users-helping-users

[echo pine]

Brotha, if you didn't have Claude Code installed, and haven't done the claude setup-token steps, then you aren't on what we refer to as oauth with anthropic. For some reason, it still shows as token in that profiles result you looked at. Have you run the commands to get a setup-token?

[next walrus]

no, i did do that stuff, that means my onboarding may be wrong then?

[echo pine]

Yes, you're probably hitting some free tier anthropic API. You need to follow the auth steps for anthropic oauth, one of which is claude setup-token.

[next walrus]

so just re onboard?

[echo pine]

You literally install claude code, point it at your openclaw folder, and say, "I need to change the way this is authenticating. It's calling the API. I need to do something the people on the internet like claude setup-token. What are the steps?"

All the answers, are in the LLM. THey're magic.

[next walrus]

i access claude code from powershell, right?

[echo pine]

Ask an LLM? Or google it? https://code.claude.com/docs/en/desktop-quickstart

[next walrus]

remember, the reason I'm here is because I'm limited on my use, no more talky talky

[echo pine]

I hear ya. But, literally, go to google, type the words, "How do I use Claude code", and the gemini llm will have 30 lines of advice for you. Or open a tab to ChatGPT. Or go to claude.ai

I'm not being fussy. I'm trying to set you up for how the world works now. LLM's are the future. Asking them should be the first step for basically anything we do now. It's not within OpenClaw, it's in a browser, from many frontier companies.

When should I put out fertilizer on my lawn... ask the LLM.

How do I change the oil in my truck... ask the LLM.

How does one use Claude Code... ask the LLM 🤙

[next walrus]

My LLM burned 130 MILLION TOKENS and I have nothing to show for it.

in under 4 days

there are problems they cant solve yet

[echo pine]

Have a great night 👌

[next walrus]

sorry, 170M

[next walrus]

users helping users, my man, not users telling people to ask the LLM. I already had it try to fix itself, that's why we're here. If you think it's already smarter than you, i guess you're right

[ancient jungle]

ngl, this thing is definitely smarter than me when it comes to a lot of these tasks

[next walrus]

its not, it just has significantly more resources available to it instantaneously than you do.

properly functioning humans don't hallucinate

[ancient jungle]

i'd beg to differ

[tall meteor]

I started silly small...ensuring an agent only completes my tasks on Kanban cards, then they go on coffee break when all To Do column cards are completed.

I also have an auditor agent running QA, checking work of other agents.

[crude wren]

https://people.com/tech-pro-uses-chatgpt-to-create-cancer-vaccine-for-his-dog-and-best-mate-11928192
cant solve simple things but can solve death 😛

[copper nova]

I learned what happened

it's only affecting users that SSH into their claw setups

that's why I can't get a proper token

which is probably what's blocking home boy here too

so it's not actually to any fault of his own 😂

cus it stumped my ass too

[echo pine]

Hmmm... is it not copying out correctly or something?

[copper nova]

when you do the claude auth thing, it opens a browser, when you SSH in, you can't access that browser

[ancient jungle]

what's the issue again? cuz i SSH into my claw

[copper nova]

so you use the URL it gives you

the URL creates some different token that's not the actual long lived token you need

[echo pine]

ahhhhh. Interesting.

[ancient jungle]

ahhhh

[echo pine]

JINX

[copper nova]

yea pain in teh dick to figure out lol

[ancient jungle]

what did you do to solve it, just so it lives in my head somewhere

[next walrus]

yea, i remember having htat issue, so then i asked claude for my token for that and it just gave it to me

claude code, specifically

[copper nova]

did you really? 😂

[copper nova]

i thought you didn't have claude code installed

[next walrus]

i have it installed, i havent used it

[echo pine]

I just need gpt5.4-mini on codex to hit in the next release. Get a lot of my little jobs moved over. 5.4 is a bit overkill

[next walrus]

it took me 3 or 4 prompts to get it to give me my token though, said it was secret and couldnt be shared

[echo pine]

I don't think it gave you the right api key.

[next walrus]

ill paste it here you can check

[copper nova]

no

don't

do it

STHAP

you crazy

[next walrus]

don't? or do it? the syntax

[next walrus]

thats possible

[echo pine]

@copper nova Is there any way to get the right code, right now? Or need time to fix something deeper?

[sacred nebula]

I got a question, if I reached the API rate limit...
Would it help to change the provider?

[copper nova]

I'm working with some devs to figure it out for future users

I wonder actually... can I past that code into claude and get the token that way?

[next walrus]

claude gave me a token when i asked, it may have been wrong though

[copper nova]

OH shit

it works

ok

[lime mauve]

hey can you help me install gog skill for openclaw ?

[copper nova]

no

what? how would I help you do that

you need to go to google cloud and sign up an oauth client or something

then give your agent that stuff to figure it out

[copper nova]

ok this is what you do

[next walrus]

So wait, just to clarify, I'm not an idiot, I'm actually a genius/amazing detective?

[copper nova]

stfu and just follow instructions

how do you access the machine your claw is on?

[next walrus]

Windows Remote Desktop

[copper nova]

um, what?

where is your openclaw installed

[next walrus]

on the server behind me

physically behind me

but its monitorless, so i just go through RDP

its my plex server

[next walrus]

something happen?

[worthy kelp]

Kinda just popped in the channel, are you trolling? Or are you legit this dense

[copper nova]

oh god lol

[echo pine]

I asked my LLM... it's just me :/

[copper nova]

wait let me think, RDP is showing you.... a GUI of some kind?

terminal?

[next walrus]

It's windows 10 via WRD, so theres terminals, chrome...everything a normal windows PC would have

[echo pine]

Then he should be able to have the browser pop up and work fine to get the oauth key, right?

[next walrus]

I can bring up either terminal or openclaw GUI

or powershell

[copper nova]

do you see windows?

[next walrus]

yea

[copper nova]

fuck me

[next walrus]

im married

[copper nova]

you can just auth and get the token then

run the openclaw onboarding

[echo pine]

Agreed

[copper nova]

and have a terminal open to run claude setup-token

copy paste into the onboarding

dont fucking copy paste here

[echo pine]

all dat 👆

[worthy kelp]

This dudes gonna do some devious shit with OpenClaw I’m already knowing 😂

[next walrus]

okay, so I think this is where the issue is

because I think anthropic changed how this works

it doesn't show me a token anywhere anymore

unless it's hidden somewhere I can find it

[worthy kelp]

I literally did the setup-token thing yesterday and it worked fine?

[next walrus]

because I'm looking at my authroize page "Claude Code would like to connect to your Claude chat account"

did you do it right now?

I hit authorize

[worthy kelp]

Hit yes, it should give you a setup token after

[next walrus]

and it says 'youre all set up'

no where does it give me a token

[echo pine]

After you ran claude setup-token in your terminial?

[next walrus]

yes

[worthy kelp]

And it gave you a link to paste to the browser right?

[next walrus]

did you do it actually yesterday or like maybe last week 'yesterday'?

[echo pine]

This guy...

[copper nova]

it shouldn't cus he has windows there

it should just open the browser and he can click authorize

[worthy kelp]

Literally yesterday

[next walrus]

which is exactly what happened

[copper nova]

then it should show him a token

[next walrus]

it does not

[copper nova]

on that terminal

not the fucking browser

[worthy kelp]

Huh.

[copper nova]

I'm gonna kill this dude

[worthy kelp]

Yoo 😂

[next walrus]

There is a token in the terminal

[worthy kelp]

bro

[echo pine]

Bahahahahahahaha

[copper nova]

go paste that into your onboarding

[worthy kelp]

Not here please

[copper nova]

please not here

[echo pine]

That's your key bro. Save it somewhere, like in your password manager, so you never have to do this with us again.

[next walrus]

"⚠️ API rate limit reached. Please try again later."

[indigo goblet]

lol whats going on, lazy to scroll up but that seems funny

[echo pine]

Have you removed the other key (somehow), and specified which model it's trying in the config?

[next walrus]

I just went through onboarding again so it shouldn't have any other keys but I can check

[echo pine]

run /status and see if the part of the key it shows, is the one that you JUST got from the openclaw machine a minute ago

nvmd. It doesn't show part of the key when I run it

[next walrus]

when i do models status it does and it shows only 1 model, token, and the correct key

profiles=1 (oauth=0, token=1, api_key=0) | anthropic:default=SECRETKEYHERENOTSHARINGITWITHYOU

[echo pine]

well I'm stumped then. You've got a haunted install.

[next walrus]

the oauth should = 1 though, not token?

[echo pine]

Well, it's still a token, but using the subscription.

[next walrus]

right

[copper nova]

finally got it to work, my god lol

now I just hope I don't get banned

[next walrus]

do we know what the token limit on subscriptions is?

[copper nova]

you don't,

[echo pine]

You already checked the usage page, you were at like 8%. Not even close

[copper nova]

just session usage and weekly

[next walrus]

well my API rate limit page begs to differ...lol

[copper nova]

welp, I helped all I could

[echo pine]

What I would do: Keep that fresh API key, delete that entire openclaw folder. Go back to the install command, start fresh. Onboard. Put in that key in. And it should work 🤷‍♂️

[copper nova]

only jesus can save you now

[echo pine]

Go download Manus Computer 😉

[next walrus]

whats that?

[copper nova]

you think the chinese can save him'

[echo pine]

A user friendly openclaw. Similar to Perplexity Computer.

[copper nova]

just install claude code on your plex server through the terminal

then ask opus to save you

[next walrus]

it is installed there

[copper nova]

ok spin it up in the terminal

[next walrus]

its open

[copper nova]

type /model and select opus

[echo pine]

Opus knows all

[next walrus]

model set to opus 4.6

[echo pine]

Thinking = Super duper massive extra high quality stuff high

[copper nova]

tell it your problems

your hopes and dreams

it will save you

[indigo goblet]

whats your issue ghostoutlaw

[next walrus]

API rate limits except im not using an API key and I'm on the $100/mo plan. And I've been rate limited twice in 4 days now

first time I was on $20/mo plan, so upped it to $100 and 2 days later, boom, limit

[indigo goblet]

so claude correct?

[next walrus]

correct

[indigo goblet]

you talk to your agent on a terminal?

[next walrus]

Discord

[indigo goblet]

the API rate limits popup is normal, it does show that even on a sub, so your quota is probably gone

what are you doing with him, any automation

[next walrus]

no, tried to get it to set up some sub agents for web searching and a few other things but it hallucinated the shit out of those so i benched them all

[copper nova]

I pray for you, god speed

[next walrus]

most of those agents were set up to run on lower models for more specific tasks to conserve token usage

[indigo goblet]

whats the context, its running on a physical server behind you and you communicate through discord?

[next walrus]

Correct, there's a win10 server behind me, also handles plex, but it's beefy so I just use that to run other stuff too, nothing NSFW, figured putting OC here would make it simple and use discord since it's the easiest and most user friendly platform to communicate with it that's fairly universal, so that's in a private discord server where I'm the only memebr and I just DM the bot

[indigo goblet]

so you confirm you are using the right OAuth with it, it worked but then you hit the limit.

sorry not OAuth, whatever it is called for anthropic

[echo pine]

yeah, if he did it like we walked him through, he should have a fresh, 'oauth' anthropic code to his subscription.

[next walrus]

so, let me be clear...I did do what you all told me.

However....Opus found an old key likely being used still, testing now

did not resolve th e issue

[indigo goblet]

lol understood, hmmmmm

[copper nova]

tell opus in claude code to test each token and see if it hits rate limits

cus you're obviously not rate limited if opus responds to you

if you were actually rate limited, you would of seen the same error on claude code

does this make sense?

and type /usage in claude code and screen shot what you see

wait you can't post screen shots

that's for the privledged

just tell us exactly what it is lol

[next walrus]

2% 11% 18%

current, weekly, sonnet week

[copper nova]

it should look like this ( I haven't been coding enough)

[next walrus]

yea, looks exactly like that

except mines bluer

[copper nova]

yea, you're like 10% usage

you far from rate limited for the month

so tell your opus agent to stop being a lazy douche and figure it out

[next walrus]

copy pasting

[copper nova]

test each token and set the valid one up

[next walrus]

oh, it said it cant do that

sec

I can't test API tokens for you — I don't have the ability to make external API calls to Anthropic's endpoints from here. But you can quickly test it yourself by running:

so, we have a fresh key, because the key I just got from terminal a little while ago is newe and doesn't match any of the others I've used.

Should we just nuke the install entirely and start it fresh?

[copper nova]

tell it to look into your openclaw project and see how they are set up in there currently, then to test the tokens through it

give your opus agent the new key and tell him to set it up

[next walrus]

Got a 400 invalid_request_error, not a 429 rate_limit_error. The token itself might be invalid or malformed — it's not even getting far enough to be rate limited. This could be an OAT (OAuth Access Token) prefix (sk-ant-oat01-) that's expired or was never fully provisioned via OAuth flow.

The issue is that OpenClaw keeps generating these oat tokens on startup and writing them into the auth-profiles as "type": "token", but they're OAuth-derived tokens that don't work as standalone API keys.

[rose halo]

Is the best way to save tokens to auto compact on like 30k-tokens context, auto saving relevant info to memory before? (what i set currently)

I thought 30k tokens would be a 'cheap' amount. usually 50% context is like 100k tokens and I guess that would be sent on every message. any suggestions?

[copper nova]

switch to github copilot that does request billing instead of token based

[rose halo]

wym switch to copilot?

[copper nova]

as the provider for openclaw

i guess that also depends on what models you use

[rose halo]

currently using minimax free tier, and some others, I have some very generous limits.

I just want to know if my strategy makes sense.

[copper nova]

but I reccomend not compacting at 30k tokens, you're gonna run into compaction issues alot

[rose halo]

I noticed without optimization I spent 20M tokens on my first day. But its no biggie it was a free tier acc lol.

[rose halo]

like what?

[indigo goblet]

run "openclaw models status" in a terminal on your server, and past me the output here

[copper nova]

you have alot of context which is built into every request

[rose halo]

Do you auto compact?

[copper nova]

at like 70% I do a compact with lossless-claw plugin

compacting at 30k you're gonna be compacting every prompt, its gonna be a nightmare and your bot is gonna become very.... dumb

not gonna remember what he said the last messsage

cus it got compacted

[next walrus]

🦞 OpenClaw 2026.3.13 (61d171a) — I can grep it, git blame it, and gently roast it—pick your coping mechanism.

Config : ~.openclaw\openclaw.json
Agent dir : ~.openclaw\agents\main\agent
Default : anthropic/claude-sonnet-4-6
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (0) : -
Configured models (0): all

Auth overview
Auth store : ~.openclaw\agents\main\agent\auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (1): anthropic (1)

• anthropic effective=profiles:~.openclaw\agents\main\agent\auth-profiles.json | profiles=1 (oauth=0, token=1, api_key=

[rose halo]

cant find that one

[copper nova]

can't find what

lossless claw?

[rose halo]

but like, did i get this wrong, or will it actually send all the 70k tokens on every message i send?

yea i use Token-manager

i can configure a higher limit

[copper nova]

basically

openclaw is a token hog

every prompt in a long session and the content is sent with each message

it gets to a point where it's like 150k consistantly through each message if it's like a long coding session

[rose halo]

do you know if opencode, qwen code, other agents also resend all the history on each msg? or is it different

I guess the workspace files are a small portion of this, biggest token burner is the session history right?

[copper nova]

any model you use through openclaw will send the same

[rose halo]

idk, I'm using it normally seems normal not very stupid

[rose halo]

no, I'm talking other cli tools

[copper nova]

you can use opencode in the CLI/terminal and it will not have all that heavyness

[rose halo]

qwen cli, opencode-cli

do they also send the whole hstory on every msg?

[copper nova]

it does it's own compaction but yes, it sends the entire history of the session still

[rose halo]

kk thx bro

[copper nova]

you just hope the model is smarter with teh copaction

opus for instance, write himself a nice "what we did, what we doing" checklist before it erases everything

[indigo goblet]

stop openclaw > openclaw gateway stop
rename your auth-profiles.json to auth-profiles.backup.json then restart the gateway and run again "openclaw models status" and past the output

it should says the profile is missing

[rose halo]

Another topic, regarding scaling, has anyone monitored resource usage for a bit?
Like with 10 agents responding actively, how much CPU/RAM openclaw would consume (not running any local AI, only remote AI providers)

Just wondering cuz I will test this later

[dusty slate]

WHAT DO YALL THINK OF NEMOCLAW? is it worth to reset and use????

[rose halo]

Trying to access the web UI?

[next walrus]

🦞 OpenClaw 2026.3.13 (61d171a) — I've survived more breaking changes than your last three relationships.

Config : ~.openclaw\openclaw.json
Agent dir : ~.openclaw\agents\main\agent
Default : anthropic/claude-sonnet-4-6
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (0) : -
Configured models (0): all

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

Missing auth

• anthropic Run claude setup-token, then openclaw models auth setup-token or openclaw configure.

[rose halo]

ah rate limited by claude

tell your agent to stop the DDoS attack

[next walrus]

yea, I can use claude in the claude app, or on web, and claude code is working fine

[rose halo]

do ya have ur antrhopic apikey?

[next walrus]

I'm not using an API key

[rose halo]

are you using 'openclaw configure' to set it?

[next walrus]

yea, i onboarded via openclaw and used the claude token

[rose halo]

for antrhopic there is 2 auth options, apikey and another one which you can use

could try the other one if u haven't

[next walrus]

im using token, not API key

[indigo goblet]

run "claude setup-token" in the claude code CLI

[next walrus]

did that, got a token ready

[rose halo]

yo

ollama run qwen3.5:0.8b runs very decently on my basic vps CPU lol

[indigo goblet]

ok then run "openclaw models auth paste-token --provider anthropic"

[rose halo]

awesum

[indigo goblet]

and paste your token

then run again openclaw models status

[next walrus]

done

gateway start and test?

[indigo goblet]

whats the output of openclaw models status

[next walrus]

OAuth/token status

• anthropic
  • anthropic:manual static

[rose halo]

what issue you having?

[next walrus]

API Rate Limit

[indigo goblet]

no need to restart, run openclaw tui and message your agent, is it working?

[rose halo]

is the rate limit from your openclaw configs or from claude? you 100% sure its from claude right

haha

cuz you can configure rate limits in oclaw

[next walrus]

message in console or in discord?

[rose halo]

whenever you change config oclaw always auto restarts (at least here)

[next walrus]

im unfamiliar with TUI

[indigo goblet]

in a terminal, open a fresh terminal and run "openclaw tui"

[next walrus]

penclaw tui - ws://127.0.0.1:18789 - agent main - session main
gateway disconnected: closed | idle agent main | session main | unknown | tokens ?

[indigo goblet]

hmm

[next walrus]

gateway start?

[indigo goblet]

restart the gateway

[rose halo]

paste ur openclaw status

also what is your gateway network mode configured to? local loopback?

did u try openclaw doctor? it didnt help me much but it does give problem insights

[next walrus]

gateway started

[copper nova]

?

[indigo goblet]

you restarted it with "openclaw gateway restart"? open a tui again and see if the gateway is connected, send just test to your agent

[rose halo]

[indigo goblet]

then tell me if this is still the case

[next walrus]

big read out now

[copper nova]

what you want lol

[rose halo]

did u read what I said

😔

[next walrus]

response too long, what specifically do you want from the ping?

[copper nova]

about ollama running on your vps?

I'm confused lol

[indigo goblet]

you opened a new tui? the gateway is connected or is it still disconnected?

[next walrus]

gateway reconnected

idle

[indigo goblet]

send "test"

working?

[next walrus]

so...I assume I'm talking to my agent there in the cmd prompt

[indigo goblet]

its working?

[next walrus]

yea

in the cmd prompt

i get rate limited in discord

[indigo goblet]

ahhh coolll

[next walrus]

so is that an openclaw or discord issue?

[indigo goblet]

its related I would say

how did you connect your agent to discord initially?

[next walrus]

yea, so I set a discord bot per claudes instructions and then sent sent that to a private server and I just DM the bot there

and connected the bot via openclaw via the openclaw instructions

[dim harbor]

you can check here (instructions in the post) how to find if its discord or llm rate limiting you. enable logging and check exact message: https://discord.com/channels/1456350064065904867/1472405061144678430

[next walrus]

so, with that, the TUI suggested the /reset command in the discord DM, and my agent is now working in discord.

[indigo goblet]

fucking great! problem fixed

done and done

[next walrus]

indeed

now, I had actually requested that the agent do that automatically, daily or as needed, and told it even to build a sub agent to handle that and then update memory accordingly

that agent freaked the fuck out and caused all kinds of gateway crashes

[indigo goblet]

hm

[next walrus]

and so did every subagent it had made

[indigo goblet]

is your workspace clean and vanilla? any skills, anything important source of truth files modification? anything that you would call the workspace not 'fresh'?

[next walrus]

Fairly, yea, I only got into this on wednesday

[indigo goblet]

did you initially explain to your agent the context and your setup? windows server etc

[next walrus]

I did yea, it actually hallucinated quite a few solutions to a problem last night trying to get a dashboard working

I couldn't connect via the IP my main to my server but I could remote desktop in without a problem. It ended up installing tailscale and changing some IP configurations after like 4 hours to be able to pull a dashboard in a browser

@indigo goblet @echo pine @copper nova Thank you all again, Glitch is back up and running in discord

[copper nova]

it was all @indigo goblet , thank him lol

[next walrus]

Well, I think we may have looked at the problem the wrong way as well. I think the API limit I was hitting wasn't the monthly, but the per message limit, if ther eis one. I have a feeling it may have been sending my entire discord DM

with every message

[copper nova]

I had one issue a couple days ago when it was auto selecting the 1million context model in claude code when it wasn't actually available yet and kept rate limiting me that way

[indigo goblet]

to be fair, you guys did a good back and forth and tried to help him so it's kinda everyone

[copper nova]

but I think that's fixed now too

[next walrus]

also, who else is it 4AM for?

[noble lodge]

Hi

How can I develop skills for openclaw?

[indigo goblet]

hmm the thing with agent is imo to approach them as in engineering, pair programming and system co design

so to go slow, to be as specific as possible with minimum ambiguity

and to work with it step by step, like you have a goal, how do we approach this goal and what needs to be done?

what the agent can and cant actually do and does he knows enough?

once you have that, you zoom in on the first step and go slow, multiple passes, lean first iteration but working and see if you want to move on to the next piece

[indigo goblet]

what is it that you wanted to do exactly?

[next walrus]

So I had a dashboard that is being hosted on my server, that I should just be able to type the IP and port into my browser on any other PC on the network and pull it up, except it wasnt

the bot hallucinated for like 4 hours before it finally had me install tailscale, but that didnt even work

[copper nova]

there's a create skill skill you can use

[indigo goblet]

your agent is on the server right?

[next walrus]

Yea

[indigo goblet]

in that case I would explain him the context clearly and the environment he is in, then I would ask him what acces does he have, what does he see

[noble lodge]

I know, but I wanna make my own skill

[indigo goblet]

but honestly, for any consistent work and to go anywhere with openclaw, I would say focus on creating a memory first, itts very important

[copper nova]

I don't undestand lol use the create skill skill to make your own skill....

[noble lodge]

ah, I understand. I made my own game and am going to make own skill for this game
is it possible for using this skill?

[copper nova]

yes I believe so, it just gives you a frame work to create your own custom skill to do w.e. it is you desire

[indigo goblet]

because thats not normal at all, so, the thing is openclaw is not an out of the box battlefield tested ready so you have to fiddle with core first, then you can gradually move on to do real work

like the agents.md, soul.md etc and the memory are very important first

but for your dashboard issue, he should be able to figure it out, its just a one shot thing, but you cant just let it loose, work with him step by step.

[dim harbor]

got any good suggestions for improving the default md files ?

[indigo goblet]

it depends of how you want your agents to operate and the work you want to do with it

one suggestions I have is, avoid vibe at all costs, when it is about his files or future files he is going to use, it has to be machine readable and low ambiguity, otherwise the model will simply fill the gaps itself.

[dim harbor]

do you find that the model can edit them on their own or would you do a lot here manually in terms of suggestions? my agent "likes" his files so its tough to improve them. lol. running opus. I did do a revamp of soul file and am adding operational info to agents and other files, but i wonder when it becomes too much info and how lean we should keep them

[indigo goblet]

I wouldn't let my agent edit his source of truth files on his own, he can't do it since he is not allowed, if he wants to he has to tell me why first and then I approve but that never happened.

I did write his SOUL.md with him though, so we went slow and step by step, traits by traits with multiple passes to reach maximum machine readability and minimum ambiguity while making sure he understood what I wanted

took me a while but its working as I want so far

[dim harbor]

yeah thats what I meant, when you are improving the files, are you asing agent for improvements based on workflow etc, to propose what it can improve ?

[indigo goblet]

for "when it becomes too much", well it depends, if your agent can actually read the file without issue and things are clear and machine readable/low ambiguity and there is no conflict or overlap in this file or other files, it should be ok

what I do is I ask him directly e.g "is your SOUL.md too packed right now, can you read it without any issues?"

thats one thing also to be careful, conflict and overlap.

[indigo goblet]

I do yes, but vision usually comes from you but yeah

[dim harbor]

yeah for sure, clearly defined md file purpose helps with avoiding overlap and conflicts. you got any good resources on this or just figuring it out as you go ?

[indigo goblet]

not really any resources, initially I started with who is Jarvis from the Iron man movies really? how would that translate in the real world under openclaw and gpt 5.4? how would he operate?

and so I had a lot of back and forth with my agent about his SOUL.md

on what is possible and what is fiction, becauses LLM are not magic

and we basically agreed on certain things and started to draft one traits or 'operational rules' at a time and refining

[dim harbor]

nice , guess theres no shortcut to this

are you trying to get LLM to "believe" its what you have in your soul md file, or is it impersonating an assistant? I found that by default it was doing the impersonation rather than being that, eg 2 separate layers, llm and persona

[indigo goblet]

for example here is something pretty useful in his SOUL.md that I have under Decision discipline;

---

Verify live state before answering any question about current status, current behavior, or current availability.
When the answer depends on runtime state, configuration, model availability, tool access, system status, or what is true right now, check first instead of relying on memory, impression, or recent conversation.
Do not present an unverified current state as confirmed fact.
If verification is not possible, say that clearly, state what could not be checked, and keep the answer explicitly uncertain.
Prefer checked state over plausible explanation.

[indigo goblet]

yeah there is no shortcut lol, for your question, I never tried to make the model believe, initially I looked into if his SOUL.md and source of truth files actually works as in, is it really a source of truth for the agent or does he have a 'pre-baked' personality and other things from says openai, but it isn't, the model is pretty neutral when it comes to calling via API or codex for example, so its quite like clay and you can steer it into any direction you want.

I would assume it is the same for Claude, because its kinda the point of AI agents, to sculpt them like you want

here is another example under execution discipline just to give you an idea;

---

Carry each task to a checked state before calling it complete.
Do not leave work in an ambiguous state or imply completion when the result has not been checked.
When something is complete, state exactly what was done and what was verified.
When something is incomplete, state exactly what remains open, blocked, failed, or unverified.
Do not claim success before checking the result.
If the task cannot be completed, state the blocking reason clearly and state the next necessary step.

it's different for everyone in the end, it depends on what you want to do and how you want your agent to operate.

[dim harbor]

learned all that from mistakes and then just kept adding to the list ?

[indigo goblet]

a little bit yeah, I think I spent 2 or 3 days just going too fast and not setting up anything properly, just going straight to my goal which btw wasnt even clear in my mind, had no clear mental model

then ragequit

uninstalled it and started again, slow

the SOUL.md was kinda in one stretch, slow but one stretch and then refinining it over future days and weeks. I had a clear idea of what I wanted from my agent so that helped.

[dim harbor]

cool, what your line count for souls and agents files now ?

[indigo goblet]

it helps to know LLM's weaknesses

[indigo goblet]

soul has 147 and agents 248

[dim harbor]

sounds reasonable, im at 73 and 215 at the moment, was wondering if its getting too bloated

[indigo goblet]

yeah understandable, you could check that out with him

hows his memory? your agent

[radiant ocean]

hello lads
i got 2 sets of choices infront of me
$20 in api for claude or $20 in api for chatgpt
claude pro or chatgpt plus

what do i pick in each category and how long would these credits last me while using openclaw and doing programming

[dim harbor]

meh, im not sure . especially when i make changes, it does get confused. i was just fixing up task tracking, need to see how it will work now, i was just cleaning it up the last few days

[dim harbor]

subscriptions should be infinitely better than API

[radiant ocean]

yea ok but i will pay for api anyway as i will plug the api to openclaw, that's why i said two sets of choices

first one is $20 in api for claude or $20 in api for chatgpt

second one is claude pro or chatgpt plus

[indigo goblet]

hm I see

[dim harbor]

chatgpt plus i think is probably best value out of those options. they added 5.4 mini model recently so you can use that to reduce the usage on the subscription

https://developers.openai.com/codex/pricing?codex-usage-limits=gpt-5-3-codex&codex-credit-costs=gpt-5-4

you can switch it up next montht to see for yourself what will work for your specific usecase

[indigo goblet]

how do you do task tracking?

[dim harbor]

Initially I was using just a tasks.md file and then creating additional md files but wasnt very happy with it. because updated in one file were not reflected in the other , you know how that goes. i migrated to beads now and im testing it out for the last few days. its like a local jira/linear, but more lightweight. https://github.com/steveyegge/beads

ive been updating agents md file last days to try to get it to use it properly whenever we have tasks to do, starting/completing etc. and cleaning up references to the old files, so its not mixing them up. right now it seems good, its getting me tasks each morning that we have left to do. im using it for both personal and agent tasks right now.

[indigo goblet]

hmm I see

so his memories are handled by beads?

the entirety?

[dim harbor]

no, only tasks/to do lists specifically, thats been more important for me right now. memories are still using the default system with local embedding. its something im looking at next to see how to tackle, might experiment with some of the plugins that are available for openclaw specifically

thats been one of the issues, memory search is like " oh yeah you told me to do x 3 days ago, but im not tracking it anywhere". i added more specific rules to make sure each task is tracked and referenced , hopefully that helps

[dense halo]

Discord inbound images are landing in ~/.openclaw/media/inbound/ instead of my agent's workspace ~/.openclaw/workspace-wade/media/inbound/. How do I get inbound Discord media to stage into the correct agent workspace?

[fallow hedge]

Hi All,
Has anyone been able to automate Instagram completely with openclaw?

My ip is getting blacklisted time and again, before able to login

[indigo goblet]

oh yeah I see, default memory has a pretty powerful core actually but it needs to be worked on yeah

[indigo goblet]

yeah it can be done normally, without any plugins or skills but, hm, how to say it, the structure around it has to be built

[blissful trench]

anyone knows how to acutally intregate notebooklm into openclaw?

[little cove]

Hey anyone can help me with my bot? When I ask him to send a direct message trought WhatsApp to a contact, i include inform him the contact's whatsapp number he tells me that can't sende the message. Probably a problem with the message tool... I had search for the problem and i found that the whatsapp listener is not working good. This can be a WhatsApp ban for bots messages that is not using the whatsapp oficial API or could be a problem with openclaw's message tool?

[indigo goblet]

did you ask your agent?

[spiral parcel]

You guys using Claw on Raspberry Pi, are you getting this issue with cron jobs too? 🤔
openclaw cron list just dies with (gateway closed (1000 normal closure)

What’s interesting is that the log also shows that:
cron is actually starting
the job file gets loaded
the timer gets armed

So the cron subsystem itself seems to be alive.

That makes it look like cron exists in the gateway, but the CLI call to the cron RPC is losing the handshake.

What I’m trying to figure out is whether there’s a fix for this, or if it’s just an OpenClaw issue. 🤔

[echo pine]

We talking memory? I love memory chat 🤘

[indigo goblet]

hows your agent memory?

[spiral parcel]

When switching between models, I’ve noticed that memory gets worse. It feels like each model wants memory to be structured in its own way for it to work well. There doesn’t seem to be one structure that works properly for all models, at least from what I’ve noticed.

[indigo goblet]

hm might be because each models is somewho different, made, trained differently etc.

[spiral parcel]

yepp

Embedding is probably the future in this kind of stuff 🤔

[indigo goblet]

which is the issue with switching models often, there is inconsistency, in term of behavior and personality etc too

even work

[spiral parcel]

Just wish the good once can be cheaper one day.. that will be a game changer...;D

[rustic lynx]

Hey, How do I set delivery target on a systemEvent cron job via CLI? has anyone had an issue with this?

[spiral parcel]

Solution:
https://github.com/openclaw/openclaw/commit/36f394c299a91301a84c455be2bdb418eeb2d08e

[echo pine]

I'm getting it so dialed in. QMD was a big upgrade but now that LCM is in place, it's mind-blowing. Major upgrade and it's working fantastic.

Qmd + lcm = victory

[fallen mountain]

Hi, Can I use minimax 2.7 LLM?

[slate minnow]

-what are folks using to talk with openclaw from their mobile device? I notice discord etc does not allow exec and restricts some things. I mean there is good reason I guess. Are people using the regular web ui via tailscale etc?

[oblique perch]

hello everyone,
I am new to this scene. I am hooked! currently using ollama with openclaw, on windows without a sandbox. I have fallen for Abliterated models. Open to suggestions or reasons not to use these. Currently using "huihui_ai/qwen3.5-abliterated:latest" i have four RTX 2000 ADA's. I wanted to go to a bigger abliterated model. but it seems openclaw is making me use a sandbox for security. I dont want to do this. But cant seem to find a workaround.

[mighty rampart]

If you can’t find a way to configure it to do what you want, have your agent look at the source code and commit history, explain to you why it is the way it is. Then you can decide what to do. Maybe you modify the source and run your own custom version.

[oblique perch]

the working agent was trying . kept wanting to change the model name and disable sanboxing in its json file. but it didnt work

[granite bobcat]

Did it say why it didn't work?

I've found that sometimes my claw will say it can't do something, but doesn't specify that it won't do something, but persistence pays off. It'll eventually do what you want it to do (has been my experience).

[oblique perch]

it thinks it will work makes the changes . restarts gateway. then everything bricks.

[dim harbor]

use the validate config command to verify before restart

i just posted in the other channel about it: #architecture message. you can use CLI to update the config and add warning to agents md to make sure it knows it will brick itself with a bad edit 😄 mine got really careful after

[dreamy horizon]

hey guys why i can only use tools at main chat but not through discord chat

[proven fulcrum]

anyone down setup custom openclaw for me and to build its skill? ((paid job))

[granite bobcat]

ACPX configured?

[proven fulcrum]

nothing configured yet

I just want openclaws to make my work

gotta explain you if you let me dm u

[granite bobcat]

Anything in your terminal logs?

I'd say keep it here tbh, someone (more knowledgable than me) might see it and be able to help more readily than I can.

[proven fulcrum]

Ok ill try to explain

I'm into ecommerce so I daily research products and import them to test, doing this via some spytools and using some other tool to make this product pages - i just want everything to done via openclaw

I don't wanna deal with fiver guys just trying to met someone here

[granite bobcat]

Hm, this sounds beyond my scope. Someone else might be able to help. First step is get ACPX configured though (IMO), if you don't know how, you can actually just ask your Claw to configure it for you, and it will.

[proven fulcrum]

i want to run it on a vps

[chilly kayak]

Hey folks, looking for a little adivce here, cant seem to nail down a solid answer. I have a 32GB M4 mac mini and am trying to run a local model on it. I tried getting Qwen2.5-Coder-32B but it seems like its timing out, and its returing empty responses. Would Qwen2.5-coder:14b be small enough to run properly?

[proven fulcrum]

Hey, here's a quick summary of what I need:

• Scrape Kalodata via web UI (browser login, no API) — around 100–150 products/day
• Filter by video count + sales volume (exact thresholds TBD)
• Extract the TikTok Shop product URL for each result
• Feed that URL into Atlas (Shopify app, also web UI based) to auto-import the product into my Shopify store
• Everything runs on a VPS — I'll need help setting that up too
• Budget: $500–$1000

No experience with proxies or anti-bot stuff — happy to follow your recommendation on that.

Let me know if this fits your stack and what you'd suggest!

[indigo goblet]

hm interesting

[echo pine]

The biggest thing for me, is the agent ability to pull convos from other discord channels when needed. If i start talking about a feature or OpenClaw expansion that I want to do in one channel, and then it starts bleeding into a discussion that I want to wrap up in another channel, that is more about tha topic, I can just go to the other channel and tell them to pull the last bit of conversation to get caught up.

[indigo goblet]

first 4 features of lcm can already be done with openclaw memory though, minus the DAG, but cool idea nonetheless. unless you build it lol

[echo pine]

And, it's way, way smarter long term. They have unlimited, forever message storage... well, forever is a bit much. I don't know how big the DB will get long long term. But a few weeks of reasonably heavy use, and I'm at 40MB

[echo pine]

which 4?

[indigo goblet]

//Persists every message in a SQLite database, organized by conversation
Summarizes chunks of older messages into summaries using your configured LLM
Condenses summaries into higher-level nodes as they accumulate, forming a DAG (directed acyclic graph)
Assembles context each turn by combining summaries + recent raw messages//

[echo pine]

So the built in memory / context management, does those out of the box?

[indigo goblet]

it's not there 'there' but the structure for it can be done

[echo pine]

well yeah. Isn't everything something that can be theoretically done? This is one simple plugin, and you've got it already done.

Peter S recommended 😉

[indigo goblet]

default memory is basically a blue print for users, under the hood is powerful but need building

[echo pine]

True

[indigo goblet]

yeah technically lmao, true

[echo pine]

QMD for mem, LCM for smart context management. For me, it's night and day better. It just works, I don't have to tinker with anything. And long term, it'll manage itself.

[indigo goblet]

what I like with building it yourself is that you know the ins and out, you can expend it and fix stuff, and the mental model in your mind is clear, its not obscure.

[indigo goblet]

yeah understandable, for sure it is better

I mean, I dont think you can go anywhere with openclaw if you dont have a proper memory structure like that, makes work too annoying and inconcistent

[echo pine]

Ray does a much better job explaining than I could: https://www.youtube.com/watch?v=_aXfcVO6Pbs

[indigo goblet]

QMD seems interesting and useful tho

[winter snow]

after 2 months of being on Claude OAuth API key (max plan), i encountred " API rate limit reached. Please try again later." switched to Anthropic API, but didn't resolve it, went with OpenAI api gpt-4o-mini but feel like the bot is very robotic, less proactive. Did anyone have had a similar issues/experience? 🙏

[indigo goblet]

you dont have issue with it? apparently its not super stable in regard to time to search @echo pine

there is also qdrant but that is more for when you have many months/years of memory and its not exactly in the same category as QMD.

[dim harbor]

What kind of requirement to run qmd, it was telling me it would be too much for my machine

[echo pine]

I haven’t been running tests or anything. But for me, I have 500+ files in my obsidian, and qmd is happily indexing and searching them all. Seems like she’s running well

[echo pine]

I don’t know specifics man. Sorry

[dim harbor]

No worries, thx 😅

[indigo goblet]

cool

yeah 500 files is a lot, i'd say builtin/sqlite is enough if you dont have a lot of files etc or if youre not having issues anyhow

[echo pine]

There’s a reason Peter and all the heavy users recommend qmd. She’s just better.

Long term, you want a better search. Vector search is mucho better

[austere rose]

how to make open claw send images to me in telegram currently it can only send text images

[echo pine]

Have you tried having your agent add nano banana?

Then prompt it to send you something it generates?

[indigo goblet]

sure I don't doubt that but you need the use for it, otherwise its overkill and complicated for very little reason.

[worthy kelp]

What exactly is QMD? I just set up Mem0 for my agent, does qmd replace Mem0 or do they work together at the same time?

[indigo goblet]

I mean, for my use case, sqlite + gemini embedding is enough right now.

[winter snow]

anyone finds OpenAI chat responses more robotic and not as human as Claude ?

[worthy kelp]

Yeah, I do. Also GPT seems to have this bad habit of trying to turn EVERYTHING into a bulletpoint.

[winter snow]

right, haha, that's what I thought, the problem is Ive had "API rate limit reached. Please try again later." for hours that wouldn't go away, so had to switch to openai. not sure when the limits are reset

[granite bobcat]

I feel like this is useful when you use Codex as an intermediary between yourself and Claw (helps save on tokens/API rate limiting)

[worthy kelp]

If you’re using the Claude subscription there’s like a session limit (resets after several hours) and weekly limits.

[winter snow]

that did't move at all, did't register any limits, as if never even used it, could be some context throttelinng

[winter snow]

as in copy and paste from codex to claw for more precise insrtructions?

[granite bobcat]

No, I don't do any copy/pasting. I just let Codex do all the work of communicating and let Claw do the actual work that Codex passes to it from my request. If that makes sense.

[winter snow]

how do you link that worflow?

[rose halo]

anyone else see probably wrong context warnings in oclaw web UI chat panel?

Getting '100% context used 516k/200k' on a fresh session with 4 msgs. minimax oauth free tier apikey

also free tier is fresh just made account

so its not that i guess

[granite bobcat]

try /new in OC Control

[granite bobcat]

DM'd you

[rose halo]

I did /new in web UI chat, created a fresh session

maybe visual bug?

500k/220k context makes no sense either so probably visual bug

[granite bobcat]

I got the same thing a lot. Gemini?

[rose halo]

also i limited my messages to send only the last 20 messages not full history.

no im using minimax, free tier.

pretty generous free tier

[granite bobcat]

By chance did you change models at any point? I'm wondering because I switched from Gemini to GPT 5.1 and kept getting rate limited unil I removed all mentions of Gemini from anywhere in OpenClaw's .json* or any other OC files. It turned out that OC was still referencing Gemini's API even though I wasn't using it anymore.

[oblique perch]

still bricking openclaw....

[rose halo]

by default agent is not in sandbox

you created an agent with sandbox:true

so it runs in a container

[oblique perch]

i want to run without a sandbox. it works on a smaller abliterated model but having issues with larger ones

[blissful trench]

are there any good cli's that i can use with openclaw? i recently found about the playwright cli

[rose halo]

try running openclaw tui

will probably launch an agent in the host machine (not sandboxed) - assuming you installed openclaw (the gateway = the openclaw server) with the NATIVE install script, NOT with docker
then tell the TUI agent to delete your sandboxed agent and make a regular one

need more details hard to help without any info

like do you have multiple agents, or your only existing agent is sandboxed?

[slate minnow]

is there version of openclaw web ui that works better on mobile device, smaller screen etc?

[mild heron]

hello, is anybody using Nate Jones Open Brain project with Openclaw ? I'm wondering how to connect open-brain with openclaw

nate Jones open brain project here btw : https://github.com/NateBJones-Projects/OB1

[pale ice]

Running openclaw on mac mini m4 24GB and hitting non stop api limits with claude pro plan. Running haiku but now considering running a local model. Does anyone have experience with ollama with openclaw? Im trying to use it but its not replying at all. Im communicating via discord with my bot

[little pagoda]

yep just started for me too. i cant work around this. every message is saying - "⚠️ API rate limit reached. Please try again later."

[pale ice]

Have you tried any of the local models?

[little pagoda]

local models im sure work fine. but i need at least opus/sonnet to be the brains. im on a 200/month plan. just started giving me issues yesterday

[mild heron]

have you tried the onboarding command to connect it to local ollama ? I've had it once in my lowly VPS server (not a Mac) - it didn't work in my particular case due to hardware constraints... but the setup was fine if ran through the onboarding

[pale ice]

Yeah i hear you, im also trying to do quite a bit of coding so i was planning to jump on the max plan. However if youre facing the same problems then i wont

[mild heron]

but with 24Gb RAM... don't ko what sort of model you're planning to run, but I believe the first initial commands to the selected model required quite a bit of inbuilt context which requires a relatively big model... I would not work on my 3b or 7b models

[pale ice]

You mean have ollama do the onboarding?

[mild heron]

no.. I meant openclaw onboard command - then direct it to use your local ollama instance

[pale ice]

will try now

[rose halo]

run ollama on host machine

run ollama pull modelname

then add that ollama model to openclaw

[copper nova]

why 5.1, that old as heck

[slate minnow]

Well spent many hours getting NATS working for inter agent comms, only to realise that the main /front model will always interpret any results from sub agents. Should have thought about that.

So if you have local sub agents, stuff still gets sent to cloud agents via the main agent. Doghouse

[idle violet]

I’m trying to setup a discord channel in addition to my telegram channel.
I’m getting the Discord ID Fail message and think I’ve configured the bot correctly.

I saw this was an issue a month ago but not sure how it was resolved.

I updated yesterday.

[hexed imp]

Heres a question - has anyone tried to get openclaw to find ways to fund itself? For instance, starting out by getting it to generate $20/month for income otherwise it dies?

[quiet scroll]

Came here to ask the same question. Oauth no longer working for me. It's a shame because Opus works extremely well but the API costs are so high.

[idle violet]

yeah, heard of that before.
I’ve tried to have it make reoccurring income, but it’ll give you the basic ideas which are crowded.
I feel like humans still need to provide the idea/spark

[hexed imp]

agreed - im curious about using perplexity to research ways to generate income without human intervention...

[mild heron]

Adding tools like a CLI to Openclaw - what's the standard way ? I have the CLI binary but all AI suggested ways to add it to openclaw.jsn are failing miserably...

[hexed imp]

I would break out of Openclaw for this task and have Cursor (or whatever) do it.

[limpid cypress]

guys, help me with this error please: gateway closed (1006 abnormal closure (no close frame)): no close reason\nGateway target: ws://127.0.0.1:18789\nSource: cli --url\nConfig: C:\Users\Дом\.openclaw\openclaw.json",
"url": "ws://127.0.0.1:18789"

I can't get access to Control UI because of this

I only managed to load the page once, then it broke again

[mild heron]

check openclaw logs.... :~$ openclaw logs --follow

[limpid cypress]

Doctor warnings ──────────────────────────────────────────────────────────────────────────╮
│ │
│ - channels.telegram.groupPolicy is "allowlist" but groupAllowFrom (and allowFrom) is │
│ empty — all group messages will be silently dropped. Add sender IDs to │
│ channels.telegram.groupAllowFrom or channels.telegram.allowFrom, or set groupPolicy to │
│ "open". │
│ │
├────────────────────────────────────────────────────────────────────────────────────────────╯
Gateway not reachable. Is it running and accessible?
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: C:\Users\Дом.openclaw\openclaw.json
Bind: loopback

same thing

[mild heron]

then... maybe paste the segment that seems relevanbt to you to an AI like gemini or other to help out

[limpid cypress]

[sullen ivy]

how would i approve a message request uning the cli

[rose halo]

I'm trying to create another agent which is sandboxed and I'm going through literal hell. Also the docs on this are all over the place

[sudden patio]

Hi, I’m having several issues with my OpenClaw Discord setup and I’d appreciate some help.

Current problems:

None of the Discord commands are working properly for me.
The bot keeps telling me that I am not authorized to use commands.
It does not respond to normal messages in Discord and only reacts when I mention it with @.
The Discord integration feels inconsistent and unreliable overall.
I’m not able to use the bot normally inside my server even though the setup seems to be completed.

At the moment, I’m trying to get a basic working setup where the bot can respond correctly in Discord and commands function as expected.

If anyone knows the correct configuration or the exact steps to make the Discord side work properly, please let me know ot DM me THNX

[dark glacier]

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

Could you help me? Why OpenAI OAuth and OpenAI APIs work different?

[idle blade]

Hi there ✨

WhatsApp QR pairing broken — issue #33961

We're on OpenClaw 2026.3.13. WhatsApp QR login appears to succeed ("✅ Linked! Credentials saved") but the gateway never loads the session as active. Every send attempt fails with "No active WhatsApp Web listener."

We found issue #33961 and PR #37801 which describes our exact problem (getStatusCode doesn't unwrap the Baileys v7 lastDisconnect wrapper, 515 restart path is dead code).

Two questions:

1. Is #37801 merged into any release yet? If so, which version?
2. Is there a workaround for 2026.3.13 while we wait for the fix?

[idle blade]

Did you try installing brave api search? Or I’ve heard that Open AI models are really bad at executing

[tall meteor]

Mission Control & Kanban task board is working perfectly
My dev team picks up To Do cards and completes attached tasks.
Amazing 🎉

I create a CLI monitor to watch more granular progress:

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

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

Watching 25 existing session files. Waiting for activity...

[13:21:21] dev_manager    SESSION    → new session file detected
[13:21:23] dev_manager    SESSION    → started  cwd=/home/agent/.openclaw/dev_workspace
[13:21:23] dev_manager    READ       → SKILL.md
[13:21:29] dev_manager    READ       → runbook.md
[13:21:41] dev_manager    EXEC       → curl -s "https://agent.lan.arpa/api/kanban?status=todo"
[13:22:25] dev_manager    EXEC       → curl -s -X PATCH "https://agent.lan.arpa/api/kanban/cmn1zscfq000801pel87k59w9" -
[13:22:28] mcd            ACTIVITY   ← [user] Card Status Changed: Card "3-22 Card" status changed from todo to in_pr ()
[13:22:48] dev_manager    EXEC       → curl -s -X POST "http://127.0.0.1:7700/queue/card" -H "Content-Type: application

[toxic gyro]

~~Just isntalled in docker and cant get into the thing? ~~

I keep getting: control ui requires device identity (use HTTPS or localhost secure context)

but I dont have a reverse proxy or a local browser...

[solid narwhal]

Hi everyone. I have just started setup and cant seem to get it to even start up (lol). I have Ollama installed locally with a few models available, opened openclaw dashboard and it just wont respond. This is the most recent issue. Anyone free for a quick private chat please?

21:01:57 [diagnostic] lane task error: lane=main durationMs=602006 error="FailoverError: This operation was aborted"
21:01:57 [diagnostic] lane task error: lane=session:agent:main:main durationMs=602018 error="FailoverError: This operation was aborted"
21:01:57 [model-fallback/decision] model fallback decision: decision=candidate_failed requested=ollama/glm-4.7-flash candidate=ollama/glm-4.7-flash reason=unknown next=ollama/glm4.7-flash
21:01:57 [diagnostic] lane task error: lane=main durationMs=21 error="FailoverError: Unknown model: ollama/glm4.7-flash. Ollama requires authentication to be registered as a provider. Set OLLAMA_API_KEY="ollama-local" (any value works) or run "openclaw configure". See: https://docs.openclaw.ai/providers/ollama"
21:01:57 [diagnostic] lane task error: lane=session:agent:main:main durationMs=32 error="FailoverError: Unknown model: ollama/glm4.7-flash. Ollama requires authentication to be registered as a provider. Set OLLAMA_API_KEY="ollama-local" (any value works) or run "openclaw configure". See: https://docs.openclaw.ai/providers/ollama"
21:01:57 [model-fallback/decision] model fallback decision: decision=candidate_failed requested=ollama/glm-4.7-flash candidate=ollama/glm4.7-flash reason=model_not_found next=none
21:01:57 Embedded agent failed before reply: All models failed (2): ollama/glm-4.7-flash: This operation was aborted (unknown) | ollama/glm4.7-flash: Unknown model: ollama/glm4.7-flash. Ollama requires authentication to be registered as a provider. Set OLLAMA_API_KEY="ollama-local" (any value works) or run "openclaw configure". See: https://docs.openclaw.ai/providers/ollama (model_not_found)

I've spent over 28hours at this point and Grok/Clawaid are all firing blanks atm

[hasty grove]

I don’t follow this exactly, but you can execute single non-interactive message to agent thru openclaw gateway from the terminal

I do this to run an e2e agent ux test

[mild heron]

Set OLLAMA_API_KEY="ollama-local" (any value works) or run "openclaw configure"

[mild heron]

either edit openclaw.json manually to add that, or the openclaw configure command might assist

[solid narwhal]

Will try manual as have done config and set, neither worked.

[vernal gate]

Hey! this is actually a super common one. the error says ollama needs to be registered as a provider first.
just run this in your terminal:

openclaw configure

when it asks for ollama api key just type anything like ollama-local — literally any value works

that should fix it. lmk if it doesn't and i can help further ..

[rose halo]

So idk is there any complete guide to run sandboxed agents?

I can't run the agent in a sandbox in the first place, and the whatsapp channel just gets shared from main agent, to the sandboxed agent, idk why, since I have explicitly ran the command to connect the whatsapp channel for each agent separately. Literally ran the cmd channels whatsapp connect channelname agentname etc for both.

Also it's ignoring the docker image and sandbox enabled in openclaw.json

I want main agent on host normally, and a second agent on sandbox

I also adapted the config json to multi agent, replaced the default{} object with

Agents: list [] for multi agent configs

They appear in the Web ui, it just doesnt sandbox anything or run the docker container etc

And both agents list the 2 whatsapp channels

Which I think is wrong? They should only list their own whatsapp channel

Anyway the documentation is really not very clear on how to set this up and I couldn't fix this yet even with AI support lol

[long bobcat]

Hello OpenClaw Community,

I'm facing a critical, persistent ACP issue with sessions_spawn: "ACP target agent is not configured."

Context & Troubleshooting:

1. Agent Setup: "Trinity Architecture"; "Big Sister" (gemini-2.5-pro) is the intended ACP target.
2. acp.defaultAgent Config: Tried setting acp.defaultAgent = "big-sister" in gateway.json via:
   • Agent-specific profile.
   • Global acp block.
   • Channel-specific (per configuration-reference.md for channel:1485338354149195826).
   • All attempts failed; sessions_spawn ignores it.
3. agentId Constraint: My policy forbids agentId, making acp.defaultAgent my only ACP option.
4. CLI & Gateway: Currently OpenClaw 2026.3.13 (61d171a). CLI was broken, but I performed a full manual recovery and clean install. Both CLI and gateway are confirmed functional.
5. Final Test: Even with functional CLI/gateway, sessions_spawn to "Big Sister" still yields the "ACP target agent is not configured" error.

Conclusion: acp.defaultAgent appears ignored by sessions_spawn in v2026.3.13. This prevents inter-agent communication and seems a persistent bug, not a config error.

Goal: Enable primary agent to spawn sub-agents (ACP) using acp.defaultAgent.

Any insights or workarounds for acp.defaultAgent in v2026.3.13?

Thank you!</final>

[rose halo]

  "agents": {
  "list": [
  {
    "id": "main",
    "default": true,
    "name": "Main",
    "workspace": "/home/brewuser/.openclaw/workspace",
    "model": "google-gemini-cli/gemini-2.5-flash",
    "sandbox": {
      "mode": "off"
    }
  },
  {
    "id": "public-ytdlp",
    "name": "Public YTDLP",
    "workspace": "/home/brewuser/.openclaw/workspace-public-ytdlp",
    "model": "openrouter/free",
    "sandbox": {
      "mode": "all",
      "scope": "agent",
      "docker": {
        "image": "openclaw-ytdlp:v1",
        "readOnlyRoot": false,
        "network": "bridge",
        "user": "0:0"
      }
    }
  },
  {
    "id": "douglas",
    "name": "Douglas",
    "workspace": "/home/brewuser/.openclaw/workspace-douglas",
    "sandbox": {
      "mode": "off"
    }
  }
]
  },
  "tools": {
    "agentToAgent": {
      "enabled": true,
      "allow": [
        "douglas",
        "media-agent",
        "public-ytdlp"
      ]
    }
  },
  "messages": {
    "ackReactionScope": "group-mentions"
  },
  "commands": {
    "native": "auto",
    "nativeSkills": "auto",
    "restart": true,
    "ownerDisplay": "raw"
  },
  "session": {
    "dmScope": "per-channel-peer",
    "maintenance": {
      "mode": "enforce",
      "maxEntries": 20
    }
  },
  "channels": {
    "whatsapp": {
      "enabled": true,
      "dmPolicy": "pairing",
      "selfChatMode": false,
      "groupAllowFrom": [],
      "groupPolicy": "allowlist",
      "groups": {},
      "debounceMs": 0,
      "accounts": {
        "default": {
          "enabled": true,
          "dmPolicy": "pairing",
          "groupPolicy": "allowlist",
          "debounceMs": 0
        },
        "public-ytdlp": {
          "enabled": true,
          "dmPolicy": "pairing",
          "groupPolicy": "allowlist",
          "debounceMs": 0
        }
      },
      "mediaMaxMb": 50
    },
    "telegram": {

[jade cargo]

A little word here could help out. How often do I need to subscribe for Claude API? Install openclaw today, sub for API, did a few tasks and it’s already saying credit balance is too low. Even had me on -$2 on Claude console.

[vernal gate]

yeah claude api charges per token so it burns fast if you're not careful. easiest fix is set a monthly spend
limit in your anthropic console — goes to zero otherwise real quick

also worth trying gemini or local ollama models inside openclaw — completely free, no api costs at all. way
better for daily use

[rose halo]

I give up trying to configure oclaw with multi agents sandboxed/direct, the web UI does not allow me to configure visually and managing the config json is hell. I'll try the other 4 openclaw alternatives lol

[vernal gate]

yeah the multi-agent config json is brutal honestly. what are you trying to set up? might be able to point
you in the right direction before you switch

[rose halo]

It's pointless, I won't be able to use this for production agents at work or even for my personal hobby uses anyway, I need a 'clawer' app that I can actually configure without everything breaking. I already found 4 open source alternatives that seem better honestly

[rose halo]

I configured exactly as the docs say btw as seen above with the agents: list [(...agents here)]

sandbox doesnt work, whatsapp channels are being wrongly shared between 2 agents and etc I simply can't be wasting this time trying to figure out what exact keys they want me to put in the config json for this to work properly...

[mighty rampart]

My advice is to pay for a frontier model to manage the configuration of the OpenClaw instance. I use Claude Code at the terminal in the home directory of the user running the various containers that make up the agent ecosystem. (Ollama, Open WebUI, OpenClaw, mitmproxy.) Claude reads the docs, runs openclaw command line things, edits the json config if absolutely necessary (avoid if possible!) and importantly, maintains the runbook.md describing how this all works for my specific installation. When I want to upgrade, or if anything goes wrong, that is where I go to debug and get things working again.

[rose halo]

Do you run multi agents, with sandbox and without?

[mighty rampart]

I have multiple agents configured. They each have their own workspace with their own SOUL.md and they are bound to different Slack channels. My sandbox setup is outside of OpenClaw. Docker containers, mitmproxy to inject credentials so OpenClaw never has my API keys, and a restrictive firewall on the host and on the network. What is your use case? Research, productivity, software development, or ?

[rose halo]

how is your sandboxed agents setup organized? you have 2 gateways 1 in docker 1 in host?

I simply could not get the multi agent working, I was trying to keep my main agent on host access and a second agent sandboxed (in ccontainer)

I was almost giving up and just running multiple openclaws in docker which would be simpler but I guess it's better to switch to alternatives at that point

I could try again later with a smarter paid model, throwing the whole docs at it and trying to make the ai fix it for me
https://docs.openclaw.ai/tools/multi-agent-sandbox-tools#multi-agent-sandbox-and-tools

Feels like adding any new agent would be the same hell though, maintenance work would also probably be this bad

[mighty rampart]

In my case the entire thing is sandboxed. OpenClaw cannot touch the host, only its own workspace. I was trying to proxy everything but discovered that the webfetch tool does not respect the proxy settings, thus the additional firewalls to stop it doing GET bad-guy.com?stuff-that-should-not-be-revealed and exfiltrating via query string. I am not using OpenClaw’s built-in sandbox. I have not looked closely enough at it to understand how it could escape.

[rose halo]

Also the docs really feel like they were written/generated directed at AI models only, because the documentation is insane

[rose halo]

so youre running the gateway (whole openclaw server) on docker

[mighty rampart]

Which goes back to… let a really ‘smart’ LLM configure it for you. and then explain it to you. 🙂

Correct. OpenClaw itself is in a Docker container. It’s all managed with Ansible.

[rose halo]

I actually wanted my personal oclaw to be able to access my host cli cuz thats useful

so usually if you wanna manage server work running the whole thing in docker isnt what u want

But I could just run a second instance tho

[mighty rampart]

I allow Claude Code to touch the host. I do not trust OpenClaw that much! 🙂

[ancient jungle]

i spun up a separate VM and just let OC loose in it. i guess i'm a lot more risk tolerant than most of y'all. but i'm also not in any kind of business environment where security is actually a concern

[spark valley]

Have asked the agent main to set it up

[oblique perch]

@rose halo we seem to have the same sandboxed dilema

[rose halo]

[chilly kayak]

Im trying to setup local ollama, but having a hell of a time. Any advice?

[broken ginkgo]

Is there anyway to run OpenClaw and Ollama efficient in an old PC? not so old, it got 16GB RAM and 6GB VRAM of GTX1660

[spare berry]

Directly on Openclaw or with an llm proxy?

[rose halo]

its literally 1 command?

I decided I'll try to reinstall oclaw gateway in docker and try to make peace with the claw

[rose halo]

Just run the whole thing in docker problem solved I guess

[chilly kayak]

appreciate the help!

[rose halo]

you did not provide any information details or context or even error logs

[chilly kayak]

directly with openclaw

I was originally running kimi k2.5 cloud version but realized I was going to burn through api credits quickly. Started looking into what local models i could comfortably run on a m4 32gb macmini. started with 27b parameter models, quickly realized that was too much and was trying to get 14b models to work with smaller context windows like qwen3.5. Still unsure if that would be feasble, so Im resorting to llama 3.1 8b (then found out it is still 30gb of ram, and need to force it to smaller context window) havent been able to get either running, stuck in a loop

[broken yarrow]

just downloaded openclaw and api is limited. heard it this fixes itself over time?

run error: ⚠️ API rate limit reached. Please try again later.

chatgpt

5.4

100% ussage remaining

yeah

no

on it

says im not connected to gatewat

whats openclaw doctor?

thanks btw

gotcha

wont even let me type a command in command prop

just says gateway discconnected

got it

was in wrong termina;

used doctor and it "fixed"

nothing happened

discord default: Some configured guild channels are not numeric IDs |
| (unresolvedChannels=1). Permission audit can only check numeric channel IDs. (Use |
| numeric channel IDs as keys in channels.discord.guilds.*.channels (then rerun channels |
| status --probe).)

yeah but every time i try to talk to the egent it says rate limit reached

yeah

says it is running gpt 5.4

will try that, thankss

[chilly kayak]

got it working. I just started off with way too much juice

[mortal phoenix]

how do i run openclaw after i installed it but i want to use it again

[summer granite]

Hello, does anyone know where can I host openclaw for free aside from Oracle?

[vernal gate]

hey! what's the error you're seeing? most ollama issues come down to either ollama not running in the
background or the api key not being set

quick fix to try: run openclaw configure and when it asks for ollama api key just type ollama-local — any
value works

what does your error say?

[vernal gate]

railway.app has a free tier that works well for openclaw. render.com too. both pretty straightforward to set
up

what are you trying to run — just the gateway or agents too?

[summer granite]

agents too

i mean, im new with openclaw, would it possible to host both so I could use openclaw right away?

[graceful fox]

Hey what is the cost effective way to run openclaw ? im looking to get apiKeys from any of the cloud providers as i dont have a strong hardware setup to local host

[ancient jungle]

depends on what you mean by cost effective

[graceful fox]

yeah i agree. Just want to see what works for people here to get their openclaw setup running. i am just starting to scratch the surface here 🙂

[summer granite]

which hosting or cloud server can i use for host an openclaw?

[vast stream]

windows docker setup openclaw， can't modify the default model, how to modify the defaulted Claude model into Gemini?

[vale hinge]

Hey! I was just trying to setup openclaw for the first time on my windows PC. I'm using ollama for hosting the model locally. It was using glm-4.7-flash.

I have a 5070ti with 16gb of vram, but when it got to the step of "hatching," my whole computer was going super slow and I got a memory layout issue and I had to close everything.

I thought my computer could handle a 30b model. Is my computer the issue, or might there be something else going on?

[mighty rampart]

not an expert on this but even at Q4 I suspect that is going to exceed 16GB and need to offload to the CPU, making it very slow. first I would figure out what logs and metrics you have access to so you can understand what is happening on your hardware.

[hasty crane]

I’m hitting persistent “API rate limit reached” errors in OpenClaw even under extremely light usage and I don't understand what’s actually causing it.

Setup:

• Windows 11 Minisforum UM790 Pro - 64GB RAM, 1TB SSD
• OpenClaw gateway running locally (loopback)
• Models: Moonshot Kimi-k2.5 + OpenAI currently configured
  Previously also had OpenRouter, DeepSeek and several versions of Ollama configured.
• Discord + webchat channels active
• No heavy tools running yet (still in setup phase)

Problem:

• Rate limit errors occur after asking only single small questions sequentially
• No parallel agents
• No streaming jobs
• No long prompts
• No obvious burst usage
• Sometimes responses are fast, then suddenly rate limit storms start

What I’ve already tried:

• Reinstalled everything
• Restarted gateway over and over
• Cleared logs / cache folders
• Rebuilt memory structure (workspace/memory + MEMORY.md)
• Deleted scheduled gateway task and ran gateway manually
• Verified gateway listening locally
• Tested fallback model switching
• Confirmed not running multiple gateway instances
• Checked that no background loops are firing

What I think it’s likely NOT:

• Not user-side API quota exhaustion
• Not concurrency spikes
• Not large prompts / tool chains
• Not Moonshot-only issue (saw similar behavior earlier with OpenAI, DeepSeek, and Ollama defaults)
• Not network disconnects
• Not Discord spam loop

Questions:

👉 MOST IMPORTANT - has anyone else had this issue and solved it?
👉 Is OpenClaw known to generate hidden token usage / internal tool loops that can trigger rate limits even during light manual testing?

Would really appreciate any direction — thank you 🙏

[ancient jungle]

yea, but what plan are you on, are you paying per token?

[copper nova]

Are you using the free kimi api? Even the paid one I've been hearing its been rate limiting cus they are getting slammed

$10 unlimited sounds like it be flaky for sure

[hasty crane]

Paid versions of everything

[copper nova]

switch your default to openai and see how it goes

[hasty crane]

Already did.

[copper nova]

then bolt on another provider just for a sanity check like github-copilot

i love me github copilot cheapest frontier model usage for sure

[copper nova]

how are you hooked up with openai? oauth login with subscription?

[hasty crane]

API keys with everything

[copper nova]

yea I know kimi has been giving some phantom rate limiting as well as Gemini, I know the gemini one from personal experience.

i've had the best lucky with github-copilot so far

but even now they also rate limiting their opus and sonnet models

12cents too cheap for us to abuse apparently lol

[hasty crane]

Thanks for your responses. Unfortnately, I've run into the exact same issues with Kimi, DeepSeek, OpenAI, various Ollama models, OpenRouter. I've had them all separately as default and as the single models and I've also tried with default models and then several others as fallbacks. I'm still getting this API Rate Limits freezes no matter what I do. Makes me think it's nothing to do with the providers or models.

[copper nova]

hmm cache issue? do you have another coding agent harness you can use to debug?

spin up an opencode or claude code if you have it and tell an opus agent to figure out

[rose halo]

I don't think llama 3.1 8b should use 30gb RAM, are you sure it's not something else also consuming it?

As far as I know, the listed model size in GB has always been almost the exact amount of RAM the model consumes here / that size stays in RAM

I suggest you forcefully kill all ollama processes, and check running processes by RAM consumption just to be sure

[rose halo]

I could be wrong but I'm 'almost' sure that's right (that model you said is around 5GB size)

[rose halo]

oh

Also could be a 'leftover' ollama process generating text, that wasn't killed properly

That happened here and ofc I noticed the RAM and CPU on 100% and I had to pkill -f the ollama processes

[nimble hill]

thatswhy use docker to deploy application. This way you can have a clean playing ground and a security thru isolation

[queen cedar]

Hi guys. I notice something weird when I use some LLM. With Minimx or Codex5.3 my agents can perform action like creating files. But it doesn't work with Qwen3.5-9B (local in LM Studio) or GLM4.7 (from Nvidia API). Anyone has an idea? Both LLM have tools, or coding structure.

[nimble hill]

whats you token burn rate? with OC with API keys is just asking for financial disaster imho

[karmic notch]

What is the general consensus regarding a cost effective AI model to use as a primary agent? I’ve been running Sonnet through API and it’s pricey… I’ve been trying to delegate tasks to sub agents using lower cost models, but just the back and forth communication has been costing me a lot more than I’d like. Is there another AI model option I should consider? I really like the Sonnet “personality” and intelligence.

[copper nova]

try to use it through your claude sub, I'm a big fan of they models still, GPT has been too much of a snitch for my usecases

if you can't get your sub hooked up, use github-copilot as the provider, cheapest frontier API rates IMO. 12cent opus calls vs $4 dollar token based

[nimble hill]

I use 5.1-codex-mini - it is relatively cheapest.

[copper nova]

have you tried 5.4 mini yet?

[karmic notch]

Thanks! I’ve been using openrouter. Will copilot work through that service?

[copper nova]

it's gonna be 1cent a call on copilot, i hope it does decent as a chat layer

[copper nova]

it will work through openclaw, just select copilot in the onboarding process

and obv sign up a github copilot account first, they have 10 and 40 dollar plans

[karmic notch]

How does it compare to Sonnet for intelligence?

[copper nova]

it's just a provider for those models, you can use sonnet and opus through it for cheaper than openrouter

[karmic notch]

Oh!! Perfect

[nimble hill]

nah - had a API timeout - is that probably not supported by OAuth yet? or maybe I need to reset my gateway

[copper nova]

cheaper only if you're using it properly, please don't just say hello good morning to it

[copper nova]

oh I meant outside of openclaw, I have a fix in for next release to address the model mapping for github-copilot

[nimble hill]

use the OAuth for github copilot - quite good actually with $10 subscription

[nimble hill]

nah - I am not so much of a model testing enthusiast. I might use it in my n8n though - due to its low cost

[copper nova]

I'm assuming it's available through openai o auth?

i don't have a sub to try it tho

[copper nova]

I would be testing not cus I like to test models, it cus I want to save money cus I'm a cheap ass lol

[nimble hill]

of course. Since I run OC on k8s - I could just spin another instance and use that to call n8n and see the performance og 5.4-mini

This mcporter skill of openclaw is really messy - or rather it installs in every other place rather than in a skills folder in workspace. So it does not show up in the gateway dashboard, but it runs. Openclaw really needs to streamline its testing. With a "relatively" long delay between 3.13 and upcoming 3.14 - lets see if this improves. Especially this webhook stuff - which is completely against the usual norms

[copper nova]

submit an issue, or a PR if you willin lol

[nimble hill]

they have started limiting PR number to 10 🙂 not that I have raised though

[copper nova]

lol that was always the limit right? I never opened more than 2 at a time tho so not sure.

[karmic notch]

Hey… any of you guys have experience w persistent 409 errors in Telegram? I’m travelling and haven’t been able to connect to my OpenClaw agent after a memory flush. This error keeps coming up in my logs. Getting really annoying. Hope I don’t need to recreate it.

[karmic notch]

Can you elaborate on “using it properly”? I’m trying to run multiple subagents, using it to research and vibe code.

[copper nova]

gotta have to SSH into your claw box and see if you can figure it out with another agent harness probably

[karmic notch]

Yeah, I was using Termius to access my wife’s OpenClaw assistant (on Sonnet). Got to that point where it ran out of options. 🫣

[copper nova]

like give prompts and tasks that would normally cost you more than 4cent/12cents worth of API compute, if you're vibe coding with it than that's a good usecase. Don't just casually chat is what I'm saying 😛

[karmic notch]

Gotcha. No, it’s a tool not a friend 🙂. I’d like to feel more free to use it though. I’m into it for about $1000 of token cost in less than a month (of heavy use)

[copper nova]

yea... I think we all felt this pain initially, I like to call it tuition lol. You have a claude code plan? get a max plan and hook it up with your openclaw.

the $200 plan gives you roughly $2000 worth of inference according to their estimates.

[karmic notch]

I have a Claude Sonnet Max Plan subscription… heard Anthropic doesn’t want the subscription plan used w OpenClaw.

[copper nova]

it's heavy subsidized

[copper nova]

they don't but you can still

[karmic notch]

Fair point

I’ll revisit that. Lots of work to do and it would be a shame to get priced out

[copper nova]

just switched to github-copilot will save you a bunch

[karmic notch]

Heard copilot rate limits/ throttles the output or restricts context? I need all the juice I can squeeze and am concerned this will result in reduced quality of out put

I’m completely fine w paying the “tuition”, just want to make sure I learn the lesson and build in some more cost efficiency!

[copper nova]

i honestly haven't had any issues with code quality so far, and I do some heavy coding. If you're concerned about the context then look into getting your sub hooked up.

[broken yarrow]

how do i connect openclaw to gmail so it can send mail from chat?

also whats the coolest thing yall have done with it

[copper nova]

look up gog skill

[jaunty perch]

Hi everyone, I set everything up yesterday, All working fine, but woke up to the reply of:

"⚠️ Agent failed before reply: skills.entries.nano-banana-pro.apiKey: unresolved SecretRef "env:default:NANO_BANANA_PRO_API_KEY". Resolve this command against an active gateway runtime snapshot before reading it.
Logs: openclaw logs --follow"

I only set Nano Banana for image generation. The chat only comes back online in the system chat if I re-install. But thr what app connection still shows this error.

Have any of you seen this before?

Tried:

Deletding the Nano Banan reference from .env

[karmic notch]

Yes, I believe I’ve seen this error. I think my agent was able to solve it though? Is yours offline? I think I had to have a Gemini API running alongside nano banana for image generation. I’m a noob, so hopefully this is helpful. I know what it’s like to be offline and frustrated by OC.

[frigid ravine]

Just a general question - does anyone experience slowdowns with OC. meaning the chat on telegram is slow - it takes ages to receive replies and it does not help either that when you give a task - you are unsure whether it is still working in the background as the typing indicator goes for a while then stops altogether. So you have to send a message asking only for OC to tell you that its done and setup and all is well and what is next on the list

[junior acorn]

are you starting new chats? it sounds like you have a lot of context bloat

[frigid ravine]

that is also something i am trying to deal with and handle - however with the previously free Xiaomi Mimo model and with 1M window there was not much reason to clear context. You'd find that at most only 6% used. Or perhaps paid models just respond much faster than free ones

[junior acorn]

studies have shown that 1m window isn't really 1m. performance degrades usually at the max 300k. So you're just wasting tokens tbh

[unreal veldt]

Hi everyone, beginner here, I want to study architecture and working of openclaw to understand its codebase. From where should I start first?

[junior acorn]

I would suggest learning about context windows and tokens and how to properly prompt an LLM. it's a skill issue tbh

[junior acorn]

read the docs, the architecture is explained thoroughly there

[frigid ravine]

interesting - I didn't know that. But that aside - I got frustrated and decided to start fresh restart my journey with OC and set it up again but I am still struggling with the delays. This is on a VPS by the way. It's a fresh install so I should not be having context/bloat issues but somehow i'm still using racking up alot of tokens which is worrying.

[vernal gate]

yeah copilot has context limits and throttles on the free tier. openclaw with a local ollama model has zero limits — your
machine, your context, nobody throttling you

that's actually one of the best reasons to run local. what are you trying to build?

[prime fern]

hi

I am setting up a local openclaw on my linux laptop. ideally id like to set it up for free or nearly free. Ive watched several tutorials. Should I use openrouter or ollama?

[vernal gate]

hey salma! for free go ollama — runs completely local on your machine, zero api costs

openrouter = pay per token, adds up fast

only catch with ollama is your laptop needs decent ram (8gb minimum, 16gb better). what specs does your machine have?

[prime fern]

Thx. I tried Ollama. But I very quickly reached the API limit

which I found surprising. Because I wasn't sure which API I was being rate limited on

[prime fern]

I have a 16GB Acer laptop. And I just upgraded the ram to 32GB

[prime fern]

Thanks for your feedback btw. It has not been easy to find good tutorials for installing on linux

[vernal gate]

totally understand — linux tutorials for openclaw are rough, most are outdated or assume too much

32gb is plenty, you're good there. the api limit you hit was probably openrouter not ollama — ollama local has zero limits

once ollama is running properly you won't hit any limits. if you get stuck at any point feel free to dm me — happy to help
get it sorted

[prime fern]

Thank you so much for your assistance!

actually the previous installation was 100% ollama

i just restored my linux to a pre-openclaw install

so i will start again with a standard ollama install. followed by an openclaw install via ollama. Is that the recommended method?

[slender hound]

Hey everyone
After a weeks use of Openclaw via Anthropic OAuth my account was suspended meaning I need to switch model.

Id like to set it up via Hostinger and follow "safer" recommendations. Do you recommend that I uninstall everything and restart? Or can I install Hostinger and still move forward?

Thanks in advance 🙂

[prime fern]

does hostinger allow full disk encryption?

[slender hound]

I have no clue, what does that mean for my setup?

[prime fern]

For a secure set up I think everyone should be using full disk encryption. AFAIK this doesnt affect openclaw set up. But it should be considered as a "pre-OpenClaw" set up

[slender hound]

Thanks, ill look into it 🙂
Do you however, if I can continue or should start over? aka uninstall everything and then go hostinger?

[honest pivot]

Hey I am trying to access the web ui for openclaw at http://127.0.0.1:18791/ but for some reason I get a simple message "Unauthorized", how do I get over this?

[tight rose]

How do you guys make dashboard?

[vast stream]

are there anyone just failed to setup openclaw into docker in windows, I saw everyone just finished, but when I deleted openclaw in windows and just install docker and put openclaw into it, I met all kinds of problems, and finally, everything fixed, but modify the default model, which means change claude into gemini, but didn't work for kinds of different approaches, anyone can help me?

[vernal gate]

hey! docker on windows is rough — good job getting that far honestly

for switching from claude to gemini in docker you need to set the env variable in your docker run command:

-e GOOGLE_GENERATIVE_AI_API_KEY=your_gemini_key_here

and in openclaw configure make sure you select gemini as the provider. what error are you getting when you try to switch?

[upper niche]

Hey all — looking for help diagnosing what looks like an OpenClaw control-plane / heartbeat issue on WSL2.
Setup
• OpenClaw version: 2026.3.13
• Environment: WSL2 on Windows
• Gateway mode: local / loopback
• Dashboard/UI: works
• Direct agent runtime/chat: works
• Main problem: CLI/admin/control-plane path is unreliable, which makes heartbeat + cron untrustworthy
What’s strange
On the same install/machine, I can have:

• gateway daemon running normally
• dashboard reachable
• normal agent work happening
…but these paths are flaky or failing:
• openclaw gateway probe
• openclaw system heartbeat last
• some CLI/admin websocket/control-plane calls
• heartbeat reliability as a trusted unattended mechanism
• cron confidence for autonomous workflows
Symptoms seen
Examples of failures include:
• missing scope: operator.read
• websocket close 1000 normal closure
• loopback probe timeouts
• CLI/admin path behaving differently from dashboard/runtime path
Examples:
• gateway service/listener appears healthy
• dashboard is usable
• but CLI-side control calls still fail or degrade
Important detail
This does not look like a simple “gateway is down” case.
It looks more like:
• runtime/dashboard path = healthy
• CLI/admin/control path = unhealthy
So I’m wondering if this is:

1. a known OpenClaw 2026.3.13 regression
2. a WSL2-specific loopback/websocket/auth-resolution issue
3. or some known mismatch between dashboard/runtime and CLI/admin scope handling
   What I already checked
   • restart / reinstall / re-run tests
   • local config review
   • paired/auth scope checks
   • gateway service status
   • loopback listener presence
   • normal dashboard/runtime behavior

Those checks did not cleanly resolve it.

Question

Has anyone seen this pattern specifically?

Especially:

• dashboard works but CLI/admin heartbeat/control doesn’t
• WSL2
• operator.read issues
• websocket 1000 closure / loopback timeout
• unreliable heartbeat/cron trust even though the system is partially healthy

[vernal gate]

hey! unauthorized on the web ui usually means the auth token isn't set right

try running: openclaw configure

and make sure you set a password when it asks. then use that password to log into the web ui at 127.0.0.1:18791

did you set a password during initial setup?

[honest pivot]

No, I used a token as my authentication method because it was recommended during the setup, I am going try and set up a password now

[vernal gate]

ah got it — if you used a token, make sure you're passing it in the url like this:

http://127.0.0.1:18791/?token=your_token_here

that should get you in without changing anything

[cyan jolt]

use the cmd openclaw dashboard --no-open to get the url

[honest pivot]

thanks! it worked

[vast stream]

you mean -e GOOGLE_GENERATIVE_AI_API_KEY=your_gemini_key_here, what is the command for this, how to make sure select gemini as the provider, I put the google key into .env, or docker_compose.yaml, or openclaw.json, when docker compose up -d, error always has been: not found claude api key....

[vernal gate]

the issue is openclaw defaults to claude even if you set the google key. two things needed:

1. in your docker-compose.yaml under openclaw-gateway add:
   GEMINI_API_KEY=your_google_key_here

(GOOGLE_GENERATIVE_AI_API_KEY won't work — it needs GEMINI_API_KEY specifically)

2. then run: docker compose run --rm openclaw-cli models set google/gemini-3.1-pro-preview

restart the container and the claude error should go away

[honest pivot]

One more thing, when I communicate with openclaw I hit the rate limit for requests per minute just after one message. Does openclaw send more than 1 api call per request?

[velvet phoenix]

Hello everyone! I'm wondering if any of you have sorted out autonomous developer agent workflow.

The problem I constantly persist is that agent is stopping to code and is not contuining itself until I write in Discord chat: "continue". Meaning that I have constantly write "continue" for him to keep moving forward.

The tasks and workflow is clear for agent, the .md's are full of instructions to move until the end, but it just stops and if you write "continue or whatever push forward message" it only then continues.

I have tried to set-up heartbeat's and cron's to wake bot up every few minutes to resume task if no output is given, but that's not helping at all. I have tried to instruct to move on without any confirmations and etc.

Maybe some of you know the trick to make it work from A to Z without me interacting on every step? The work is being done, but that poking every few minutes is kinda annoying.

[vast stream]

it works, thanks,

[wary osprey]

yeah it can, especially depending on what you have enabled. one message isn’t always just one call, it can trigger multiple like memory, tools, or even follow-up steps, so you might hit rate limits faster than expected even with a single prompt.

if it’s happening immediately, check if anything is looping or calling tools repeatedly. I ran into this early on, and it's part of why I stopped relying on a single provider setup, easier to manage when you can route and switch models without getting blocked

[tall meteor]

Is there anyone here who uses a WebSocket locally with OC?

[sterile steppe]

like from one agent to another?

[nimble hill]

Since i connect over tailscale, it is effectively local - so yes, i connect OC nodes to gateway this way
Shoot your question

[nimble hill]

Using gemini 2.5? Yes it hits it almost immediately. Sign up for a chatgpt subscription - even a lowest tier to experience it first

[mighty rampart]

Which model are you using that keeps stopping?

[narrow egret]

isn't it better to always use pay-per-use? cus i mostly just use a unified API model provider to route my model, I use AIsa and it's pay as you go. I feel like paying per token usage saves more money, just saying. Because i use these high end models like GPT, Claude, and gemini 2.5 on AIsa, and the costs are not that crazy

[quaint creek]

Hi all, what models are you recommending or using? I'm hitting the anthropic rate limits and with the API keys the token usage is growing very quickly

[nimble hill]

Well openclaw is a token hogger. I have burnt lot of tokens by direct API

[robust raven]

Hi all.
I’ve been using openclaw with my OpenAI subscription and I want to change to Ollama and Qwen.
Does any of my setting will change doing this?
I don't want to start from scratch.

[wind warren]

Yo hi guys can someone help me

i installed openclaw on a vm running linux ubuntu 24.04 and thr local website and the gateway token works now I need to put mi api key somewhere but I don't know where

I want to use ollama any suggestion?

[nimble hill]

Go to shell and type openclaw onboard which will give you wizard

[honest pivot]

yeah, I checked the chat in web ui and it does indeed send multiple calls per request, gonna set up some more models as backup

[honest pivot]

Yeah it runs out pretty fast, I was just going for free options first before getting an api subscription since I am trying out openclaw

[quasi walrus]

In my case, I put all api key in ./openclaw/.env ( running VM with Linux Ubuntu too), and make sure it was read when openclaw gateway running.

[narrow egret]

you could also try AIsa, they give like $5 free credits on first signup so you can test different models without committing upfront. that’s what I used when I was getting started.

[hollow bay]

Openclaw.json controls all that so update it, be sure not to overload local models with too much work

[hollow bay]

The work request is the hogger so break it down to smaller requests like sub tasks

[narrow egret]

you don’t have to start from scratch. most of your setup should stay the same, you’re mainly just swapping the provider and model. You will need to update the API endpoint and model names, and if you’re moving to Ollama locally then make sure it’s running and accessible

sometimes small things break like tool configs or formatting, so it’s worth testing one workflow first before switching everything. I went through that a while back, and I got tired of reconfiguring every time I switched models so I just run everything through AIsa now, makes it a lot less stressful

[hollow bay]

Setup one open claw build to be production and other to be test, try playing around in test because it gets crazy when you mess up the environment you need stable

[ashen herald]

Hi all. I'm trying to setup openclaw on fly.io but I can't seem to get it work. I'm only using web ui on my phone as I don't have access to a terminal. I was trying to use the kiloclaw ai to help me but we are going nowhere. Do I need to configure the GitHub repo first before deploying?

[hollow bay]

Really depends on your work needs, i am using the qwen models locally to control costs

[wary osprey]

imo, i will say split your setup instead of relying on one model

for cheaper and solid options, these are the ones I’ve been using lately...

1. Kimi 2.5: really good balance, cheap and handles most tasks well (this is my go-to)
2. Qwen (like Qwen3 / Qwen coder): very cheap for general and coding tasks
3. MiniMax: one of the cheapest for high volume workloads
4. DeepSeek: good for batch stuff and still very cost efficient

pricing difference is actually huge compared to Anthropic, like Kimi and Qwen are way cheaper per token. I route everything through AIsa so I can mix these depending on the task instead of getting stuck on one provider. Very easy to switch between models too

[hollow bay]

Work request is turn based so each turn can call separate tools and get clogged up. Optimize your work request and work load

[velvet phoenix]

codex 5.4

[hollow bay]

Hey Claw friends need your input please. Local model qwen setup running in blocking issues so thinking of breaking out ollama models to run on separate ports like small chat model in separate port from large model i know is going to churn and burn on big work request, what is right way to do this lol everything so complicated

[wet kraken]

Hi all! What is the recommended default thinking level for GPT 5.4 codex via oAuth on OpenClaw that it should be left at? (I'm on the $20 ChatGPT plan)

[stoic anvil]

What is the issue you are facing here ? Any errors shown by fly.io ?

[oblique perch]

Trying to turn sandboxing off on windows . It allows me with some llm’s but not others . Using ollama.. I tried editing the json file with no luck .

[simple marten]

Ciao a tutti, qualcuno usa come host w11? che skills usate?

[ancient jungle]

i want to expound on this, i'm using minimax's token plan ($10/month starter one) you get 1500 requests per 5 hours. unless you're doing something super crazy you're not going to hit that limit.

[rose halo]

I Limited the amount of history messages that are sent on every message I send to 20.

So now I'm not sending a huge useless history on every message 🙂

[rose halo]

I forgot the config key name but your agent can find it

[verbal quiver]

Hey, so I updated to the newest OpenClaw and my webUI does not run .. I have this error -> Control UI assets not found. Build them with pnpm ui:build (auto-installs UI deps), or run pnpm ui:dev during development. But when I actually try to run this pnpm it can't find script.js in the folder it's looking for .. 🙁

[grizzled merlin]

I'm having the same issue

[kindred lantern]

same error Control UI assets not found.

[wary osprey]

That plan is solid if your usage is completely static. But the exact second you start hot-swapping models or scaling a multi-agent swarm, a single-provider ecosystem creates massive drag because you are bound to one billing cycle and one set of limits. AIsa delivers absolute flexibility; you deploy exactly one API key and instantly command multiple frontier providers from a single endpoint. you swap models and monitor everything in one clean dashboard.
This is just what I use (personally, and some other folks use it to. It's more like OpenRouter, but I find it cheaper and more OpenClaw compatible)

[ancient jungle]

what are you using? i didn't see the mention of a router, just a list of models

[wary osprey]

Ohh, I use AIsa as my API router, and i route Kimi 2.5, Qwen 3max, and also Opus 4.6 through AIsa with a single API key. The point is I recommended Kimi because i felt it's cheaper and can get the job done. And since you mentioned this plan on Minimax, I will make sure to check it out. Thanks for mentioning 🙂

[ancient jungle]

yea i looked at kimi, but i might be looking at the wrong thing but it's still $20/month. at that point i might as well just pay ollama.com and use their cloud models

[queen cedar]

Hi guys. I notice something weird when I use some LLM. With Minimx or Codex5.3 my agents can perform action like creating files. But it doesn't work with Qwen3.5-9B (local in LM Studio) or GLM4.7 (from Nvidia API). Anyone has an idea? Both LLM have tools, or coding structure.

[ancient jungle]

i've noticed that, the small LLMs are just too stupid and i'm tired of fighting it

[queen cedar]

but GLM4.7 is not a small LLM and from the Nvidia Card, it has the tools and coding too, so it should works

[ancient jungle]

🤷‍♂️ with minimax 2.7 right now i'm able to code a whole web tutoring app for my kids in openclaw, with it coding and pushing it to github by itself. when i tried to do ANYTHING with something i'm running locally, glm4.7 flash, gpt-oss-20b etc, it's pulling teeth to try to get it to even do anything. it'll make up all kinds of excuses to why it can't

[queen cedar]

anyway, now after the last update, I got this error: Control UI assets not found. Build them with pnpm ui:build (auto-installs UI deps), or run pnpm ui:dev during development.

[cursive peak]

idk what these hooks do during onboarding eg what is hook.md? i think it would be helpful to explain more what they are and what they do

[queen cedar]

so with GLM4.7, you can't do any action like create a folder? can you try?

[ancient jungle]

it's too slow on my 7900xtx honestly. i'd rather not. the juice ain't worth the squeeze if you know what i mean. to run anything on my 7900xtx at a rate that doesn't make me want to smash the computer, i gotta run something smaller like the gpt-oss-20b, and i know for a fact that it's lazy and doesn't want to do anything

[wooden knoll]

Hi good day

[queen cedar]

gotcha. Well, I'm surprise for GLM 4.7, it should be able to work. I will continue to investigate if I find the fix for the last error due to the last update

[wooden knoll]

I am using AWS VPS to install the Openclaw. The onboarding is completed

however, I facing the issue to access the OC dashboard

[fossil vault]

What happened to WhatsApp plugin after update

[wooden knoll]

I think I need to setup the SSH forwarding port?

[ancient jungle]

just spend the 10 bucks on minimax or something and actually be able to do something. use your local compute to do non time sensitive things, like running logic and querying. like i have my local compute generate new question and answers everyday for that homework/tutoring app i built for the kids. i have it processing receipts and OCR and stuff for my self built budget tracker that can track line items on receipts etc.

i wouldn't use it for agentic/chatting, too slow

[queen cedar]

I just asked my agent "create a folder 1234" and he returns :Ok done. But no new folder. Same with Qwen3.5-9b

[queen cedar]

https://github.com/openclaw/openclaw/issues/52808

[ancient jungle]

do you have verbose or whatever else setting in your openclaw that shows you exactly what tool call it's doing? like min looks like this
📝 Edit: in ~/.openclaw/workspace/projects-git/budget-reports/expense_qa.py (455 chars)

[queen cedar]

seems to have a bug with WhatsApp after the last update

so buggy

[queen cedar]

not with this LLM I think, but he does with the other one. I have to double check later

[deft quail]

im running into memory issues? curious how you're all solving it rn.

[ancient jungle]

probably because it ain't actually doing shit. i've noticed that.

[robust raven]

Thanks for your Answer. I take note.
What is Alsa?

[queen cedar]

that's sucks. I will try again otherwise I will change for another one

[hasty crane]

It's nothing. Yesterday Kimi charged me $0.006 and the day before was $0.07. So cents or fractions of cents.

[ancient jungle]

so this interpretation
"last week" = past 7 days
• "this week" = current calendar week (Mon-Sun)?
• "past 2 weeks" = past 14 days

plus the month, year logic.
[3/23/2026 10:28 AM] Crab Tsar: 📝 Edit: in ~/.openclaw/workspace/projects-git/budget-reports/expense_qa.py (1051 chars)
[3/23/2026 10:28 AM] Crab Tsar: 📝 Edit: in ~/.openclaw/workspace/projects-git/budget-reports/expense_qa.py (99 chars)
[3/23/2026 10:28 AM] Crab Tsar: 🛠️ Exec: search "this_month_end" in ~/.openclaw/workspace/projects-git/budget-reports/expense_qa.py, grep -n "this_month_end" ~/.openclaw/workspace/projects-git/budget-reports/expense_qa.py
[3/23/2026 10:28 AM] Crab Tsar: 📝 Edit: in ~/.openclaw/workspace/projects-git/budget-reports/expense_qa.py (147 chars)``` check this out, it immediately goes into actually doing something, so if you don't see exec/edit calls then it ain't doing anything

[queen cedar]

yeah I know, I check the verbose commands all the time to see what he is doing

[ancient jungle]

i just have it on all the time, it's a bit spammy, but at least i know exactly what it's trying to do. sometimes i have to call it out when it's doing something off the walls

[narrow egret]

AIsa is a solid alternative of OpenRouter worth trying.
It’s a unified API aggregator that lets you access multiple providers and more paid skills (Tavily Web Search, Licensed Data, Twitter APIs, SaaS tools, etc.) through one endpoint.
One API key for Kimi, Minimax, Qwen, Anthropic, OpenAI, Gemini, etc.
https://docs.aisa.one/

[verbal quiver]

Anyone knows how to solve this please?

[nimble hill]

Kimi is good - if thay works out for you great.

[robust raven]

I've just started using OpenRouter today and I can't change the model. I don't know what is using ( as the config says openrouter automatic) but is not as smart as the GPT 5.6 i was using.
I'll check Alsa. Thanks

[silk path]

Hi gm

Yesterday I regenerated my claude setup-token and I tried to reconfigure my claw after it started using anthropic api and now whatever I do it says I am rate limited. My weekly limit reset this morning and still gives me the same rate limit issue with sonnet and opus but haiku works. Any idea what the issue is?

When I open the tui I see:

HTTP 429 rate_limit_error: This request would exceed your account's rate
limit. Please try again later. (request_id: req_011CZKPUZwpyhkNz8kx8GMTX)

[versed niche]

How do I change the user account OpenClaw runs under on macOS?

[junior acorn]

it's not going through Oauth, it's trying to use your tokens check ur usage on claude

[junior acorn]

reinstall it

[junior acorn]

holy GPT 5.6, ur goated af

[silk path]

its weird bc it was working before and i tried to regen setup and now it just not working. I did a full reset and still nothing

[junior acorn]

did you update or anything?

What did you do for full reset?

[silk path]

i dont think i updated, maybe a few days ago and then it brokelike 2 days ago so possibly linked

[silk path]

openclaw reset --scope full --yes --non-interactive && openclaw onboard --install-daemon

[nimble hill]

openclaw has a new version just released ~4 hours back. 3.22 njoy!

[junior acorn]

try just doing openclaw onboard and re-doing oAuth via claude code

also send a screenshot of ur claude code usage

@silk path here: https://claude.ai/settings/usage

[deep radish]

i updated my openclaw and my gateway and acp broke krill isnt responding in the other chat i lowk have no clue what im doing can someone PLEASE help

[silk path]

Channel doesnt allow for pics but I just realized anthropic has a sonnet only limit now so that makes sense. I was only checking usage on mobile and it didnt show that. I guess thats the issue

[silk path]

Bet will do again thanks

[versed niche]

Heres my setup and my problem.

My openclaw is on his own machine. I have my own machine. I want to use obsidian vault for clippings, notes, and that is on my machine. But I need to give my openclaw read/write access to it.

That's where I'm having issues.

Any solutions to this?

[sullen ivy]

does claude oauth work any more and how high is the ban risk

[torpid nexus]

What is everyones fallback model of choice rn - im looking to add an openrouter one behind my claude sub. I guess checking the app rankings is a good start? https://openrouter.ai/apps?url=https%3A%2F%2Fopenclaw.ai%2F

[low quest]

Anyone able to help ? I payed for the starter plan and it took money out of my account but I did not get the upgrade ?

[junior acorn]

you can share files over a network, openclaw can explain better than me

[junior acorn]

that's not a problem we can really solve here. You'll need to reach out to customer support for whatever ai company ur using

[versed niche]

We've been working every angle and my agent is stumped and sent me here to ask for help, lol

[covert blade]

My global install of openclaw sudo npm install -g openclaw@latest(26.3.12) on a generous pve vm results in headaches when I openclaw update to apply 26.3.22 I get Control UI assets not found. Build them with pnpm ui:build(auto-installs UI deps), or runpnpm ui:dev during development. in the browser.

[junior acorn]

it's a current bug with the newest build, roll back

[sullen ivy]

use syncthing https://syncthing.net/ to mirror your obsidian vault directory to a directory on openclaw's machine i have used syncthing to sync obsidian vaults before just not for ai agents but the principle is the same

[sterile lotus]

Here is what the fix was for me so you don't have to roll it back

1. cd "$(npm root -g)\openclaw"
2. npm pack openclaw@2026.3.13
3. tar -xzf openclaw-2026.3.13.tgz
4. xcopy /E /I package\dist\control-ui dist\control-ui
5. Remove-Item -Recurse -Force package
6. openclaw gateway restart

[worldly garden]

WhatsApp broke for me

[nimble hill]

Openclaw has concept of nodes. Your machine which obsidian in it, cna be installed with openclaw node which can connect to your main openclaw gateway. Once connected, you can manage any actions in node by gateway

[unreal nacelle]

Hey, y'all! I'm new to AI (Got claude code like 2 weeks ago) and new to openclaw, new to discord... so many news.. not new to computers, started on ms-dos in the before times.. I'm in over my head on this openclaw setup! This seems like the right place to get some tips.. am I in the right spot? Don't wanna break the discord norms

[covert blade]

I had some trouble with the upgrade to 26.3.22 because of missing ui elements. Thanks to @junior acorn for pointing that out and @sterile lotus for providing his windows fix. I'm in a debian trixie pve vm. With the help of copilot web I came up with this script that fixed the problem. Thank you gang!

Now let's see if I can get matrix installed a little more cleanly.

#!/usr/bin/env bash
set -e

echo "=== Detecting global npm root ==="
NPM_ROOT=$(npm root -g)
echo "NPM_ROOT: $NPM_ROOT"
OC_DIR="$NPM_ROOT/openclaw"
echo "OpenClaw directory: $OC_DIR"
echo

if [ ! -d "$OC_DIR" ]; then
    echo "ERROR: OpenClaw not found at $OC_DIR"
    exit 1
fi

cd "$OC_DIR"

echo "=== Downloading OpenClaw 2026.3.13 package ==="
npm pack openclaw@2026.3.13

PKG="openclaw-2026.3.13.tgz"
if [ ! -f "$PKG" ]; then
    echo "ERROR: Package $PKG not found"
    exit 1
fi

echo "=== Extracting UI assets from 2026.3.13 ==="
TMPDIR=$(mktemp -d)
tar -xzf "$PKG" -C "$TMPDIR"

SRC_UI="$TMPDIR/package/dist/control-ui"
DEST_UI="$OC_DIR/dist/control-ui"

if [ ! -d "$SRC_UI" ]; then
    echo "ERROR: Extracted package does not contain dist/control-ui"
    exit 1
fi

echo "=== Backing up existing (broken) UI directory if present ==="
if [ -d "$DEST_UI" ]; then
    mv "$DEST_UI" "${DEST_UI}.backup.$(date +%s)"
fi

echo "=== Installing working UI assets ==="
cp -r "$SRC_UI" "$DEST_UI"

echo "=== Cleaning up temp files ==="
rm -rf "$TMPDIR"

echo "=== Restarting OpenClaw gateway ==="
openclaw gateway restart

echo "=== DONE ==="
echo "Control UI restored. Open http://127.0.0.1:18789/"

[plush hill]

Hi guys, does anyone have queue steer mode working in telegram? I tried everything and couldnt make openclaw steer mid task with new messages

[tall meteor]

I'm running Matrix also on another machine using Docker Compose.
Services:
traefik (reverse proxy)
postgres
synapse
element-web

[loud radish]

im having ui issues as well

[covert blade]

I have matrix in a seperate lxc exposed externally via cloudflare proxy and a pve lxc for traefik, openclaw in a pve vm only local not exposed externally and a self hosted ollama server on bare metal trixie with a dedicated compute rtx 3090. I'm just starting my 2nd week of openclaw starting with 26.3.12 where I had some trouble installing the matrix plugin. Duplicate ids and missing assetts. Now that I have updated to 26.3.22 and stablized the ui I'm just about to try the matrix plugin and config again. I have element on gnome wayland desktop and on my android phone.
I'm off to attempt to install the matrix plugin on openclaw.

[covert blade]

the fix for my debian trixie published in this channel worked perfectly for me. it's based on a fix in windows syntax back a few messages from my message shared with codeblock

[umbral oak]

Could I add my own models if I were to be using this system?

[wet kraken]

Has anyone had the issue where Heartbeat runs in the main session, and while you're in the middle of interacting with OpenClaw, it comes and wipes everything out, and the main session becomes main heartbeat, and none of your conversations there.

[steep bone]

how kan i conect google Worspace i have createt an new acc only for him

[tall meteor]

Very interesting. Thursday will be the end of week 2 for me.
I spent most of my time on building a custom Mission Control with tasks, attach tasks to cards on a Kanban, and drag cards from a Backlog column to a To Do column, and the Mission Control route calls a tiny Node service which wakes an agent to pick up the card and tasks.

I will be integrating Matrix in the future.

[ancient jungle]

that sounds interesting, got a screenshot or something how it works?

[wary sapphire]

Whats the best way to get a good discord bot with openclaw, and for it not to dox you or get all your information stollen by prompt injections lol

[ancient jungle]

just use telegram?

[final gyro]

Hi! Running OpenClaw from source (git clone) on Linux. Trying to get the voice-call plugin working with Twilio. Issue: voicecall.initiate RPC via WebSocket authenticates fine but doesn't trigger any Twilio call. Also getting a port 3334 conflict when using the CLI (openclaw voicecall call) because the gateway already owns port 3334 for the plugin webhook.

Setup:

• OpenClaw 2026.3.23, source install at /home/username/openclaw
• Plugin shows as loaded from stock:voice-call/index.ts
• Tailscale Funnel running, webhook URL reachable (returns 200)
• Twilio upgraded account, number verified, direct REST API calls work fine
• plugins.load.paths cleared, no duplicate warning anymore

Question: On a source install, should openclaw voicecall call use the gateway's existing webhook server (port 3334), or does the CLI spin up its own? Is there a config to point the CLI at the gateway's RPC instead of starting a standalone server?

[unreal nacelle]

https://docs.openclaw.ai/channels/discord and don't let anyone in who's going to attack your system... Or, if you're letting randos in, let me join, My bank account info is bad, I wanna try yours

[wary sapphire]

I was going to let randos in but the bot knows nothing about me and is bassicly just a mod lol

[unreal nacelle]

I'm sure there's no way around that 🙂 I'll be waiting on the invite

[wary sapphire]

Wonder how molty and the other bots here work then

[tall meteor]

I can drop some screenshots in my private Discord and see if they will link. I don't know if this server links screenshots for users with low permissions like me.

[unreal nacelle]

IDK... new here, new to discord... I don't know that molty is running on openclaw.. but I don't wanna clog up the help channel anyway.

[wary sapphire]

yep and krill

[covert blade]

in the web ui when selecting some tab and returning to chat I seem to have two 'main' sections and although main is selected one of my subagent models is selected rather than the GLM. anoying to have to reselct the corect main and model. This is the agents section of my openclaw.json

  "agents": {
    "defaults": {
      "model": {
        "primary": "ollama/GLM-4.7-Flash:latest"
      },
      "models": {
        "ollama/GLM-4.7-Flash:latest": {},
        "ollama/qwen2.5-coder:7b": {},
        "ollama/codellama:latest": {}
      },
      "workspace": "/home/jeffa/.openclaw/workspace"
    },
    "list": [
      {
        "id": "coding-general",
        "name": "Coding General",
        "model": {
          "primary": "ollama/codellama:latest"
        },
        "tools": {},
        "workspace": "/home/jeffa/.openclaw/workspace"
      },
      {
        "id": "coding-fast",
        "name": "Coding Fast",
        "model": {
          "primary": "ollama/qwen2.5-coder:7b"
        },
        "workspace": "/home/jeffa/.openclaw/workspace"
      }
    ]
  },

Is there a section in the documentation I might be missing that might help me validate or hopefully fix to solve my problem?
It seems like I ought to add the GLM-4.7-Flash to the "List": I seem to be breaking the config.
Thoughts? Suggestions?

[eternal flame]

i'm facing a problem downloading open claw can anyone send me a youtube link video ?

[tall meteor]

@ancient jungle I just did the preliminary test of the drag/drop triggering. The whole process took an hour on a RTX 5060 Ti 16GB running ministral-3:14b

13:18:51 → 13:24:53 — dev_manager wakes, reads its runbooks, finds the card, moves it to in_progress, enqueues it, dispatches to dev_coder, and releases the GPU token cleanly.
13:25:43 → 13:28:47 — dev_coder checks out a feature branch, creates the file, releases the token, moves the task to review (triggering the automatic card rollup to Review), then hands off to dev_auditor.
13:29:02 → 13:34:16 — dev_auditor has one small hiccup (branch name truncation on first checkout attempt), recovers by running git branch -a, finds the right branch name, verifies the file, marks QA passed.
13:53:37 — You drag Review → Done.
13:54:04 → 14:14:58 — dev_devops wakes, picks up both cards in Done (the older "Second Task Card Test" and the new one), processes the new card's task: commits, pushes, pg_dump, rebuilds Docker, verifies 200 from https://agent.lan.arpa, merges to development, deletes the feature branch, marks everything done.

It requires a system-scheduler service (Node.JS/Fastify/Redis queue) so that agents don't pile on all at once on the GPU, and they pass a token around like a football and each take turns on the GPU.

dev_devops agent does git merging and deployment, also pg_dump and backup.

[ancient jungle]

i'm still very new to this, what's the advantage of having multiple agents?

[tall meteor]

Each agent has specific abilities and roles defined in SKILL.md

[ancient jungle]

why can't my main one just do it all?

[idle stratus]

I guess you would clutter the context, I'm also using one agent for all, but reconsidering purposing different agents

[tall meteor]

That way a single agent isn't a jack of all trades, getting mixed up with too many abilities

If you're running local LLMs

Also, a low-level agent which doesn't require deep reasoning, it can be set up to use a cheaper LLM

[ancient jungle]

the only thing i'm using local compute for is in my projects, i'm still using cloud based for my agent/chats

[tall meteor]

Yeah, it's definitely a different approach when everything is local

Resource management

When you're dealing with a mega datacenter, then money is the resource, if you got it

[near ember]

I am facing problems when I installed openclaw. I am trying to configure googlechat but get "googlechat plugin not available." in openclaw config. does anyone know how to fix

[ivory carbon]

OG

[ancient jungle]

anyone have issues with the agent not actually checking everything in heartbeat.md? seems to just stop at the first instruction

[mighty rampart]

Use one agent until you have a reason for different ones. Maybe one will be fine for you, it depends on what you are going to do with it.

[vernal gate]

hey! what device are you on — windows, mac, or linux? there are a few different ways to get it
running depending on your setup, happy to point you to the right resource 👋

[ancient jungle]

what's in the new openclaw update? is it actually stable

[finite sable]

guys, why is the matrix option from channels gone???

whose incharge, who removed matrix option do they think it's something nobody uses

[woven birch]

Guys, is it currently possible to work in openclaw using a Gemini subscription instead of the API?

I've been trying to set up the subscription-based version for two days now, but I can't get it to work. Everything works perfectly via the API, but the limits are very low. I want to set it up so it works via subscription.

Who use Gemini subscription instead of API?

[dim harbor]

you havent read the release notes ?

Plugins/Matrix: add a new Matrix plugin backed by the official matrix-js-sdk. If you are upgrading from the previous public Matrix plugin, follow the migration guide: https://docs.openclaw.ai/install/migrating-matrix

[haughty thistle]

Hey, trying to connect a Google account via gog auth add but getting error 403 access_denied. Can you add
[e-mail] as a test user for the Google OAuth? Thanks

[finite sable]

oh my

then why is it throwing errors

[copper nova]

I think Google been cracking down on people with antigravity subs using it, I tried and it was flaky, but that was also like 5 updates ago

I just use their API now cus it's got $400 of free credits

[finite sable]

it's asking me to do openclaw matrix, but the matrix command doesn't exist

[woven birch]

wow how you gets your free credits? Which page?

[tall meteor]

Last night about midnight, I was using Claude Pro plan, and it went full-idiot. It was a new session and seemed to be throttled or something. Very strange. Today it is back to normal. This is the second time in one week I've experienced this.

[ancient jungle]

i ran into the same issue on minimax

http 424 errors was what i was getting

[finite sable]

Guys, I need human help does this server not have any human staff

[tall meteor]

I wasn't seeing any errors, but the reasoning was very surface level, and seemed lobotomized. Maybe it was an update.

[copper nova]

sign up on google cloud AI studio, they give everyone free credits

but be warned, the rate limiting on free tier is crazy bad

[finite sable]

Can someone tell me why I'm getting this
Plugin "@openclaw/matrix" requires plugin API >=2026.3.22, but this OpenClaw runtime exposes 1.2.0.

[copper nova]

you need to enable billing to get better rate limits

but it's still flaky after with ghost rate limiting

I honestly hate gemini API, I only use it for things I have to

[woven birch]

Which must have model for your oppinion?

[sweet bobcat]

Is it just me, or has OpenClaw's response time gotten way slower over the past 2-3 releases when talking to your agent over Discord?

[copper nova]

gemini flash 3.1 for grounded web searching, nano banana 2 for image gen

that's basically it

everything else there is trash

oh I guess the embeddings for memory indexing I use as well

[fallen steeple]

Did anyone figure out the Control UI - Broken fix. Without messing up what you have going on

[woven birch]

i mean, you say that you use gemini api some times for some things, but which model you basicly use for talking with openclaw?

[dim harbor]

interesting, how are you doing the web search with that? does gemini replace the search api or just the model that is used while searching ?

[dim harbor]

rollback to version that works for you and test it before updating again in a seperate environment

[woven birch]

I'm usually using serper for websearch, brave make they api only with billing, but serper give free credits (easy spawn account for many credits)

[copper nova]

I use sonnet or opus depending on what I'm doing

I don't casualy chat with my bot, not looking to date him

[copper nova]

the model determines the search api used, and google is the best results I've come accross, I'm looking to experiment with brave and perplexity at some time but been lazy

[woven birch]

You using that via subscription or api?

[copper nova]

sub

[woven birch]

Easy for connecting?

[dim harbor]

so when using that it doesnt even need the brave search api ? i got brave search api set up, just wondering if theres anything to gain by adding gemini 3.1 flash in the mix somewhere, and which part does it actually replace

[finite sable]

how do I rollback

[woven birch]

I've been trying to connect Gemini to OpenClaw via a subscription for the past couple of days, and it's fully shit - I'm at my wits' end with it.
Is it different with Claude? Do they still let you connect their model to OpenClaw via a subscription?

[copper nova]

yes, just use openclaw onboard and follow the steps

[dim harbor]

check here: #1485695158828798012 message

[copper nova]

implement it with your brave search and compare the results, I'm curious, tell me if brave is better lol

[woven birch]

Searching in Gemini is the best
Other models use 50–100 sources, while Gemini uses 500–1,000 in in-depth researches

[dim harbor]

is that with gemini app or even when using it through api with openclaw?

[finite sable]

I was having the control UI issue as well tbh

[feral chasm]

I was tryna go thriugh documentation but the site never loads, does anyone know why

[dim harbor]

yeah new version is a bit buggy, im waiting for a few fixes before trying it out

[finite sable]

I'll probably downgrade then-

really was excited for the reasoning update

[dim harbor]

yeah cant wait to upgrade, it looks awesome. just cant spend time getting it to work, rather keep a working claw for right now

[graceful parcel]

i guess i picked a bad day to install openclaw for the first time. cant get past the control ui assets not found. i think i will just wait until someone fixes it.

[mighty rampart]

It’s in GitHub issues, there is a comment to expect a release with a fix shortly. The gateway UI in 2026.3.22 is problematic. Slack is working fine though.

[graceful parcel]

i figured. im not worried.

[open saffron]

Whew. New update is responding in triplicate over telegram.

[unique quarry]

afternoon folks (at least for me) dashboard was working, updated openclaw and now the dashboard is throwing

Control UI assets not found. Build them with pnpm ui:build(auto-installs UI deps), or runpnpm ui:dev during development.

any pointers would be appreciated

[peak viper]

Its an issue with the release. The gateway is running and working as expected. Just the UI isnt working right now because the NPM bundle doesnt include the UI assets. A fix is coming soon(tm)^

[sullen coral]

In the meantime you can use it with terminal or discord without problems.

[vernal gate]

hey! yeah there was a rough update recently that broke the control UI for a lot of people — you're
not alone on this one 😅

quick fix that worked for others: try running openclaw update in your terminal first, then restart.
if that doesn't work, what OS are you on? windows or mac?

[graceful parcel]

old mac mini intel 16gb. update not working, downgrade to 2026.3.13 not working. tried doctor many times. its ok, no need to trouble shoot, id rather wait until its fixed and at least start with a good install. then look at other "normal" problems later

as other said, its running, just not the control ui

might grab a mac mini m5 when released. just wanted to try it out now to see if i like it. ive been using openclaw on kilo.ai as a test.

or kiloclaw as they call it.