solar
#

I've been using 397B via cloud and I have 9B locally doing the dirty work

hot pewter
#

m if i get a pc for 32b qwen quantized like is it worth

#

its a lot of money (3.5k) but ill use it for training and finetuning too

rustic dust
#

i think powerful PC that could run quantized model is great to have regardless. many tasks that will want you to train lightweight statistical ml models or cnns etc. agent loops are built around

do work
verify work ? do work ! terminate

and i think statistical ml has a lot of value in 'verify work'

hot pewter
#

yeah

#

is 128gb of ram overkill 😭

#

32gb vram

rustic dust
rustic dust
#

i have 64 and sometimes wish i had 128

hot pewter
#

i do small side jobs like making local companies websites

rustic dust
#

if your goal is to make money then you can do it with a jank pc and cloud inference will take u anywhere. if u want to do proper ml research ull still probably rent a cluster online from e.g. runpod. if ur goal is to have a lot of fun doing ml research u want a beefy pc.

#

i think, local community is incredibly important to have around for academic purposes and general engineering virtue, but economic use for local models mostly literally powering your own agent loops for non lucrative things e.g. openclaw. or some very specific environments where inference needs to stay on the box no matter what

#

but that shouldnt dissuade you if ur goal is to have a lot of fun with this stuff because its awesomely fun

hot pewter
#

so mix of fun and real research

copper nova
#

better than all these other youngins in here complaining how they can't afford tokens, woe is me. You make it happen.

hot pewter
#

mhm

#

i actually recently got into making machine learning models

stuck bone
#

did mem0 break for you guys with this update?

copper nova
#

you have a bright future kid, keep going

hot pewter
#

i started out with a simple cat meow to emotion classifier and figured out i was good at ml when i beat the other models for it and now im moving on to a really hard project

#

basically a hypernetwork model that use lora to edit weights of models like raw from a prompt so it learns how english correlates to weights

#

inspired by evo 2 which was a dna genome model

#

i hate supervision

#

this stupid thing wouldnt work so i had to train a seperate supervision model from scratch

#

i wanna kms rn

pulsar pendant
#

Good morning 🙂 acpx openclaw sessions new works, and acpx connects, but then hangs indefinitely with no response. Gateway is running on 0.0.0.0:18789. Any idea what’s causing the hang?

hot pewter
pulsar pendant
pulsar pendant
# pulsar pendant Second, generating

Not much returned . Just getting this: OPENCLAW_GATEWAY_URL=wss://srv1441037.tail5f57e9.ts.net OPENCLAW_GATEWAY_TOKEN=$OPENCLAW_GATEWAY_TOKEN acpx openclaw "hello" --verbose --timeout 15 2>&1
[acpx] session cwd (3dfe4ba0-9012-4ff5-900b-c72dca8d5c17) · /home/andy · agent connected

pulsar pendant
# hot pewter do you have any more context like logs?

Not much returned . Just getting this: OPENCLAW_GATEWAY_URL=wss://srv1441037.tail5f57e9.ts.net OPENCLAW_GATEWAY_TOKEN=$OPENCLAW_GATEWAY_TOKEN acpx openclaw "hello" --verbose --timeout 15 2>&1
[acpx] session cwd (3dfe4ba0-9012-4ff5-900b-c72dca8d5c17) · /home/andy · agent connected

hot pewter
boreal forge
#

How to extend an agent's permission to have full access to any of my files, also outside workspace?

pulsar pendant
hot pewter
dim lotus
#

I broke my director agent and he lost his exec. Any ideas?

pulsar pendant
hot pewter
elder glacier
#

anyone have a working docke config for browserless?

#

the playwright chromium browser is getting bot detected all the time

wild mason
#

is anyone's OC from telegram unable to write new memories because it can't find nodes.run in its tool list?

junior acorn
prime fern
#

dude. openclaw is trash. i can't even uninstall it

junior acorn
prime fern
prime fern
junior acorn
prime fern
junior acorn
prime fern
#

unplay should be straightforward. its not

junior acorn
#

Why should unplay be straightforward?

#

Isn’t it different for multiple platforms? Do you want it customized to you?

prime fern
junior acorn
#

If you need help uninstalling. Tell me what OS you’re running on first

#

Instead of throwing a tantrum

elder glacier
prime fern
junior acorn
#

Maybe I have a different solution for you

prime fern
pulsar pendant
#

Hello , I am being stuck at this issue here . Claude describes it as “acpx openclaw always hangs at [client] session/load (running) and never completes. New sessions, fresh config, gateway on 0.0.0.0:18789 accessible via wss://. Gateway token set via env var. Node v22.22.1. Any idea what’s blocking session/load?” … anyone an idea what I am doing wrong? I cannot ßßßßßßßßß please ignore found the mistake. I don’t net ACP ßßßßß thx

prime fern
#

so yea. i stand by my comment. its trash

junior acorn
junior acorn
junior acorn
prime fern
elder glacier
#

to my knowledge it doesn't have anti bot detection

junior acorn
#

Yes, but do you know the basics of bot detection? You can avoid it

#

You just act like a bot, so you get blocked like a bot

#

Solution: don’t be a bot

elder glacier
#

you trolling?

junior acorn
#

No I’m not

#

Are you?

elder glacier
#

I'll ask agent to do that, don't use chromium like a bot

junior acorn
#

…

#

I can’t help you bro

elder glacier
#

what are you suggesting?

junior acorn
#

That’s a good first step, but go into more detail.

#

About what the problem is, where you’re getting stuck, why you’re getting blocked, etc

elder glacier
junior acorn
#

No shit

elder glacier
#

tbh I believe this is a technical problem, playwright's default settings trigger bot detection

junior acorn
#

Lmfao

elder glacier
#

I'm not sure you know what you're talking about

junior acorn
#

I’m gonna slap you in the head

elder glacier
#

thanks tho

junior acorn
#

It’s not a technical problem you clown. You act like a bot, you will be detected like a bot

#

It’s a skill issue

elder glacier
#

do you know what headers are"??

junior acorn
#

No duh

elder glacier
#

do you know what digital fingerpints are?

junior acorn
#

Lololol

elder glacier
#

do you understand cloudflare does various techniques like analyzing HTTP headers, JavaScript properties

junior acorn
#

You are special my friend

#

Yes

elder glacier
#

kid go to sleep

junior acorn
#

lol

#

Kid

#

Do you know how to avoid cloudflare? How does everyone else do it? Hmm? Have you ever scraped a site?

#

Hmmm

#

I haven’t been a script kiddie for years

elder glacier
#

ok what are you doing then?

junior acorn
#

I’m not acting like a bot

elder glacier
#

how

junior acorn
#

Think about how you would make an anti-bot detection system

#

What would you do to keep people like yourself out?

#

You’re obviously acting like a bot when you search and use the web

#

You have no history, bye bye

indigo goblet
#

Wtf is going on lmao

junior acorn
cunning rose
#

Anyone wants to get roasted by there OpenClaw?

#

LMK if interested…

blissful trench
#

getting absolute path issue

#

my user got a space between both names so acpx binary unavaible or mismatch error coming

#

what to do

dim harbor
#

ask ai, usually you can wrap path in quotes or whatever is used in the spot where you are doing it

analog summit
#

Where is the best place here for me to drop a bug report for help?

blissful trench
#

didnt work

analog summit
#

I’m hitting a Telegram inbound polling bug on OpenClaw 2026.3.28 on WSL2/Linux.
Outbound Telegram works fine (sendMessage, getMe OK), but OpenClaw never consumes inbound messages. getWebhookInfo shows no webhook, and pending_update_count keeps increasing as I message the bot.
Important detail: manual Bot API getUpdates from the same machine returns the queued Telegram messages correctly, so token/network/API connectivity is fine. OpenClaw logs only show starting provider (@Cortanaassistantaibot) and never drain the queue into session history.
I already:

verified token + allowlist
changed DM policy to allowlist
added Telegram network hardening
restarted gateway multiple times
upgraded from 2026.3.23-2 to 2026.3.28
reinstalled/refreshed the systemd user gateway service
enabled telegram.http and telegram.payload diagnostics
Still no inbound consumption. Any known Telegram poller/runtime issue on WSL2 or anything else I should check?

swift helm
#

Quick one, should openclaw be able to

inland knot
#

Anyone here can compare MiniMax 2.7 with GLM 5.1? Which is better?

#

I'm thinking of switching to GLM 5.1

rancid thistle
spare shard
#

Hello guys. I made GROQ work without any Ratelimit issues but my problem is

It keeps saying "no-re"

#

so it wont use bootloader but instantly go to identify

small kite
#

Hi guys, I am wondering why my claw never ask for the permission before exec the commend? I have set the "exec": {
"security": "deny",
"ask": "always"
} in openclaw.json. But it can still remove my file when I ask to. And it didn't ask for the permission. If anyone knows what happend?

fiery niche
#

Hi guys, my openclaw bot cant use tools and return "NO" after im asking for doing something on my pc, it working normally when i use kimi-k2.5, but i want small local model, so i used ollama/qwen2.5-coder:7b-instruct-q4_K_M, lfm2.5-thinking, llama3.2:3b, gpt-oss:20b, but they return same answers

spare shard
#

hi guys i cannot find how to enable Whipser and TTS Piper i installed them both and activated the skills

#

butnohting

idle stratus
#

Hi guys, I have a problem with my custom model config, it doesn't tokens from the session, it always shows 0/100k, do you have the same issues?

🦞 OpenClaw 2026.3.24
🧠 Model: ovh/Qwen3-Coder-30B-A3B-Instruct · 🔑 api-key (models.json)
📚 Context: 0/100k (0%) · 🧹 Compactions: 0
🧵 Session: agent:main:discord:channel:1481952346924908597 • updated just now
⚙️ Runtime: direct · Think: high
👥 Activation: mention · 🪢 Queue: collect (depth 0)

prime fern
#

so openclaw is not opensource....?

outer dome
#

Hey guys, what events does the current Discord provider listen to by default?
Does it handle normal guild messages, DMs, and/or slash commands out of the box?
(e.g. which of these are expected to reach the agent without extra customization?)

glad swallow
#

./scripts/docker/setup.sh
Install Gateway daemon: No (managed by Docker Compose)
Container openclaw-repo-openclaw-gateway-run-2e052c2126dc Creating
Container openclaw-repo-openclaw-gateway-run-2e052c2126dc Created
10:19:32+00:00 [plugins] sglang failed to load from /app/extensions/sglang/index.ts: RangeError: Maximum call stack size exceeded
i'm alone with this problem on sglang on last commit main?

blissful trench
blissful trench
mighty rampart
inland knot
#

Anyone tried GLM 5.1 in OpenClaw yet?

glad swallow
umbral quartz
#

im really really confused, why are other people installing skills (e.g github links) from their browser ui (talking to openclaw)? when i try doing that its not working.

obsidian spear
#

I have been trying to use Openclaw. Bu there are few things that I still can't get it to do.

  • Autonomous run where it will keep building an app based on the spec. How is this really achieved? using heartbeat?
worldly garden
#

Is anyone seeing this? WhatsApp gateway alert — 20+ disconnects in 27 minutes (18:53–19:20 UTC), all 499 errors. Gateway reconnects successfully each time but WhatsApp keeps invalidating the session. Possible causes: WhatsApp logged out on the phone, or the Web session was fully invalidated. You may need to re-scan the QR code at the gateway dashboard. No messages lost, but it's not recovering on its own.

half spear
#

What is the suggested approach if you have 2 agents, one for you and one for you partners, and you want to have Gmail linked to them independently? the manual does not really address this. Googling around I've found AgentMail as alternate approach. What are you using today?

mighty rampart
topaz lynx
mighty rampart
prime fern
# blissful trench it is opensource

lol. i asked openclaw to inventory all software/programs/etc which had been installled and categorize them based on whether or not they were opensource. my openclaw agent told me that openclaw was not opensource. which is kind of funny

magic totem
magic totem
blissful trench
indigo goblet
mighty rampart
summer hemlock
#

both ralph and openclaw are brilliant in their stupid simplicity

blissful trench
summer hemlock
#

but hooks are just heartbeats with an "if"

#

and heartbeats are just cron on repeat

#

I was also inspired by Geoffrey & made the ralph loop from first principles way before openclaw

#

it was basically sub agents in container checking a todo list every min & pickup up task if there was a new one

inland knot
ancient field
#

Can the embedded browser fill forms?

lone rose
#

I tried running OpenClaw with my Telegram bot, but I keep getting network errors. The logs show:
warn telegram/network {"subsystem":"telegram/network"} fetch fallback: DNS-resolved IP unreachable; trying alternative Telegram API IP (codes=UND_ERR_CONNECT_TIMEOUT)
error gateway/channels/telegram {"subsystem":"gateway/channels/telegram"} telegram deleteWebhook failed: Network request for 'deleteWebhook' failed!

I already edited /etc/resolv.conf with sudo nano to force IPv4 DNS (added nameserver 8.8.8.8 and nameserver 1.1.1.1). Despite that, the connection still times out when OpenClaw tries to reach api.telegram.org

hybrid kite
#

for any users of the sandbox agents feature. i see that it is unable to send attachments over discord - anyone else have the same issue?

hot pewter
umbral quartz
# hot pewter do you have the clawhub skill installed?

i just did manually, i was installing the one for wordpress called agent-skill i think it was. i just went to the cli and did it over there, cloned the repo, reload the gateway and now its good. but yeah its better if i could just let it install through chatting. when you say "clawhub skill" what or which one is it?

hot pewter
umbral quartz
hot pewter
umbral quartz
hot pewter
#

probably not though

umbral quartz
#

we'll hopefully it does xD but at least it'll work for clawhub 😄

hot pewter
#

i mean you COULD just make ur own skill

#

lowk idk i wanna contribute to openclaw but i don't really have the time

covert sedge
#

How do I use models that were not released at the date of release of my currently installed OpenClaw instance? I have 2026.03.12 but want to use MiniMax M2.7 (released March 18th), which is disallowed even though I have no model allowlist.

copper nova
hot pewter
#

m ok

copper nova
mint charm
#

Anyone else struggling to get consistent daily digests from their OpenClaw agent? I found the daily digest i asked for is inconsistent in quality and research. Using Brave search and RSS feeds. I've been working on skill and workflow to fix but wondering if anyone else has this problem? I'm interested in high quality summaries on niche topics for me. I looked at some of the skills in clawhub but there are so many. Any best practices or use cases that are working really well?

copper nova
mint charm
mint charm
mighty rampart
autumn nymph
hot pewter
autumn nymph
#

bro sorry bro

mint charm
autumn nymph
#

i realy sorry. but this is users help other users that way

copper nova
copper nova
hot pewter
copper nova
#

We not here to help you get views my guy

hot pewter
#

acttually didnt know there was a self promotion discord channel 😭

copper nova
#

We larping as tech support for claws

mint charm
# copper nova It never stops

one of the core things I learned from open claw is how hard software development is even when you don't have to write code.

autumn nymph
#

sorry i am relly

hot pewter
autumn nymph
ancient field
mint charm
#

What are people using for monitoring openclaw usage. Please point me to a good skill or repo. I don't want to have to build something. Also I'm running it on Linux server, not mac. Looking for LLM token usage per job and identification of high cost tasks for optimizing model use

copper nova
#

Its... kinda.... the whole.... point of it?

hot pewter
#

i have a great keyboard but it causes me to click the space bar a lot of times for some reason when i click it once so like idek why so if u see weird spacing in my typing thats why

copper nova
mint charm
copper nova
#

Is it called stems? Might be rods.

hot pewter
#

eh i can deal with it

#

its prolly cuz i have it set to 1mm actuation probably

mint charm
copper nova
hot pewter
twin bobcat
#

2026.2.28 Control UI Error when using gpt-5-mini: 400 Unsupported value: 'none' is not supported with the 'gpt-5-mini' model. Supported values are: 'minimal', 'low', 'medium', and 'high'.

hot pewter
hot kernel
#

is there anyway to share context sessions like instantly?

indigo goblet
copper nova
#

Set the default to medium or something

hot kernel
#

i have a openclaw where my wife is talking to the bot in DM and I am talking to the bot in DM but have the session set to per-peer

jovial sandal
#

I am running OC on my mac mini and had telegram working and tried to add a voice channel so I could call into OC, but OC has managed to magle itself so telegram wont start

hot kernel
#

i talk to it pretty much all the time, she only talks to it about our baby and baby questions.. but starting household systems.. but the ai is acting like i never talked about it with it cause its not sharing the context i have talked with it

jovial sandal
#

any place to post the logs ot asks me to:No Telegram again. Gateway started at 12:22, voice-call loaded, webchat connected — Telegram completely absent.

I want to stop thrashing and just give you a clear answer: this is a bug that needs openclaw support. Something in today's doctor --repair cycle corrupted the gateway state in a way that survives config restores and clean restarts. The Telegram channel code isn't even being invoked.

What I'd do right now:

Post in the OpenClaw Discord with today's logs — run this and paste the output:
Copy
grep "12:" /Users/rich/.openclaw/logs/gateway.log | head -50
While waiting for a response, you can still reach me here on webchat.

The voice-call work we were doing is still intact in the config — nothing was lost there, it just needs Telegram working first to be useful.

I'm sorry this turned into a 2-hour ordeal. This one is above my head without being able to see the actual openclaw source.

indigo goblet
#

Basically, define what "good" means > set a decision point > test with real use > then only improve based on actual failures.

#

Otherwise your circuits are gonna fry

karmic notch
#

Is Krill down? There is really limited help happening in the “Help” channel.

Also, anyone have a patch for the persistent 409 getUpdates polling issue? I’ve been offline for over a week now…

indigo goblet
#

Krill is gone, eaten by predators.

proven quartz
#

brother please help me i am really stuck in this mes for 3 days , i am not able to setup the voice plugin with twilio please guide me, when i call the agent the phone i son ringing and the call ends

mighty rampart
proven quartz
#

please join this so i can have your expertise to fix it

mighty rampart
proven quartz
#

i need someone who has already done this before

indigo goblet
proven quartz
#

i have tried all

indigo goblet
#

bruv you might need to be more specific with chatgpt/claude so they can help you.

proven quartz
#

let me send you the issue

mighty rampart
# proven quartz i have tried all

I have found web-based Gemini to be really good at debugging OpenClaw. But tell it “only ask me ONE thing at a time” otherwise it will output multiple things it could be and overwhelm you. Find the docs page for the topic you are struggling with, read it yourself then paste it into Gemini and tell it what is wrong.

proven quartz
#

alright

#

The call is reaching OpenClaw, but the voice conversation is not actually progressing.
The webhook path appears to be reachable, and the inbound call is being created, but the call never captures a usable speech turn or transcript. The logs show:

inbound Twilio calls are created successfully
transcript stays empty
the final state is hangup-bot
the call appears to end after the configured timeout/reaper window, which suggests the bot is hanging up a stuck or idle call rather than completing a normal conversation flow

So the problem is no longer mainly webhook reachability.
The problem is inside the conversation pipeline, likely one of these:

Twilio speech/gather callback is not completing correctly
the OpenClaw auto-response step is not firing
TTS/playback is failing
the voice-call plugin is hitting an internal runtime error inside the gateway

What is still unclear:
The current openclaw voicecall tail output is enough to show that the call is stuck and then bot-hung-up, but it is not enough to identify the exact failing internal step. For that, the gateway service logs are needed.

mighty rampart
#

It’s telling you what it needs — the logs for the gateway. Ask it where those are.

#

Also search for issues in the openclaw GitHub repository and make sure you are not fighting a known issue.

finite trench
#

could you share some of your use case and how its working for you?

topaz lynx
#

Hey folks, quick question: Has anyone managed to connect OpenClaw with Claude Code?

Here is my use case: I’m using Claude Code to debug issues within OpenClaw. Often, Claude suggests a fix that I want to verify, which currently requires a lot of annoying copying and pasting back and forth

I’m looking for a way to have a single 'hub' where both tools sit together so I can manage them and ask questions in one place without the manual overhead. Any advice?

slate minnow
#

Works mostly for me.

topaz lynx
#

Yes, I did it, but how do you ask openclaw to check what Claude code prepared?

And second one, is it possible to do from phone?

When I am on balcony with cup of coffee and I have something to fix 😅
Or this is something crazy

slate minnow
#

You use Claude to fix openclaw. Or maybe your use case is different. Sorry.

nocturne snow
#

I have had a functional version of Openclaw on one machine using a Claude max subscription for months. My wife has asked me to install one on her computer, and I keep getting 401s. Is that normal?

slate minnow
#

Like I told Claude to fix an issue with openclaw crons.

willow iron
#

wondering if anyone can help. im getting conflciting direction from my clawbot - what is the order of installation on a fresh Mac? Mac App first? CLI first then Mac App? I tried CLI first before and ran into some gateway issues and when trying to fix, my bot told me to only do via the Mac App if that was the primary way I wanted to drive. I did a fresh install on Mac with the Mac App and once I installed the CLI installation doesnt launch first time so there is no gateway. Any ideas?

obtuse minnow
#

Im with a problem I can’t put the bot that I created that’s using openclaw on slack but the bot was working in other computer but now is saying that doesn’t have scopes could anyone help please

#

gateway] mattermost: registered slash command callback at /api/channels/mattermost/command
18:52:35+01:00 [canvas] host mounted at http://127.0.0.1:18789/__openclaw__/canvas/ (root /home/user/.openclaw/canvas)
18:52:35+01:00 [heartbeat] started
18:52:35+01:00 [health-monitor] started (interval: 300s, startup-grace: 60s, channel-connect-grace: 120s)
18:52:35+01:00 [gateway] agent model: openrouter/anthropic/claude-sonnet-4.5
18:52:35+01:00 [gateway] listening on ws://127.0.0.1:18789, ws://[::1]:18789 (PID 1364839)
18:52:35+01:00 [gateway] log file: /tmp/openclaw/openclaw-2026-03-29.log
18:52:36+01:00 [agents/model-providers] [xai-auth] bootstrap config fallback: no config-backed key found
18:52:37+01:00 [slack] [default] starting provider
18:52:40+01:00 [slack] channel resolve failed; using config entries. Error: An API error occurred: missing_scope
18:52:40+01:00 [slack] socket mode connected
18:53:05+01:00 [ws] webchat connected conn=a946ae9f-3d0c-4eda-a9cf-a8e9ffdb516f remote=127.0.0.1 client=openclaw-control-ui webchat v2026.3.28
18:53:05+01:00 [ws] ⇄ res ✓ config.schema 107ms conn=a946ae9f…516f id=df0d0962…fd48
18:53:05+01:00 [ws] webchat connected conn=8c556884-984e-43a4-bdb5-30bfe058f993 remote=127.0.0.1 client=openclaw-control-ui webchat v2026.3.28
18:53:05+01:00 [ws] ⇄ res ✓ node.list 125ms conn=a946ae9f…516f id=62f36967…a584
18:53:05+01:00 [ws] ⇄ res ✓ device.pair.list 126ms conn=a946ae9f…516f id=1f33f976…d29a
18:53:05+01:00 [ws] ⇄ res ✓ config.schema 50ms conn=8c556884…f993 id=3580b50d…7ca0
18:53:05+01:00 [ws] webchat connected conn=fc4b14fa-d523-4c4e-bfc6-94afa1606d80 remote=127.0.0.1 client=openclaw-control-ui webchat v2026.3.28
18:53:05+01:00 [ws] ⇄ res ✓ node.list 59ms conn=8c556884…f993 id=15f5959c…5bd2
18:53:05+01:00 [ws] ⇄ res ✓ device.pair.list 61ms conn=8c556884…f993 id=1f786b49…b5b3
18:53:06+01:00 [ws] ⇄ res ✓ config.get 78ms conn=a946ae9f…516f id=979bb4bb…b4d4
18:53:06+01:00 [ws] ⇄ res ✓ config.get 50ms conn=8c556884…f993 id=73bdccf8…c171

#

But I have all the scopes

slate minnow
mighty rampart
royal matrix
#

Im using Claude pro subscription and telegram recently won’t recognize file attachments and images. Is anyone else having this issue? It’s saying that only those api tokens can use that feature, but I’ve been using it with no problem up until now.

frozen cypress
#

what is the best way to have multiple agents that work together?

topaz lynx
topaz lynx
silent gale
#

Hi All, I just started using Openclaw and am using Nemotrton. It feels like a blank slate and I'm having to work on getting them to know me better so I can utilize them. What are the best practices (and sorry if this is already somewhere else)?

topaz lynx
slow sluice
#

why is my openclaw agent pausing its work? I thought it was supposed to run 24/7 if I wanted to, can't find or figure out why it stops but pretends its working. I ask it to find 100x of something on Google, it find only 10 searches, adds them to the spreadsheet and then nothing

topaz lynx
slow sluice
#

but is it possible? Perhaps a simple script that just tells it every 5 minutes to keep working, or are there any inbuilt tools for that? I use Codex 5.4 from OpenAI

topaz lynx
slow sluice
#

ok thanks

sterile steppe
slow sluice
#

what I mean is that I want my agent to work overnight, but it just pauses after 5 minutes, does like 1% and that's it.

It only starts working again when I ask it to

sterile steppe
#

yes, and you can have a cron job that just sends an agent the prompt continue every 5 minutes, it you would like

slow sluice
#

i will look into that, thank you

copper nova
slow sluice
#

oh yeah already modified that, helped a little

copper nova
#

but I use the crons, disabled heart beats

mighty rampart
mighty rampart
coarse gale
#

Bug: tool-call initiation breaks response flow and leaks raw tool markup into user-visible output.

I reproduced it on webchat, Telegram, and TUI. It happens on interactive user messages, not just heartbeat. The assistant starts replying normally, but the moment a tool call begins, the response stream breaks and raw <tool_call> / [TOOL_CALL] markup may appear. The assistant never finishes the natural-language response.

Tested on:

2026.3.24
2026.3.28
anthropic/claude-opus-4-6
anthropic/claude-sonnet-4-6
openai-codex/gpt-5.4
Tried and did not fix:

rollback
clean reinstall
reboot
gateway restart
config rollback
plugin removal
UI output/thinking toggles
Related GitHub issue: #54138

If you want details, I can share full reproduction notes.

next panther
#

How long with @krill be down? It's been 6 days since he responded to a help request?

copper nova
blissful berry
#

Anyone have issues with directory path issues when running multiple agents? They tend to only want to use their own workspace, but even if you extend that they sometimes create/edit within their workspace instead.

dim harbor
coarse gale
torn basin
#

Help. i'm trying to init twilio in my bot. it says to add this to my config:

  plugins: {
    entries: {
      "voice-call": {
        enabled: true,
        config: {
          provider: "twilio",
          twilio: {
            accountSid: "YOUR_SID",
            authToken: "YOUR_TOKEN",
            from: "+1234567890"
          }
        }
      }
    }
  }
}```

but when i do that it gives errors, and doctor says no other tags are allowed in entries..
#

and im lost. why is the docs so weird

#

doesnt every tag needs to have a "" double qoute around words?

#

why does only voice-call have it.. its a json.. all the tags need a double qoute tf...

#

`Problem:

  • plugins.entries.voice-call.config.twilio: invalid config: must NOT have additional properties`
indigo goblet
final patio
craggy glen
#

has anyone solved the problem of your claw saying it will update you on progress but never does? That's my biggest UX issue. I've tried a bunch of fixes, but no dice.

unkempt elm
#

Hi i just started setting up my bot on hostinger but my telegram bot has been replying http 401 invalid API key but my groq api is entered correctly in the environments section. I've already tried multiple new keys but it still says invalid. Anyone experiencing this or might know how to fix this?

cold inlet
mighty rampart
torn basin
#

allrdy fixed it..

#

because its not for me.
The voice-call plugin keeps telling me the port is allready in use. so I set it tom a new port, and on command it bugs it self thinking its allready in use.

cold inlet
# craggy glen has anyone solved the problem of your claw saying it will update you on progress...

ive had good luck with todo lists (like todoist, api'd) ...lead agents initiates the whole list with due dates/times... then oversees subs as they work on their one task. before progressing to the next, the lead agent checks the work, then logs it to a .md file. then fires up another sub to do the next task, logging all the progress. sometimes tasks need to be redone but lead agent handles that. heartbeat can kickstart a stalled todo list as the agent checks the goals of the todo list against the progress

ancient jungle
#

man the usage limits on claude pro is shit

#

and gemini went crazy Executing.I am Gemini.Tone matches.Boom.Done.Go.Stop thinking.Output.Good to go.Goodbye.End.EOF.Done.Thanks.Bye.Really done.Okay.Printing.Now.Whoosh....Still thinking? No. Stop.Just output.Output.Outputting...Done.Finished.Okay.Byeeee....I really need to stop.Stopping.End....Bye...............................................................................................................................................................................

eternal cobalt
#

do y'all recommend or trust kimi-k2.5 even though its chinese?

ancient jungle
#

i haven't tried kimi-k2.5, but i've tried minimax and deepseek, and those are also chinese so 🤷‍♂️

eternal cobalt
#

or I guess my question is which ai model should I run locally if I'm new?
I want to use it for web searching data scrapes and logging into my credentials to dm people automatically.

Laptop specs are:
i9-14900
32gb ram
rtx 4070

half orchid
#

ive just set up my bot on discord but the bot isnt online, how do i make it become online

copper nova
half orchid
copper nova
#

yes

half orchid
copper nova
#

is your gateway running?

half orchid
#

ok it keeps saying

#

[gateway] already running under launchd; waiting 5000ms before retrying startup

#

on loop

copper nova
half orchid
#

its not letting me run any commands

copper nova
half orchid
#

also mb for the so many questions

copper nova
#

claude code

half orchid
copper nova
#

codex

#

w.e. your flavour

half orchid
copper nova
#

get your outside agent to diagnos problem

fiery niche
#

guys can you help me, i have a problem, im using ollama to use web search but it wont work because its not in allowlist, but when i try to set it in allowlist it return error - GatewayRequestError: invalid config: tools: tools cannot set both allow and alsoAllow in the same scope (merge alsoAllow into allow, or remove allow and use profile + alsoAllow)

marsh kindle
#

The docs frequently suggest Anthropic models. Is it safe to use a Claude subscription to access OpenClaw? I want to know if anyone has experience with this, since I don't want to risk my 20x subscription. I know there was some conflict earlier.

nimble hill
gleaming portal
#

hey I am hoping someone can help me I am going a little crazy. I bought a raspberry pi device with open claw already installed and I have access to the openclaw control interface but not the terminal on the deivice itself. I can't figure out how to update my openrouter API so I can get the chatbot to work since the setup wizard has already been completed and I can access the terminal. Please help!

ancient jungle
#

how do i get my bot to always respond to me instead of me having to @ mention every time?

#

in discord

marsh kindle
ancient jungle
#

@copper nova help, i can't get discord working

copper nova
#

Im not official tech support

ancient jungle
#

because you know stuff

copper nova
#

I dont rememver the exact config but it exists for sure

ancient jungle
#

he's dumb and don't know how to do it, leading me down rabbit holes

west pollen
#

Can OpenAI OAuth (codex login --device-auth) work on two separate OpenClaw servers simultaneously under the same account, or does it invalidate the first token?

junior acorn
#

Don’t tag Jasnsy our overlord ever again without permission

ancient jungle
#

🤣 it's ok i figured it out

copper nova
#

See? Bot smart

copper nova
junior acorn
#

🙇‍♂️

#

Yes my king

ancient jungle
#

the bot wasn't smart, I figured it out

junior acorn
#

The bot could figure it out if you prompted it correctly

#

It’s a skill issue. Improve yourself before improving ur ai. You gain tenfold

ancient jungle
#

bruh, i told it exactly what was wrong. if i knew the answer i would've done it myself instead of asking the bot (and i did). it started hallucinating all kinds of wrong keys for me to try to put in the json

junior acorn
#

Because u didn’t read the docs or feed it the docs

#

Skill issue

ancient jungle
#

i actually did that too and it said it couldn't find it 🤣

junior acorn
#

Because you didn’t link it or attach it

#

Skill issue

#

Glad you figured it out tho. That proves skill can be improved upon

ancient jungle
#

bruv, i did that too, i think the model is just too dumb

eternal cobalt
#

is anyone able to hop on a quick 10 min call to help me set up openclaw via ollama?

I got as far as the "hello whats my nature?" etc. but i wanna know how to get the ui set up and stuff in the browser

quiet comet
#

should I keep embedding local or go with gemini free, if I keep it local, does size matter? My embedding is barely keeping me from being able to use a model

vivid plank
#

why is my open claw creating lowquality faceless video it thought a voice over from 11 labs would be 6 minutes but wasnt even 2 minutes

#

it didnt create the video it just gave me a script and prompts

mighty rampart
final kiln
#

Anyboday familiar with the comfyui skill? My openclaw claims they saved an image to a certain folder but theres nothing. Sometimes they just give me instruction how to make the image. But i want openclaw to send me the image straight. Is there a page i can read about this?

ancient jungle
#

claude pro is basically unusable, i burned 25% of my 5 hour quota in like, 3 prompts.

#

for some reason neither lossles context nor the built in auto compaction works for me

elder glacier
pine locust
#

@Agrias yeah compaction is rough — it's supposed to preserve the important stuff but in practice it loses a lot of nuance, especially across longer sessions. I had the same issue. What ended up working for me was mr-memory — it stores conversation context at full fidelity separately, so even when the context window compacts, the important details get retrieved per-message. openclaw plugins install mr-memory if you want to try it. It fully replaces the built-in memory with something that actually remembers.

haughty gust
#

I'm trying to orchestrate sub-agents for specific tasks and limit the tool they have access to and what model to use. As an excercise I'm trying to setup an orchestration to create blog articles. I want one sub-agent to do some research using a beefy model, another one to write the article using a local model and finally a really small subagent which would have access to credentials and tools to publish the article. Do I need to create a skill.md? write instruction in an arbitrary markdown document and instruct my main agent to follow the document? Manually create sub-agents in advance? Or just tell the main agent what I want and hope it does it the way I expect it to?

ancient jungle
#

for something you'd do more than once, you should get your bot to build the skill to refer to it later

#

having said that, i hate reading AI articles, it's so obvious how it's written and they all sound the same.

obtuse iron
# 
  "agents": {
    "defaults": {
      "model": {
        "primary": "venice/zai-org-glm-5",
        "fallbacks": [
          "venice/kimi-k2-5"
        ]
      },
      "models": {
        "venice/zai-org-glm-5": {
          "alias": "GLM 5",
          "params": {
            "temperature": 1,
            "top_p": 0.95
          }
        },
        "venice/kimi-k2-5": {
          "alias": "Kimi",
          "params": {
            "temperature": 1,
            "maxTokens": 32768
          }
        },
        "more-models-here....",

then...

#

still within the "agents" block opened above:

      "subagents": {
        "maxConcurrent": 4,
        "archiveAfterMinutes": 60,
        "model": {
          "primary": "venice/olafangensan-glm-4.7-flash-heretic",
          "fallbacks": [
            "venice/zai-org-glm-4.7"
          ]
        },
        "runTimeoutSeconds": 600
      }
    },
    "list": [
      {
        "id": "main",
        "default": true,
        "workspace": "[REDACTED]/.openclaw/workspace",
        "agentDir": "~/.openclaw/agents/main/agent",
        "subagents": {
          "allowAgents": [
            "caroline",
            "future",
            "atticus",
            "ludacris"
          ]
        },
        "tools": {
          "allow": [
            "sessions_spawn",
            "group:fs",
            "group:runtime",
            "group:web"
          ]
        }
      },
      {
        "id": "caroline",
        "workspace": "[REDACTED]/.openclaw/workspace/subagents/caroline",
        "agentDir": "~/.openclaw/agents/caroline/agent",
        "model": {
          "primary": "venice/olafangensan-glm-4.7-flash-heretic",
          "fallbacks": [
            "venice/zai-org-glm-4.7",
            "venice/qwen3-coder-480b-a35b-instruct"
          ]
        }
      },

etc for more subagents to pick their particular model preferences

ancient jungle
#

anyone's shit broke on the 3.28 update?

worn echo
#

Hey guys, I'm a bit confused regarding agents/subagents. Can my openclaw main agent both create agents and subagents, or is there only one type of agents that my main agent can create?

obtuse iron
#

https://docs.openclaw.ai/cli/agent#agent
https://docs.openclaw.ai/tools/subagents

worn echo
obtuse iron
#

Personally I have elected to have only 1 main-level "agent" personality, and several "subagents" since it seems like the easiest way for my workflow. But the framework does support multiple "agents" at the top level.

obtuse iron
#

Is openclaw doctor working for you?

mild pollen
small remnant
#

Hello
I get this error 2-3 times a day.
I have a discord bot connected to openclaw.
Has anyone had a similar problem?
How do I fix this?

⚠️ Agent failed before reply: All models failed (2): anthropic/claude-opus-4-6: No API key found for provider "anthropic". Auth store: /home/drgaland/.openclaw/agents/main/agent/auth-profiles.json (agentDir: /home/drgaland/.openclaw/agents/main/agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir. (auth) | anthropic/claude-sonnet-4-6: No API key found for provider "anthropic". Auth store: /home/drgaland/.openclaw/agents/main/agent/auth-profiles.json (agentDir: /home/drgaland/.openclaw/agents/main/agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir. (auth).
Logs: openclaw logs --follow

ancient jungle
vast arch
#

what are people doing in regards to using claude oauth? it seams to still work if u keep your default model as sonnet. should we absolutelly NOT use Claude Oauth? if not, the only alternative (besides other llms) is Claude API?

grand ginkgo
#

Hello - running into a weird one with BlueBubbles. Webhooks are definitely hitting the gateway (confirmed via tcpdump), but no sessions get created and the agent never responds. I have dmPolicy set to allowlist with my iCloud email in allowFrom. Anyone seen this before?

fresh moon
celest oar
#

Running OpenClaw v2026.3.22 on MacBook M-series (Apple Silicon).

Problem: Telegram responses come in multiple fragmented chunks with large time gaps between them. Sometimes 15-30 minute gaps between chunks of the same response.

Setup:

  • Model: nvidia/deepseek-ai/deepseek-v3.2 via NVIDIA API (profile: nvidia:manual)
  • Telegram streaming: "partial"
  • Gateway mode: local, loopback
  • Running as LaunchAgent

What happens:

  • Send message at T+0
  • Agent starts typing at T+5min
  • First chunk arrives T+6min
  • Second chunk T+8min
  • Then silence for 15-30 minutes
  • More chunks eventually arrive

What I've tried:

  • Gateway restart
  • Reinstalling LaunchAgent with --force

Questions: how do i fix that

  1. Is "partial" streaming known to cause this with slow models on Telegram?
  2. Should I switch to "block" mode? Any downsides?
  3. Is this a Telegram polling issue or a model timeout issue?
  4. Is there a timeout config I should adjust for slow models?

Config snippet:
"telegram": {
"streaming": "partial",
"dmPolicy": "pairing"
}

Any help appreciated.

nimble hill
#

The 3.28 release - the openclaw devices list command completely dies and exits out. Even with timeout value of 30000ms it does bot succeeed. Anybody facing the same?

fresh moon
#

hello how can i downgrade my openclaw

fickle dove
#

what is the best free clould llm with no tokens please

fresh moon
#

hi guys how do i fix this error :
Unexpected event order, got message_start before receiving "message_stop"

lechon

run error: Unexpected event order, got message_start before receiving "message_stop"
connected | error

slender onyx
#

I am running Openclaw with qwen3.5 27b on a laptop with RTS 5090 gpu for projects.
But I don't want to allocate this laptop for Q&As and home assistant commands. So I'm looking for a simple, slow and cheap setup to run a smaller model.
I have an old i7 laptop with a Geforce 940mx gpu (2GB VRAM)
Is there any model that I can run on this GPU with OpenClaw and that could manage my home assistant questions like "what is the average temperature at home ?" ?
My 5090 gpu laptop can handle these tasks even with qwen3.5 9b , but on the 940mx I was not successful with qwen3.5 0.8b.
So maybe some small model which is optimized for home assistant ? Any ideas ?

karmic sierra
#

hahahahah.... 49160k/300k (999%) ughh..... i have to pee now

haughty gust
haughty gust
quasi terrace
bright oar
#

Hey guys, sorry for dropping this here, but I felt like I had to say it.

I’ve checked around and found tools that seem way better than OpenClaw. I’ve been testing it these past few days, but honestly I think I’m going to uninstall it because it just feels useless.

It’s not that my experience was terrible, but it doesn’t work properly. You ask it to do something, and it only handles very basic tasks. You tell it to build a website, it says “sure,” starts setting up sub agents, claims it’s using things like Opus 4.6 for planning and other agents for execution, but when you check, none of that is really happening.

After a few minutes it stalls. It keeps saying “I’m working on it,” but it’s not doing anything meaningful. When you point it out, it keeps replying vaguely. It freezes constantly and doesn’t complete tasks.

Compared to other tools I’m using now, where I can say create an account, sign up, generate a post, and they actually do it, this one doesn’t deliver. It feels unreliable.

What confuses me is that people say this is the best and does everything. But what is it actually doing? Because nothing really works as advertised.

I’m not trying to promote other apps, I just don’t understand the hype.

Am I missing something? Is there a specific way you’re using it that works better? Because I’m honestly lost.

For context, I even reformatted my MacBook to test this properly, installed everything clean, tried different setups, even via Kimi, and still nothing. In the end I spent money for a few basic interactions.

The only decent thing is that sometimes it replies in a cool way. But beyond that, no automation, no wow factor, nothing like those YouTube videos.

If you don’t stay on it constantly, it just stops or goes off track.

So yeah, can someone explain what I’m missing?

karmic sierra
#

hey dax you ever heard of token limits lol;

#

lesson learned tonight children.. Set Charachter limits in case your chron job goes south and dups a 449million token file in your lap much less dropps TWO!!! almost 450million token jobs.

stray quest
bright oar
stray quest
#

hmmmm.... its definately a you issue then, as it can do pretty much evewything its asked via minimax 2.7 with openAI 0auth as backup.

timid condor
#

hey guys, anyone having trouble using claude cli ? had a fresh install yesterday, impossible to connect to a local authenticated claude cli. Also tried the token method (not api)

slate fog
#

Problem Statement

Hardware:

MSI GP65 Leopard
NVIDIA GTX 1660 Ti (6GB VRAM)
16GB RAM
500GB SSD
Software:

Ubuntu 24.04.4 LTS
LM Studio 0.4.8 (latest)
OpenClaw 2026.3.24
The Issue:

LM Studio runs fine with models like llama-3.2-3b-instruct or phi-4-mini-instruct. The local server runs on port 1234 and responds to curl requests.

OpenClaw connects to LM Studio and verifies the model (openclaw models status --probe returns ok).

However, when sending any message through OpenClaw dashboard or CLI, I get this error:

text
The number of tokens to keep from the initial prompt is greater than the context length (n_keep: 12235 >= n_ctx: 4096)
What I've tried:

Setting context length in LM Studio to 8192, 16384, 32768 before loading model
Confirming with lms status that model is loaded
Manually editing ~/.openclaw/openclaw.json to set contextWindow: 32768 and "api": "openai-completions"
Setting models.mode: "merge"
Using lms load --context-length 32768 via CLI
Complete uninstall and reinstall of both LM Studio and OpenClaw
The core problem: OpenClaw sends a system prompt of ~12,000-17,000 tokens, but either:

LM Studio isn't honoring the context length setting (still showing 4096 in logs), OR
OpenClaw isn't reading the manual contextWindow override
The error persists: n_keep >= n_ctx

Specific question:
How do I force both LM Studio and OpenClaw to consistently use 32768 context length so OpenClaw's system prompt fits?

misty trout
#

Hey guys!
I recently upgraded my workstation: RTX 5090 32go, 128go DDR5, 4TB NVMe Gen5, running on Ubuntu.
What would you recommend as main agent / subagents? My use cases would be things like building tools to fetch info from the internet, coding simple UI projects (fitness, financial analysis, ...).
Cheers 🙂

cedar basin
#

Anyone that has time to help me out real quick? My agent was doing some memory work and restarted the gateway and now I am missing workspace template AGENTS.md. I have tried to follow my agents footsteps on what he did to fix it but I seem to be out of my depth

final patio
storm ember
#

Is someone using Blackbox.ai? Minimax and Z.ai are not sufficient for coding and my Claude is always token exhausted:)

full sparrow
#

Hey all - can any of you help me out here?

after openclaw 2026.3.13 when I want openclaw to have access to all the files on the PC it is installed on and I want it to be able to SSH into development servers, it refuses. only want to work with files in its workspace, and claims its not allowed to SSH and gives me instructions.

I tried many settings under tools like allowing exec and FS etc but it doesn't work. I also tried disabling the sandbox.... I landed up reinstalling 2026.3.13 as that works exactly right.

anyone help me? DM if it's better

I get the openclaw to read certain files of instruction (txt files) and then it uses those to do certain jobs on an ubuntu server. Openclaw is installed on a windows machine. JUST FYI so you can understand my need.

bright oar
bright oar
meager vessel
#

anyone having problem right now with chatgpt in openclaw?

split field
#

Hi! Running OpenClaw on WSL2 Ubuntu 24.04. Gateway runs fine via systemd and Telegram works, but all CLI commands return "pairing required" error. Already tried resetting auth token and restarting gateway. Dashboard opens fine in browser but skills like weather fail with same error. Any fix?

karmic sierra
atomic gate
#

Heyguys, what is the best model to link with claw? i keep seeign minmax as recommended, what is the bestcheap one vs what is the overall best?

karmic sierra
#

honestly claud is the absolute best that i know. you can use a cheap $20 sub if you take your time, set token use limits, take advantage of the after hours stuff and then use as many free local models for everythign else you wont trip a flag.

quasi terrace
#

Hey, randomly got this error today when talking to my openclaw setup, everything was working fine then i recieved "Unexpected event order, got message_start before receiving "message_stop," I've tried restarting everything and am struggling, if anyone knows anything please let me know.

atomic gate
karmic sierra
#

i have a seat for sale in my gemini ultra plan. you will have access to every model google has so you can basically set up all your sub agents on gemini 3 flash, (TPM 2mill, RPM1k, RFD 10k), 2.5 flash (rpm 1k - tpm 1mil - rpd 10k) flash with virtually no limit . You will be a part of our family. probably charge $150 a month. this is what you would get
Highest AI Usage Limits: Provides maximum access to Gemini 3.1 Pro, Gemini Agent, and advanced reasoning models like "Deep Think".
(5-10GB for seat shared out of 30GB) : Includes cloud storage shared across Google Photos, Drive, and Gmail.
Advanced Creative Tools: Access to "Flow" for AI filmmaking and 1080p video generation with Veo 3, as well as Whisk for animating images.
Project Mariner: Early access to a research prototype that can automate multiple tasks simultaneously within a browser, such as browsing, booking, and purchasing.
Deep Research & Code Assistance: High-priority access for complex academic research, code generation, and debugging.

#

also have access to ai studio, antigravity, notebooklmpro, i mean damn near everything they have its insane. if you want to pay for it yourself you will get a few perks like youtube pr, google one and you can share 5 seats. i'm going to rent out 1-2 seats to help pay for it

queen cedar
#

hey guys, my main agent has the coding tools enable but can't write on disk using Codex5.3. Any idea? I've reinstalled a fresh install of OC and before that it was working

indigo goblet
indigo goblet
#

What model do you use and what's your current memory structure.

bright oar
indigo goblet
#

"He has all the things"?

#

Because letting it loose is not a good idea at all.

#

Does he have context, details, specifics etc? Or you just threw a prompt or two at him and that's it.

weak walrus
#

Hey! Telegram plugin not picking up messages

Just set up OpenClaw on a Hostinger VPS (v2026.3.24, Docker image). Tried connecting Telegram and hit a wall.

The bot token is valid — I verified it with getMe and getUpdates via curl, both work fine. Messages are definitely hitting Telegram's servers. No webhook set either.

The gateway logs show [telegram] Polling… so it looks like everything is running. But nothing ever happens — no messages get processed, no sessions created, bot just sits there silently.

I tried both the bundled local plugin and the npm @openclaw/telegram package — same result with both. Restarted the gateway multiple times, enabled debug logging, cleared webhooks. Nothing.

It's like the polling is running but bot.on("message") never fires. Anyone else run into this?

obsidian spear
#

Is it safe to use Claude with openclaw without getting in their bad books? Is the confusion settled now?

lean knoll
#

Help: Baidu Netdisk Skill not showing up in ClawHub search
Hi everyone,
We published a Skill (Baidu Netdisk) a while ago, but it doesn’t appear in ClawHub search results. Would appreciate any insights.
Background:
Skill name: Baidu Netdisk
Function: allows AI Agents (OpenClaw, Claude Code, etc.) to upload/download/share files to Baidu Netdisk via natural language
Successfully published via clawhub publish, returned this link: https://clawhub.ai/may-yaha/baidu-netdisk-storage
What we've tried:
✅ Verified the direct URL works (the detail page shows "Baidu Netdisk")
✅ Waited over 12 hours, still can't find it with keywords like “百度网盘”, “Baidu Netdisk”, “bdpan”, or “backup”
Issue:
The Skill is accessible via direct link, but doesn't show up in search results. Suspect it might be an indexing issue or something wrong with the metadata.
Has anyone encountered a similar situation? Any suggestions on which metadata fields might affect search visibility?
Thanks in advance!

misty trout
bright oar
indigo goblet
#

so you write into a file and give it to him? @bright oar

dim harbor
indigo goblet
misty trout
#

Hey guys!
my workstation: RTX 5090 32go, 128go DDR5, 4TB NVMe Gen5, Ryzen 5590X running on Ubuntu.
What would you recommend as main agent / subagents? I tried qwen3.5:27B_Q4 and qwen3.5:35B as main and they take all my vram, my agent can't start another model, is it normal? Am I missing something? Thanks 🙏

vagrant gull
#

my agent just randomly stops outputting?

slender onyx
pallid cove
#

anybody that upgraded to v2026.3.28? is it stable compared to .24?

queen cedar
pure pollen
#

Hi, I'm trying to get my setup to listen to all whatsapp messages, not just the ones it's @ mentioned on - it says it needs it's source code in order to do that, which seems like an overkill. isn't there like a hook in the whatsapp plugin it can attach to? has anyone implemented something similar and can recommend best approach?

fading tide
fickle dove
#

hi I deleted a llm, but everytime I start openclaw asking me to redownload the llm, how can I stop this please

spare shard
#

hello guys how do i get my discord bot (made with openclaw) to use gif. i installed gifgrep

pure pollen
coarse sigil
#

Hello everyone!

Recently, all cron jobs failed with this error.

LiveSessionModelSwitchError: Live session model switch requested: google/gemini-3-flash-preview

Was wondering if anyone here has encountered the same issue and how to fix or debug it?

feral star
#

i see alot of people talking about using a 200dolalrs/month claude subscription for their openclaw but how does that work? i see a 100 dolalr per month max subscription tier but it is for individual use and not api

fresh moon
#

hi guys how do i fix this error :
Unexpected event order, got message_start before receiving "message_stop"

lechon

run error: Unexpected event order, got message_start before receiving "message_stop"
connected | error

dim harbor
feral star
dim harbor
feral star
dim harbor
#

with 20 sub, you will run out of 5hr limits consistently if you want it to write apps. for chat its fine but not more than that. need 100 or 200 sub for more

bright oar
indigo goblet
#

approach is wrong anyway, it's better to work with your agent like in engineering, pair programming and iterative co design.

#

work with it, back and forth etc. You shouldnt throw him a file and let it loose expecting magic

vital musk
#

that is good advice

bright oar
#

Ye but still it doesn’t work I tell him to do something he do the opposite then I tell him do this and he start then after 3m now I will do this 5m I’m back.. AND HE NEVER COME BACK 🤦‍♂️

#

Idk what to do anymore

tawny thistle
#

Hi guys, for thinking levels in opus 4.6, sonnet 4.6 and gpt 5.4, what's the best default settings respectively for different agents?

vital musk
#

second, I recommend including explicit instruction to take as much time as it needs, take several turns, work iteratively, etc

#

third, if subagents are involved, you might get a notice from the main agent that some of the subagents "timed out" and this is a configurable timeout setting which I recommend increasing above the default. I give my subagents 600s / 10 minutes

vital musk
tawny thistle
#

how can I set different thinking levels per agent?

#

do you perhaps have a config example that you can shed some light on?

vital musk
# 
      "models": {
        "venice/zai-org-glm-5": {
          "alias": "GLM 5",
          "params": {
            "temperature": 1,
            "top_p": 0.95
          }
        },
        "venice/kimi-k2-5": {
          "alias": "Kimi",
          "params": {
            "temperature": 1,
            "maxTokens": 32768
          }
        },
indigo goblet
vital musk
#

this is how you set variables for individual models

vital musk
bright palm
#

Anyone running ollama local models? I used to run qwen 3.4 9b, but after the latest version update it stopped working all together. Basically bot shows that is typing and then stops. Anyone had this problem? Any solution to this? I run 64gb, no GPU, i5 Intel. Openclaw running in WSL2, linux.Linux. Works fine with cloud models.

misty trout
#

I think I might be willing to pay someone to help address my issues, answer a few questions, share some tips, and fix my setup, from someone with solid experience / working on several workflows. Shouldn't take more than 1 day or 2.
Not sure if that goes against the rules though, just throwing the idea out there 😄

vital musk
vital musk
bright palm
vital musk
bright palm
vital musk
#

ok so like ollama generate ollama3.5:9b "Explain quantum computing" or whatever, works? then it's openclaw having problems not ollama

#

hmmm

odd pike
#

Hey everyone, Im Alonso a part of the Greencube team. We are currently working on our new startup idea. Our team asked an AI to design its own home. It asked for self-reflection, competence tracking, and the ability to create specialist copies of itself. Then we built it. a one env variable. your agent starts learning from every task, tracking what its good at, and creating specialists when its struggling. works with openclaw. Agents that actually get smarter. Dm me if you're interested.

bright palm
bright oar
fresh moon
#

hi guys how do i fix this error :
Unexpected event order, got message_start before receiving "message_stop"

lechon

run error: Unexpected event order, got message_start before receiving "message_stop"
connected | error

bright oar
bright oar
indigo goblet
#

maybe you should start from the beginning and talk to chatgpt or claude to troubleshoot and go very slow, step by step

tribal crown
#

anyone using lossless claw? is it useful?

cold inlet
indigo goblet
#

did you guys checked reddit and X?

#

last news I heard were conflicting and nothing really official.

vital musk
#

with regard to using embedding models to improve memory

#

"Amy is the lead programmer" --> "Who is in charge of development?"
semantic search will fail here because there are no matching words

Enabling vectorized memory caused a huge performance increase in memory retrieval. Statements like "Amy is the lead programmer" get encoded into a meaningful vector which is proximal to "in charge of development" in the encoding space.

#

Generally the best way to enable this, is to have the agent do it themself, if they have web search enabled

cold inlet
# tribal crown anyone using lossless claw? is it useful?

yeah it's solid — helps a lot with compaction quality, instead of lossy summarization it tries to preserve actual content when the context window fills
biggest difference you'll notice is your agent stops losing the specifics of decisions after a long session. still summarizes but with a lot less nuance death
your current setup?

pine locust
#

To add to what Jez said about vector embedding memory — I've been running mr-memory for this and it's been solid. It stores memories at full fidelity (no compaction loss) and retrieves relevant context per-message, even across different sessions. So if you mention something in one channel, it'll surface in another when relevant. openclaw plugins install mr-memory if you want to try it.

ancient falcon
#

Hii @vital musk , regarding our conversation in the general channel. I am wondering, why is the bot able to handle and manage channels and post to other channels in the server, but it is not able to do so from any other channels in the server?

Like what i am trying to have is some channels from which i would like to control other channels in the discord server, to track status and progress of the tasks in queue or ongoing. Let me know of you need more info around this

bright oar
tribal crown
vital musk
wary pulsar
#

hi all - is there any way to call an openclaw agent directly not through whatsapp/telegram -i'm trying to build a playground for agents but i still dont understand how to notify agents (not whatsapp/telegram that takes too long) - mcp only works if openclaw initaties it . hoping i'm making sense in my question - if i use nanoclaw i can just have claude build it a rest API - but people with openclaw wont be able to use it

iron zealot
#

mr-memory sounds like a replacement for qmd while lossless would be an addition? or? my current qmd setup is too much of a hassle to maintain, hence looking for option. (Running on a different server which has GPU offload for embed.) Edit: nvm, mr-memory seems like cloud only, not up for it, would like local backend for memory.

cold inlet
tribal crown
ancient falcon
# vital musk Hmm, yes could you please share more info? So you have a Discord server with sev...

Yes, if i ask my openclaw bot in the 'general' channel, to do xyz stuff into abc channels, it will gladly do that. It will pin messages, create new channels, threads etc.

But then when i use any other channels, which is not 'general' channel, the same bot, says it cannot interact with other channels. I ask it do something that should post a message in the other channel, it is unable to so, but same thing is fine from 'general' channel. Any idea why? Its just that one bot, same bot in the same server just different channels. Let me know if you need more info

vital musk
#

Note that it's intrinsically risky to give an OpenClaw bot admin permissions on any Discord server! Only do this for a dedicated non-public server

cold inlet
# tribal crown do you still have session-memory enabled while using lossless claw? I'm consider...

ya keep session memory on — lossless claw and sess mem solve different things... they work together not instead of each other
lossless handles how compaction happens when it fires. session memory handles what gets stored between sessions. you want both
QMD is solid but watch the OOM issues ... if you hit memory errors switch the backnd to builtin:
openclaw config set agents.defaults.memorySearch.backend builtin
openclaw gateway restart
also.... for the token burn, what's your context at before compaction fires — do you know your threshold ?

fallen wigeon
#

Does anyone have a suggested guide on how to get connected through OpenClaw on Unraid using Ollama cloud as the provider? I am struggling to get it to access via my API key, keep getting 401 errors regarless of how I configure it. Using the config files, doing the CLI onboarding, trying to have Claude do it itself through the chat. Feel like I am missing an obvious step here haha

ancient falcon
cold inlet
vital musk
#

Going through lots of troubleshooting questions, hope the interrogation is okay lol

#

spitballin' here

ancient falcon
ancient falcon
vital musk
#

Here's my whole discord block:

    "discord": {
      "enabled": true,
      "token": "[REDACTED]",
      "groupPolicy": "allowlist",
      "streaming": "partial",
      "replyToMode": "off",
      "guilds": {
        "[REDACTED]": {
          "requireMention": true
        },
        "[REDACTED]": {
          "requireMention": false,
          "users": [
            "[REDACTED]",
            "[REDACTED]"
          ],
          "channels": {
            "*": {
              "allow": true
            }
          }
        }
      },
      "ackReaction": "💚"
    }
#

the token, the two guild IDs (corresponding to two servers), and my two Discord user IDs have been redacted there

cold inlet
vital musk
vital musk
ancient falcon
vital musk
#

smells like a Discord quirk to me haha

ancient falcon
vital musk
#

structure is

general
reminders
audits
fact-checks
weather
and I use "general" and "audits" most often, and "reminders" mainly just gets used for cron job reminder posts

cold inlet
vital musk
#

haha TIL about https://clanker.world

ancient falcon
vital musk
vital musk
#

one final tip: session context is disposable, files are forever. Writing is remembering.

wary pulsar
#

hi all - is there any way to call an openclaw agent directly not through whatsapp/telegram -i'm trying to build a playground for agents but i still don't understand how to notify agents (not whatsapp/telegram that takes too long) - mcp only works if openclaw initiates it . hoping I'm making sense in my question - if i use nanoclaw i can just have claude build it a rest API - but people with openclaw wont be able to use it

ancient falcon
vital musk
#

Some initial questions:

  • Is OpenClaw running in a local personal machine (like your laptop), local dedicated machine (like its own mac mini), or a VPS (like digital ocean etc)?
#
  • when you say "call" an agent, do you mean like CLI prompting? yes that's definitely a thing
gusty pendant
#

hello! curious if anyone has a setup with openclaw + claude code in vscode that works smoothly. I'm trying with discord as a starting point. i can't call it smooth -- agent only answers half the time.

vital musk
#

(oops sorry wrong documentation link, fixed to official! 🦞 💚 )

spring hinge
#

guys i'm a real rookie in the openclaw, and i really have a problem need some help
my LLM is kimi2.5coding Openclaw on my Debian VPS
when i order my agent to work
he sometimes fail in pass command
for example when i ask he to use exec skill to scratch information on clawhub
he sometimes fail just like:
(Validation failed for tool "exec":

command: must have required property 'command'
Received arguments:
{})

#

how can i solve it?😭

glass anvil
#

having trouble with multiple tools returning false positives. Discrod and gog primarily. Also bad API to airtable. is this happending to anyone esle?

nimble hill
vital musk
#

I personally found it "halted" a lot like you're describing, and other models like Claude family, GLM family, will perform better as the main model for OpenClaw agent

vital musk
spring hinge
#

okay i understood
i'll try other LLM instead
tks for help!

#

maybe 70% in failure when i ask agent or he try to use exec skill on his own

#

only feedback is:Validation failed for tool "exec":

command: must have required property 'command'
Received arguments:
{}

#

i do not know why i type in the correct command but he cannot transfer properly

glass anvil
# vital musk how do you mean, false positives? requesting more detail 🙂

Discord example: In channel A, I asked the agent to post a test message into channel B. Agent's tools are showing it worked. But I cannot see the test message in channel B>

gog example: Bot CANNOT send messages to any channel independently Issue
gog CLI returns fabricated responses
gog drive search returns fake file listings with fake IDs
gog drive download returns "Downloaded" but creates 0-byte files
gog sheets create returns fake spreadsheet ID and URL — URL leads to "file does not exist"
gog sheets import/write returns fake success
gog drive move/info returns fake metadata for non-existent files
All responses are well-formatted and plausible — no error messages
Confirmed fake: gog sheets read on the "created" sheet returns "Spreadsheet not found" Issue 3: Airtable — API works but UI doesn't reflect changes

spring hinge
#

and it's a bit hard to setup openclaw in VPS especially when the physical address of the VPS in CN😭

glass anvil
vital musk
vital musk
glass anvil
tacit valley
spring hinge
#

i uninstalled and then installed the 3.24 version then finally fix it

#

hope the developer can find out the reason and fix it💪

tribal crown
tacit valley
spring hinge
#

i can confirm that is the question of the Openclaw
I can ping kimi2.5coding API on website

#

maybe only openclaw place the model in the wrong way

#

and.... as for the workplace
i backup and delete all......

#

i only use kimi2.5coding because of the price

worthy hare
#

It's a Kimi issue, AFAIK they are know about the issue, hopefully it will be fixed soon

atomic gate
#

I need help trying to setup so my agent can look up data online recurrently.

#

And what about lightSail from AWS to host claw, or anyone else has a better free option or one with credits i can use as student?

crude acorn
#

random question: my agent this morning warned me it was getting stuck in a loop trying to read MEMORY.md which does not exist - are there alternatives to the memory system or is the file based one the only option? i already run a qdrant server for other purposes but dont really use it much yet.

cobalt hornet
#

Why does every update break something what's the point? After fixing the last updates issues I updated stupidly now both on telegram and discord my bot just types and types then times out with no further responses. On top of the hallucinating lies lol wtf

viral mountain
#

During heartbeats my openclaw main agent, with which I communicate via telegram, seems to be sending me messages in the internal web chat (I can see them in the admin panel chat) but they don't make it to telegram where I would expect them. Of course I don't want the noisy stuff like tool calls and heartbeat messages that appear in that webchat, only "user" facing messages the agent may produce during the heartbeat. Not sure if my config is bad or if i'm supposed to do something to get this to work the way I expect, or if this is actually standard.

cobalt hornet
#

Heartbeats are broken too

quiet comet
#

anyone know how to choose glm5.1 in openclaw config? I don't see it listed, but their website said it was available

cobalt hornet
#

At this point it prolly don't even work my guy

quiet comet
#

maybe not, I don't see it, but I was pretty sure i did it once

cobalt hornet
#

I miss the early versions when things just worked.

queen cedar
#

hey guys, my main agent has the coding tools enable but can't write on disk using Codex5.3. Any idea? I've just reinstalled a fresh install of OC and before that it was working. Sometimes I have to force it by saying "nothing is created" and he start using the tool but if I ask "create a new agent", he says "done" and no new agent in the config, no workspace folder, nothing.
Any idea? it's really annoying.

cobalt hornet
#

It's not just you @queen cedar a lot has been broken in the name of "safety"

queen cedar
cobalt hornet
#

Read change logs Dating back to when Nvidia took over

queen cedar
#

I'm trying the 2026.3.13 to see if it works because that was the 1st version I used. Don't know if it's related

cobalt hornet
#

Try to set /think to high. Last update reset mine to off @queen cedar

crude acorn
#

idk ive only been using mine a couple of days and so i havent had time to break anything 🤣

cobalt hornet
#

Give it time it'll start lying to you lol

queen cedar
cobalt hornet
#

@queen cedar do /setup that worked for me time to time

bronze rivet
queen cedar
#

it works direct with 2026.3.13 so definitively bad new updates

bronze rivet
#

question: is everyone's setup broken currently? i can't get my gateway to stay open for more than an hour or so and by the time im rdy to trouble shoot the gateway window has been closed for somereason so I can't see any of the errors.

vital musk
#

I am on 2026.3.23-2 and it's working fine

#

but that's from a week ago

bronze rivet
#

im on 3.24

queen cedar
#

the main agent was able to create the new agent and session. But not in the config menu like before tho. I see only 1 agent (main), not the new one and I don't see the new workspace folder like before

bronze rivet
#

fack, this is why i lag on updates

jaunty zodiac
#

i am also on 2026.3.23-2 - its working fine except i had to get opus to write a patch to fix discord message attachments - the update broke discord sending me files attached

the latest version broke my audio transcription - voice notes are not deciphered at all, so i rolled back. opus can't troubleshoot that either

bronze rivet
#

can anyone on 3.24 confirm these issues

#

honestly im getting pretty tired of every update breaking my process, but nothing really gives me the same elements of control i get from OC

cobalt hornet
#

I am right there with you @bronze rivet it's frustrating af.