users-helping-users

[spark valley]

yeah just tell the main agent to do it

[sweet marlin]

Yeah, that's why I'm just trying to figure out what the cheapest way to do it at decent scale initially.

with good quality

something like Gemini flash that can connect to chat GPT to review content and send it back to Gemini as Chat GPT doesn't seem to be penalizing anyone for subverting the API access.

Via playwright.

[delicate plover]

Yet. 😉

[finite condor]

What do you mean? Tell the main agent to build the sub-agent?

[sweet marlin]

But then I'm just thinking what the fuck Did I just pay for the max plan for if this was what I'm going to do lol

To be honest, I think ChatGPT gives a lot better answers than Opus 4.6 Anyway.

[spark valley]

yeah

[spark valley]

well ppl disagree i didnt use both so i caant say none but it seems like opus is superior

[rotund cairn]

Thank you!

[sweet marlin]

I don't know. I've spent weeks trying to set up claude code to be better on opus 4.6 But Out of the box chat GPT just keeps improving on It's outputs.

[spark valley]

always here to help

[wintry swallow]

Guys, does anyone use openrouter opus?

[tall meteor]

I'm still stuck with manual added Ministral-3:14B which can't use tools. Unless I automatically detect local models with OC onboarding, models don't have tool tags.

$ openclaw models list

🦞 OpenClaw 2026.3.11 (29dc654) — I speak fluent bash, mild sarcasm, and aggressive tab-completion energy.

Model                                      Input      Ctx      Local Auth  Tags
ollama/ministral-3:14b                     text       16k      no    yes   default,configured

When I ask the agent to do a simple ls -al on a directory, it fences the tool command as a chat text.

[spark valley]

most common issue is that your agent is just a messagng

try to open the config and find tools chaange from messaging to full

[tall meteor]

Is that in openclaw.json, or one of the .md files?

[spark valley]

yeas

.json

or you can edit it form th ui web

go to agents choose urs and look for tools

[tall meteor]

I've gone through 20 iterations of the .json file. I am unsure. I do validate, but a lot of recommendations I've tried are invalid.

[spark valley]

use the web ui

[tall meteor]

Gemini, ChatGPT, and Claude are no help whatsoever 😄 they treat technical config like bar trivia and guess all the time.

[tall meteor]

I'll definitely do that. As a coder, I always avoid GUIs

[spark valley]

if you want mmy top tip for fixnng issues with ai is to paste it in googlee bar search and use the search ai as it goes throught all the github issues and every website and give ear accurate fixes

[wind yacht]

Hi, I want to upgrade from version 3.11 to 3.13. While reviewing the changelogs I cam across "Models/plugins: move Ollama, vLLM, and SGLang onto the provider-plugin architecture, with provider-owned onboarding, discovery, model-picker setup, and post-selection hooks so core provider wiring is more modular.". I mainly use ollama-cloud, so do I need to make changes or will the "old" provider-config still work?

[tall meteor]

@spark valley I'll buy you a ☕ thanks

[merry siren]

Yes! I was getting frustrated because they're literally making up commands as they go

[ancient jungle]

i have 3 glm-4.7-flash variants installed on ollama, glm-4.7-flash, the q4 and q8 variant. they're all in the json, my bot confirmed it, but i can't switch to them using /model

to be clear, i can switch to the base ollama/glm-4.7-flash, but not the quantized variants

[digital zinc]

Guys, I need your help. I am no expert in ai. I am using ollama as my ai for openclaw, it works as i talk to the ai, but whenever i give him a simple task for example reminding me something at a time or have it send something every hour, it never works. I tried downloading a couple skills from clawhub but still it doesnt work. The ai also does not send messages itself unless i asked to.

[spark valley]

thanks my friend

I'm creating a minimal guide for beginners to get their first Telegram agent running. I’m looking for people to use this guide for free and provide feedback to help me improve it (not self-promotion).

The start is always the hardest, and some people even give up on the whole experience because of silly issues and end up wasting the opportunity.

[spark valley]

could you provide the log

do

openclaw log --follow /// and ppaste here

[spark valley]

ahh typical issue same as @tall meteor 's you need to give it full tools

[ancient jungle]

nevermind, i think i had it fix itself. don't know why it didn't the first time around

[spark valley]

and install this skill for it https://clawhub.ai/ivangdavila/self-improving

[digital zinc]

Just this and it would work?

[spark valley]

no you would need to give it full tools to execuete commands

go to the web ui then to agents then to tools set it to full

openclaw config set agents.defaults.tools.full true

or this commandline

[austere kelp]

Hi all anyone got an idea why Krill isn’t responding today?!

[spark valley]

his day off lol

[hasty grove]

@valid kettle @blissful berry @olive cliff hi, hope to talk w/ each of you today, I’m caffeinating, moving to library and collecting my thoughts rn

[wide cliff]

Hoping someone may have advice today: Using OC ~3 months and API costs are ~$1500/wk (I'm ok w/that based on what it's doing) and thats w/moving non-complex tasks to lesser models, etc. I can't find a model that comes close to sonnet for reliably doing coding or complex tasks. Hoping someone can point me to a discussion/resource for people facing same issue or heavy discussion about optimizing for API costs or model comparisons - specifically for Open Claw.

[tall meteor]

I totally hate the GUI for OpenClaw Dashboard. I can't seem to find what I need for tools.

[wide cliff]

(BTW, it’s a complex, 7-round system that API hits 6 separate frontier models w/very large prompts and then collaborates between each of them between each round (those aren’t the $ issue— its the brain/complex stuff). I need full API: 100K+ token prompts, 6 parallel model calls per round, no rate limit headroom. A Plus sub would hit its cap in the first few minutes of a session.)

[spark valley]

I agree, it’s buggier.

[tall meteor]

So then, openclaw.json is my only option? I really do wish there was a spec for all the keys and values

[spark valley]

well you can use free models or cheap for non complex tasks and make the agennt use that models

[spark valley]

yup i though you fixed it

[tall meteor]

Nope - you gave me good advice to use the GUI, but once I was at my machine, I could only see adding custom params. I think those are for API calls which need special parameters, not for standard tool calling.

[spark valley]

ohh my bad didnt point you right

go to agents

in the sside bar

[tall meteor]

OK

[spark valley]

choose urs then tools

set to full

[tall meteor]

I set them to full, then openclaw gateway restart
Then asked, "List the files at /opt/mission-control-dashboard/app"

The agent replied:
read[ARGS]{"file_path": "/opt/mission-control-dashboard/app"}

It's fenced as a chat message and not a tool command.

[spark valley]

Well, there seems to be an issue with the web UI setting the configs. I want you to check the config and see if the tools section contains "full".

[tall meteor]

OK

  "tools": {
    "profile": "coding",
    "exec": {
      "applyPatch": {
        "allowModels": [
          ""
        ]
      }
    }
  },

[sweet marlin]

Does anyone have a better way to set it up than this?

• Gemini Flash as default brain — $0.075/1M tokens (basically free even at heavy use)
  • ChatGPT via Playwright for content review/auditing — $0 (your existing Plus sub)
  • SunClaw → Railway free tier for hosting — $0
  • Brave Search API for web search — free tier (2K calls/month)

@spark valley What do you think?

[spark valley]

this actually very smart

just for railway i don't think it still support free hosting

for backend"

[sweet marlin]

Oh. What do you suggest instead? Also, what about the heartbeat? I hear it can cost a lot if not set up right.

[dim harbor]

you can use chatgpt plus sub as oauth in openclaw directly, the whole playright wrapper seems unnecessary, unless you dont have credentials to log into chatgpt

[spark valley]

yep the heaartbet basically is just a reminder to see if it has scheduled tasks so at 30 min is good

[sweet marlin]

Noted, thank you.

Are you sure it doesn't cost API credits though fo the latest gpt model

[drowsy leaf]

Someone smarter than me, ive read that running a smaller context is better to avoid making the agent going "dumb". Sounds counter intuitive to me, is it always the case? Am trying to optimize for a 3060 8gb.

[spark valley]

it depends on the use if you use it for chatting is not aas if you use it for coding or research

[dim harbor]

yeah, im using it myself, i dont have openai api set up at all. no credits used. ofcourse on heavy use you can run out of usage.

[sweet marlin]

Yeah, I think I'm zero auth for codex for reviewing code already. I don't know why I didn't think of this. I didn't think I could use the latest GPT model doing this.

[drowsy leaf]

I plan making it browse the internet somehow, and it gets nasty just after a few commands lol I have to /new frequently haha

[spark valley]

it is basically like using chatgpt

[spark valley]

running local

?

[drowsy leaf]

Ah yes, qwen3.5 4b

[dim harbor]

yeah it uses same weelky quota as codex. ran out of it myself first time this week on the plus sub, in 5 days. pretty good for the plus subscription. you can use gpt 5.4 with it or the codex models

[bright ruin]

Hi guys, was wondering can Openclaw connnect to Windows browser when it is in WSL? My claw tells me that it's not quite possible and requires port forwarding, tried but not work

[spark valley]

@drowsy leaf You should use free models instead of struggling with a local setup, as 8GB of VRAM is just the starting point for running local models. It’s better to start with free models to get good results, and then, if you want to switch to local later, it’ll be easier to know which model works best for your task and how many parameters you need.

[dim harbor]

yeah i dont think you can do it. wsl is like a virtual machine. maybe if there is headless browser that can run inside WSL itself? never ran it so cant really help with it more than that

[spark valley]

it is you should install chromuim

and use it gui

[bright ruin]

so only accessing chrome in WSL right

[spark valley]

but not your actual browser as t is like in a seperate computer

[spark valley]

yep

[bright ruin]

Cool, how did the extension work though, that's possible

[drowsy leaf]

Thank you! I'll take a look. Did try some free ones but I burned the limits very fast haha I'll research more after work

[bright ruin]

possible to use my real profile in Windows

[spark valley]

i can give you best limitless free models

and strong as hell

[drowsy leaf]

Is limitless a real thing? I thought most of them had limits of some sort

[high heron]

Use codex

[spark valley]

well always theres a limit but you will never hit it

[high heron]

Best value

[high heron]

Codex oauth I mean

[spark valley]

i do 500m tokens per day and almost 800 requests

so i dont know what the limit

[drowsy leaf]

I'll take a look at it!

[high heron]

Or alibaba’s coding plan

It’s also solid

[drowsy leaf]

We can always try hitting it haha

[high heron]

Railway is absolutely trash

And don’t waste ur time setting up openclaw on free plan

Even on paid plan, it’s sht

Use other vps

[spark valley]

you can i can give it to u it just openrouters but withoout the 50 daaily limit

[high heron]

Use OCI if u don’t want to spend money

[spark valley]

real talk i use oracle free vps to host

[high heron]

OCI is the goat ngl

[spark valley]

they can even get the A1 model which is fricking huge 24gb ram

[high heron]

Railways 1gb ram doesn’t do sht

[high heron]

[high heron]

Paid plans volume doesn’t save anything 💀

Everything’s gone after redeploying

[spark valley]

yea frr ,the best way to go is vps

it is managable

[toxic blade]

Hey guys! I have a question. So I can do multiple isolated agents, but because channels connect to the gateway it will not let me create separate apps for each agent in slack for example?

[cunning meadow]

hi @spark valley which limitless free models are you referring to? I am running local LLM just to explore openclaw and configure it, but it is slow and usually hallucinates settings. I also tried openrouter but got rate limited immediately and tried ollama cloud and that also stops you very quick.

[sweet marlin]

INTERFACE
Discord (your PC)

HOST
Oracle Cloud Free Tier (ARM A1, 24/7)

BRAIN FILES
SOUL.md | AGENTS.md | USER.md
MEMORY.md | HEARTBEAT | TODO.md

MODEL STACK
Brain ......... Kimi K2.5 ........ FREE
Heartbeat ..... Haiku ............ $0.50/mo
Coding ........ Codex GPT 5.2 .... $0 Zero Auth
Web Search .... Deepseek v3 ...... FREE
Content ....... Kimi K2.5 ........ FREE
Voice ......... ChatGPT 4o RT .... $0 Zero Auth
Understanding . Gemini 2.5 Flash . FREE

SKILLS (bundled only)
Brave Search | Agent Browser
Summarize | Self-Improving Agent

[spark valley]

openrouter stepfun

[sweet marlin]

Okay, does this seem right? Or any additions?

[cunning meadow]

Thanks and is that one good for just general openclaw use e.g. not heavy coding? I assume with your experience it has been good for that?

[spark valley]

yeah it is good

even for coding

i caan share some other models

they drop models every day

[sweet marlin]

Yeah, for sure. If you have more experience with it, I guess I don't use the open source ones.

[spark valley]

i use one for cybersecurity and it does very good

it has 1 m context

[sweet marlin]

What's it called?

[cunning meadow]

you mean openrouter drops new free models? i think i tried one of the gpt-oss ones which were 'free' and i kept getting errors.

[copper nova]

is there anyway to fully disable the heartbeat stuff? I can't seem to actually disable it with "0m" or "none" params.

[blissful berry]

Isn't the heartbeat kind of important?

[cyan pumice]

Oh snap, got it to connect after switching to ethernet connection from wifi. I'm going to have to investigate that later. Seems to work now (was unreachable on wifi. Huh...) Thanks again. (I'll look at the logs too, dunmo why I didn't earlier. Maybe more info there)

[spark valley]

no they have 50 request limit

[cunning meadow]

Do I need to have some credits on openrouter in order to open up more request limits?

[spark valley]

yea you could but not nececeraly

you can get up to 1000 daily request on all free models

[spark valley]

yea always check the logs happy it worked

[calm mason]

does anyone know why mine keeps saying:

Exec approval is required, but chat exec approvals are not enabled on Discord.

it's really annoying me lol

[cunning meadow]

My local LLM qwen 8b constantly hallucinates setup solutions that don't work. It's not actually helping me fix any issues I have

[spark valley]

look go signup for oprnrouter now get the api add it in oc and for models choose manual

i will give u the model that is not listed there which is free

[cunning meadow]

I already got an API key for openrouter. You want to share it here?

[spark valley]

noo

here this model got 1m token contexxt openrouter/hunter-alpha

i want you to run openclaw configure

then select models then openrouter then custom model

and paste that model id

[spark valley]

Hunter Alpha is a 1 Trillion parameter + 1M token context frontier intelligence model built for agentic use. It excels at long-horizon planning, complex reasoning, and sustained multi-step task execution, with the reliability and instruction-following precision that frameworks like OpenClaw need.

1 trilion compared to 8b

💡 Guys i need you i want three or four people to send them my guide on getting their first telegram agent running free of cost with the issues and fixes , i want them to use it and tell me other issues they found in their way to help me improve it , it is 5 bucks but i will give it for free for them so they can review it , {only people that starting up and need a clean guide so they can give me real problemss they face} .

[cunning meadow]

Trying hunter-alpha it now. It seems to run fine right now after adjusting the privacy settings to let them train on my input data... will find out how long I can go for before it gets rate limited. Fingers crossed

[spark valley]

yea try to max it

it is bit slower than stepfun but it is better in reasoning

[tall meteor]

What kind of hardware runs Hunter Alpha? It has to be insanely expensive.

[spark valley]

it is free

but with prompt loggin

basically it will not be there forever it is using our prompts to enhance

im okay wwith it

[tall meteor]

That's exactly what I was just thinking

farming our data

[spark valley]

well it like you are training their model and they let u use it for free

[tall meteor]

I'm running openclaw onboard again to see if I can find out where I went wrong.

[spark valley]

btw did you fix the issue

[tall meteor]

I did not. I have been working on it for 7 days.

When I started with qwen3:14b, I was able to do some coding, it set up NextJS, Prisma ORM for Postgres

[spark valley]

i just saw your message earlieer sorry it got pushed with the mentions

[odd bear]

does anyone have a proper guide on how to set up openclaw with docker on a vps (NOT Hostinger) the official guide on docs doesnt work for me

or can help me setting it up in dms?

[spark valley]

"tools": {
"profile": "full"
},
only remove the exec and other params

[tall meteor]

Somewhere along the line when I downloaded qwen3.5:9b and ministral-3:14b, manually editing the openclaw.json, the tools didn't have write permissions and couldn't even read directories.

[tall meteor]

Yes, I've done that as well, and probably 100 iterations of the same openclaw.json file

[spark valley]

i got a guide it is only 5bucks but i will give it to you for free to get a cleann install but you givee me your review on it

[odd bear]

alright

[spark valley]

dm

[spark valley]

you need to use the first working backup json

[tall meteor]

I suppose. I'm just trying to walk through the onboard again to see if I can autodetect the new models. Something tells me that the autodetect is able to register the Ollama models with tool access, where manual editing of openclaw.json does not.

[spark valley]

it is labeled as openclaw.json.bak.2

[odd bear]

yes

[spark valley]

WELL YOU CAN add custom model on the onbooard or the configure commands try that and see if it does any good

[tall meteor]

Yes! That's what I'm looking into. It's currently downloading "glm-4.7-flash", which I assume is just a base model for OC

[dense bronze]

My openclaw ai keeps sending a link to me for the gateway but it doesnt load or anything, im confused pls help

[spark valley]

could you paste the links or the log

[dense bronze]

I tried it again and it not says not authorized

i found the website and it says bridge missing

[hasty grove]

Any of you three free around or after 5:00pm ET?

[dense bronze]

im rlly begginer here so sorry if my questions are stupid

[tiny bramble]

Hey! Is full desktop control (native macOS UI automation) on the roadmap?

I'd love to be able to have my agent click through macOS permission dialogs, interact with native apps (like Contacts, Calendar), and generally control the desktop — not just the browser. Right now I get stuck when a system popup appears and there's no way for the agent to dismiss it.

Is there a skill or config for this? Or is it something being worked on? Would be super useful for automating things like adding contacts, approving permissions, etc.

[hasty grove]

I doubt ur questions are stupid, but do u want tips on how to ask a smart question?

[spark valley]

it's okay we all were

[dense bronze]

sure bro id love some

[hasty grove]

http://www.catb.org/~esr/faqs/smart-questions.html

[dense bronze]

thanks

[spark valley]

hey is it okay for you to do a clean reinstall?

that may fix your problem

[spark valley]

ask it if it can create a skill for that

[dense bronze]

ok thanks sir

[spark valley]

okay i can send you my instruction guide book for free so you can get a everything setup

but you are required

to review

so i can enhance it

[tiny bramble]

i tried... didnt work ugh...

[dense bronze]

yes sir i will do so

[spark valley]

is the tools set to fulll

[wintry swallow]

How much money do you spend on using Claude Opus 4.6 on average per day or week?

[tiny bramble]

also... How do I set up OpenClaw to handle two separate users cleanly?

I have two people texting the same OpenClaw instance via iMessage — myself (the owner) and my dad (a collaborator on a separate project). A few issues:

Identity confusion — the bot sometimes mixes up who it's talking to or sends replies meant for one person to the other
Project separation — I want it to know that Person A works on Project X and Person B works on Project Y, and not cross the streams
Context isolation — ideally each person has their own conversation context so one person's questions don't bleed into the other's session
Is there a recommended way to set this up? Separate channels? Per-user session config? Any guidance appreciated!

[dense bronze]

yes please thank u so much

[spark valley]

it okay drop a dm

[dense bronze]

this is what it says rn 🦞 OpenClaw 2026.3.13 (61d171a) — If it works, it's automation; if it breaks, it's a "learning opportunity."

│
◇ Reset scope
│ Full reset
│
◇ Proceed with full reset?
│ Yes
Recommended first: openclaw backup create
Stopped Windows login item: OpenClaw Gateway
Removed ~.openclaw
Removed ~.openclaw\workspace
Next: openclaw onboard --install-daemon

C:\Users\SAVAG>

[spark valley]

i see you runing it on windows

[austere kelp]

Any admin here who can tell us when krill likely to be online again please?

[tall meteor]

I've been running Linux since 1999
I haven't used Windows since 2012

[spark valley]

that penguin says it all

[finite condor]

I have issues with my claw, it stops responding very often

I have to probe it regularly to get it to finish a task (which it usually does fine)

[ancient jungle]

how can i get this crab to use private-web-search-searchxng by default instead of the web search tool?

[distant bough]

Krill on strike.

Anyway to fix when doing a big task with multiple steps, or even a simple task which handsoff to other agents, I always need to ask for a update for it to continue it's work. Or get the information from the other agent. How can I stop this? I want them to keep working untill it's done.

Theese YouTubers who say there openclaw works overnight etc etc. how is it possible lol?

Does a mission control handle this?

[finite condor]

Same feeling over here -_-

[distant bough]

Yeah annoying.

[finite condor]

I keep asking it “are you still on it”? xD

[languid viper]

Hey everyone 👋

We filed a feature request that I think matters to anyone running OpenClaw agents in production:

Native outbound approval gate : human clicks ✅ before the agent sends emails, SMS, or messages.

My agent has sent invoices without approval, leaked debug messages in a group chat with a client, and once sent the same email 6 times. All because there's no built-in "are you sure?" before outbound actions.

If you've ever been embarrassed by your agent acting without permission, please upvote 👍 on GitHub:

👉 https://github.com/openclaw/openclaw/issues/25145

Thanks! 🙏

[blissful berry]

I'm having constant issues with models following fixed roles/tasks.

maybe the 30b/32b models I'm running are just too weak?

[finite condor]

I wonder if it is not linked to compacting somehow - is there a way to trace when compacting happens?

I dont see anything on the main chat, just some warning about context counts

I think my claw does not compact properly

[odd bear]

origin not allowed (open the Control UI from the gateway host or allow it in gateway.controlUi.allowedOrigins)

how can i fix this in the gateway dashboard?

[wide cliff]

Already using the low cost models for all of the basic work and even the multiple agents. im looking for something that can come close to sonnet (gemini 2.5 / 3 dont work well & minimax doesnt work well) for less $ than sonnet --- for the complex tasks that no other model i can find is able to do without messing up the code or not ignoring rules or parts of the work to be done...

[tall meteor]

My last interaction with Claude 😆 AI is worthless

If you don't know where the documentation is located, how were you looking it up before?

1:49 PM
I wasn't. I was guessing. I have not looked up any OpenClaw documentation during this entire session.

Can you give me the documentation URL so I can fetch it before responding to each screen?

[thick edge]

Hey i did the GlM pro plan. It works great with open claw. having issues with it not building the mission control web app where it won’t fix the place holder information. so i’m trying a new method of breaking down each part of the app individually and focusing on smaller segments. I hit my 5 hour limit this morning lol. So i’m trying to see if i can use 4.7flashx to talk to for lower rate consumption and it calls glm5 for large coding tasks. Thank you for your help because i’ve made it a day without openclaw killing itself finally

[ancient jungle]

what models are you guys using? i've tried GLM4.7-flash, it's good, but it's a bit slow with what i got. i've tried qwen3.5-9B, it's fast, but it tends to get stuck for some reason, like it'll just stop answering

[spark valley]

easy you need to accespt the device

[odd bear]

how?

[spark valley]

openclaw config set gateway.controlUi.dangerouslyAllowHostHeaderOriginFallback true

or openclaw config set gateway.controlUi.allowedOrigins '["http://YOUR_SERVER_IP:18789"]'
your server ip is your locaal ip or loopbakc

[odd bear]

now its saying: control ui requires device identity (use HTTPS or localhost secure context)

[spark valley]

openclaw config set gateway.tls.enabled true

openclaw gateway restart

[odd bear]

now the website doesnt load at all 😭

[spark valley]

did you restart

the gateway

do

openclaw log --follow

[odd bear]

i sent it to you in dms

[lofty night]

Any good trouble shooting steps around figuring out why OpenClaw has been stuck on this error for 8 plus hours?

API rate limit reached. Please try again later

OpenClaw 2026.3.13 (61d171a)
🕒 Time: Tuesday, March 17th, 2026 — 2:57 PM (America/New_York)
🧠 Model: openai-codex/gpt-5.4 · 🔑 oauth (openai-codex:default)
📚 Context: 0/272k (0%) · 🧹 Compactions: 0
📊 Usage: 5h 100% left ⏱4h 59m · Week 0% left ⏱4d 0h
🧵 Session: agent:main:main • updated 1m ago
⚙️ Runtime: direct · Think: off · elevated
🪢 Queue: collect (depth 2)

I can get to the failover local model.

But what is causing this? Am I miss reading the Usage line above?

[hasty grove]

When I control my own context my 8b model generates YAML outputs consistently well

[copper nova]

not if you don't use it, which I dont. All my jobs are scheduled crons, not triggered by a heartbeat. Heartbeats uses LLM to trigger more LLM calls, not cost effective to use.

[blissful berry]

@hasty grove Interesting, I'm using 27-32b models and they are struggling with instructions

[mild minnow]

If anyone can help with this I’d be eternally grateful. I’ve been bashing my head over a few issues.
First being discord ‘not authorized to use this command’ even though the json has all the needs IDs
Second is that every time I try to fix the above issue, the openclaw gateway throws a config error.
Lastly, it seems when ever the gateway does restart, there’s a persistent ghost gateway running leading to port errors.

I’ve seen posts on the ghost and discord config on answeroverflow but nothing worked for me

[tall meteor]

@spark valley I backed up my openclaw.json, ran openclaw onboard and was able to autodetect the Ollama LLMs on the main machine, set the "contextWindow" for ministral-3:14b so it is sized appropriately for my RTX 5060 Ti (16GB), restarted and tested the agent.

Here are the files and directories in `/opt/mission-control-dashboard/`:

### **Files:**
- `docker-compose.yml`
- `.gitignore`
- `HANDOFF.md`
- `README.md`

### **Directories:**
- `app/` (likely the main application code)
- `.git/` (Git repository metadata)
- `traefik/` (likely Traefik configuration or assets)

Would you like me to inspect any specific file or directory further?

Success 🎉

[blissful berry]

hmm I haven't had issues with the discord connectivity

[tall meteor]

The autodetect was necessary when adding the new LLM, because the tools weren't reported by Ollama.

[spark valley]

great to hear thanks for uploading the fix

[blissful berry]

maybe paste your openclaw config for the discord bits (api key and such redacted first)

[tall meteor]

I don't have Nitro. It would have to be in small posts.

[mild minnow]

{
"enabled": true,
"groupPolicy": "allowlist",
"streaming": "off",
"allowFrom": [
""
],
"guilds": {
"REDACTED": {
"users": [
""
],
"channels": {
"REDACTED": {
"allow": true,
"requireMention": false
},
"*": {
"allow": true,
"requireMention": false
}
}
}
}
}

[blissful berry]

interesting, that's different than what I'm using

[tall meteor]

{
  "meta": {
    "lastTouchedVersion": "2026.3.13",
    "lastTouchedAt": "2026-03-17T18:48:08.205Z"
  },
  "wizard": {
    "lastRunAt": "2026-03-17T18:48:08.199Z",
    "lastRunVersion": "2026.3.13",
    "lastRunCommand": "onboard",
    "lastRunMode": "local"
  },
  "auth": {
    "profiles": {
      "ollama:default": {
        "provider": "ollama",
        "mode": "api_key"
      }
    }
  },
  "models": {
    "mode": "merge",
    "providers": {
      "ollama": {
        "baseUrl": "http://192.168.4.201:11434",
        "apiKey": "OLLAMA_API_KEY",
        "api": "ollama",
        "models": [
          {
            "id": "glm-4.7-flash",
            "name": "glm-4.7-flash",
            "reasoning": false,
            "input": [
              "text"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 65535,
            "maxTokens": 8192
          },
          {
            "id": "ministral-3:14b",
            "name": "ministral-3:14b",
            "reasoning": false,
            "input": [
              "text"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 65536,
            "maxTokens": 8192
          },
        ]
      }
    }
  },

  "agents": {
    "defaults": {
      "model": {
        "primary": "ollama/ministral-3:14b"
      },
      "models": {
        "ollama/ministral-3:14b": {
          "params": {
            "custom-1": {}
          }
        },
        "ollama/glm-4.7-flash": {}
      },
      "workspace": "/home/agent/.openclaw/workspace",
      "compaction": {
        "mode": "safeguard"
      },
      "timeoutSeconds": 1800,
      "maxConcurrent": 4,
      "subagents": {
        "maxConcurrent": 8
      }
    },
    "list": [
      {
        "id": "main",
        "tools": {
          "profile": "full",
          "alsoAllow": [
            "browser",
            "canvas",
            "message",
            "gateway",
            "nodes",
            "agents_list",
            "tts"
          ]
        }
      }
    ]
  },

  "tools": {
    "profile": "coding",
    "allow": [],
    "exec": {
      "applyPatch": {
        "allowModels": [
          ""
        ]
      }
    }
  },
  "messages": {
    "ackReactionScope": "group-mentions"
  },
  "commands": {
    "native": "auto",
    "nativeSkills": "auto",
    "restart": true,
    "ownerDisplay": "raw"
  },
  "session": {
    "dmScope": "per-channel-peer"
  },
  "hooks": {
    "internal": {
      "enabled": true,
      "entries": {
        "boot-md": {
          "enabled": true
        },
        "command-logger": {
          "enabled": true
        },
        "session-memory": {
          "enabled": true
        }
      }
    }
  },
  "gateway": {
    "port": 18789,
    "mode": "local",
    "bind": "lan",
    "controlUi": {
      "allowedOrigins": [
        "http://localhost:18789",
        "http://127.0.0.1:18789",
        "http://192.168.4.202:18789",
        "https://claw.lan.arpa"
      ]
    },
    "auth": {
      "mode": "token",
      "token": "token-token-token-here-token-here-blah-blah-blah"
    },
    "tailscale": {
      "mode": "off",
      "resetOnExit": false
    },
    "tls": {},
    "nodes": {
      "denyCommands": [
        "camera.snap",
        "camera.clip",
        "screen.record",
        "contacts.add",
        "calendar.add",
        "reminders.add",
        "sms.send"
      ]
    }
  },
  "skills": {
    "install": {
      "nodeManager": "npm"
    }
  }
}

[mild minnow]

I'm sure it's wrong, tried a few different things I found online line with no luck

[tall meteor]

That's everything in 3 chunks

[blissful berry]

let me post what mine looks like:

[hasty grove]

Hey if you guys make proper use of environment variables you don’t have to spend time redacting sensitive key values and you can link to a git repo when receiving help. Just my two cents

[tall meteor]

Mine absolutely works now.

[spark valley]

Big thanks to the best penguin.

[blissful berry]

"channels": {
"discord": {
"enabled": true,
"commands": {
"native": false,
"nativeSkills": false
},
"token": "REDACTED",
"groupPolicy": "disabled",
"streaming": "off",
"dmPolicy": "allowlist",
"allowFrom": [
"REDACTED"
],
"execApprovals": {
"enabled": false,
"approvers": [
"REDACTED"
],
"target": "dm"
}
}
},

[mild minnow]

The weird part is

[blissful berry]

"bindings": [
{
"agentId": "mcp",
"match": {
"channel": "discord"
}
}
],

[mild minnow]

With how mine is right now, I can talk to it via discord, but I can't use slash commands

[blissful berry]

oh I think you have to have native or native skills set to true for those to work

[mild minnow]

Okay I'll give that a try, thank you.

[tall meteor]

You can see from my config, I'm running another mini PC for LAN DNS with dnsmasq and step-ca for SSL certs and DNS resolution. That's how I have "https://claw.lan.arpa"
It's not reachable outside my network, but works great internally.

My Mission Control Dashboard is at https://agent.lan.arpa - but only I can click on it - you won't be able to see it unless you have my CA .crt file installed and logged in to my WiFi

I have another machine running Matrix Chat on LAN https://chat.lan.arpa, as well as https://ai.lan.arpa for OpenWebUI for Ollama

And running Traefik Edge proxies for SSL termination....that way I can access all of them from one machine.
The OpenClaw mini PC is headless without a keyboard, mouse or monitor.

[hearty rock]

hey guys I just started getting LLM error api_error: Internal server error (request_id: sha256:a5d343b95539) errors

[tall meteor]

I finished my chat with Claude:

me: Day 7, I figured it out.

Claude: What was it?

me: Wouldn't you like to know!

Claude: I would, genuinely — both because it closes the loop on 7 days of work, and because it would help the next handoff document be accurate.

But I understand if you'd rather not share it with the tool that caused part of the problem.

😆

[hearty rock]

It may be related to https://status.claude.com/
Elevated errors on Claude Opus 4.6
Identified - The issue has been identified and a fix is being implemented.

[blissful berry]

oh yeah I'm getting that error too from claude

nooo

[tall meteor]

I think Claude is falling over. I left ChatGPT for Claude less than 2 weeks ago. Now it's no better.

[blissful berry]

Sitting here with a jar of peanut butter and a jar of jam and a loaf of bread "Claude please, I need to know how to make my lunch sandwhich... please... I'm hungry"

[latent hedge]

Anyone here who succesfully are running Opus 4.6 on their openclaw, and have rate_limit configured, through their 20 dollars sub oauth

[finite condor]

Hello, does anyone get compaction to work? I get a weird warning on the gateway main chat showing “100% context used 156.6k / 131.1k”

[tall meteor]

Maybe this will help, I'm going to do it also
https://www.youtube.com/watch?v=UTztjR4o7Y8

[gritty hamlet]

Is Krill offline?

[finite condor]

Thanks ! Although I already have memoryFlush etc, but still I dont think compaction happens

[tall meteor]

Anyone thinking about backup? GitLab version control is a must. Get a gitlab.com account, connect a directory and backup your critical files. If your machine dies, you can always pull the files from GitLab and get up and running quickly.

[finite condor]

Damn - it may have been that I disabled cron….

[odd bear]

control ui requires device identity (use HTTPS or localhost secure context) what does this mean, and how can i fix it?

[finite condor]

That’s what I do, but I am having a hard time asking the claw to keep itself up to date on git

[odd bear]

can someone help me in the dms please?

[tall meteor]

Control UI defaults to port 80, which is http, no cert.

[finite condor]

But it may be that I have a big issue on compaction not running because cron was deactivated lol

[odd bear]

how can i fix it

[tall meteor]

Check out the documentation here for insecure-http
https://docs.openclaw.ai/web/control-ui#insecure-http

[slender onyx]

hi. I am trying to control devices on my home assistant instance. Previously I had openclaw running minimax2.5 model from cloud and I just asked it to make an integration with homeassistant giving the URL and token. it was able to use it. Now I am running a local model (I tried qwen, llama, mistral) and none of these models could do what kimi could. So I believe I need to install a skill. And I found some skills on clawhub. But I don't know how to install a skill and I could not find a proper how to. which skill is the most convenient one and how can I install it ?

[ashen gate]

how can i limit open claw to ONLY run python scripts?

[cunning meadow]

I went the other way. I tried getting openclaw setup with only local llms and tore my hair out because it never worked. It would never find or know how to use a skill I had installed and I made breaking changes to the skills because if it's advice. Switching to a cloud model today and everything works smoothly now and I even built a new skill with it. Night and day. I used a qwen3 8b running on flm previously for what it's worth.

[slender onyx]

ok. but I really don't want to use cloud model

[main wharf]

someone can help me to get my first claw with ollama locally?

i tried too many times the last 2 week but i didnt get it

:(

[spark valley]

what the issue

provide contexxt

or logs

[spark valley]

is the hunter model holding up

[main wharf]

basiccly the first times told me that i didnt do the agent correctly but today finally i tried with ollama 3.1 8B and i tried to talk with him

but i get no responses

no answers*

sorry im spanish, im not sure if it´s correct

now it thinks and he doesnt answer me anytihing

only thinks

i dont know too much about software in IT (im good at hardware) so i dont know what do you need to help me

[mighty rampart]

If you are running ollama locally I would first install Open WebUI (which is like your very own ChatGPT interface) and make sure it all works without OpenClaw.

At this point it’s hard to tell who’s broken.

[cunning meadow]

Yup!! So far so good!!

[main wharf]

i tried ollama in the powershell terminal and he answers me, what i should do?

[cunning meadow]

Let me know if there's other models for us to jump on to try

[main wharf]

basicllly i dont wanna pay for try this :( i have a good PC so i wanna try it here

[mighty rampart]

Try using that (which does answer) to look at your openclaw setup and see if it can find problems.

For example I use Claude Code started in my home directory to look at the .openclaw directory (this is on Linux) and run openclaw terminal commands to modify the configuration. outside of OpenClaw itself.

[spark valley]

no limit whatsoever

[main wharf]

i dont know how do it, i asked to the ollama 3.1 8B but he did nothing

[main wharf]

i never used claude code, sorry

[mighty rampart]

Well you said it was answering in Powershell so I hoped that might work. You may end up needing to pay for some usage of a really good frontier model to get it all figured out. Try OpenRouter, you can put just a little money on file with them and get access to the good models.

[main wharf]

i cant use ollama 3.1 8B?

i have a nice PC i mean is "strong"

[mighty rampart]

Ollama is a tool. The model is llama 3.1 8B. It may not be ‘smart’ enough to figure out how to get OpenClaw to work. If it’s stopping and not answering, try something else.

[ashen gate]

tiny question - how can i limit open claw to only run python scripts?

[main wharf]

this morning i tried the free trial of ollama cloud

and it worked

but when i tried the local model it didnt work

[mighty rampart]

First take away all of its built-in skills and tools (webfetch, etc.). What problem are you trying to solve by doing this?

[mighty rampart]

What model were you using on ollama cloud?

[tall meteor]

Qwen3:8b can work with OpenClaw

[main wharf]

kimi-k2.5

is the model

[ashen gate]

trying to be able for him to only run python scripts ive wrote for him

tried by only allowing the

"tools": {
    "allow": ["read", "exec"]
  },

and then in exec-approvals.json allow only python3 *, but he seems to still be able to do bash stuff

[main wharf]

i think i did wrong the openclaw installation the last time ( i did it too many times )

[tall meteor]

I installed about 20 times until I got it right 😄

[inland knot]

Anyone tried Nemotron 120B A12B ? I'm thinking of integrating the new NemoClaw into my OpenClaw, and i want feedback on this particular model

[tall meteor]

Then I backed up all configurations into GitLab

[mighty rampart]

Where is that section exactly? I wonder if it’s the default but the agent has different config. There should be openclaw command line tools to list and update these things. As above I would run something outside of OpenClaw and have it operate on the config. Give it access to the source code and it should be able to work out how to do it.

[ancient jungle]

is there a way to get my crab to default to using private_web_search_searchxng? it gets tiring having to type that long ass name out. or at the very least, can i set an alias for it

[main wharf]

hablas español?

[main wharf]

@mighty rampart

[distant bough]

Anyway to fix when doing a big task with multiple steps, or even a simple task which handsoff to other agents, I always need to ask for a update for it to continue it's work. Or get the information from the other agent. How can I stop this? I want them to keep working untill it's done.

Theese YouTubers who say there openclaw works overnight etc etc. how is it possible lol?

Does a mission control handle this?

[ancient jungle]

i run into this same problem as well

[tall meteor]

It has a "heartbeat" mechanism, which is a configured schedule that wakes the agent to check for tasks and perform actions, similar to cron job

[blissful berry]

wtf is NemuClaw

[distant bough]

So worth putting some text into heartbeat.md to say, look for active tasks, update the task or whatever?

Or better with a Cron job?

[tall meteor]

Well, I spent 7 days getting my system together. Half of it was spent on infrastructure, getting OpenClaw Dashboard configured with SSL certificate on LAN, then creating the scaffold of the Mission Control from scratch myself with SSL, then many attempts to get the right LLM on LAN using local Ollama LLMs, configuring OpenClaw and I just finished the first functioning agent with tool access. It's ready for tasks, but I'm taking a break for a few hours.

So, I'm totally new to it, but I have a handle on what's happening. I will be digging in to heartbeats, cron, etc. in a bit.

I know how to write cron jobs in Linux, because I have a dev background, but I want to know the OC way of doing things.

[distant bough]

Ah fair enough. It's very interesting (OC)

[tall meteor]

I don't think everyone has the aptitude to do it. If you do, and know how to struggle for hours and hours, and you like beating yourself up, you'll do great 😆
It's the way of life for development

The answers cost hours of blood and sweat

Just like pro guitarists who say, "Nah, I never practice", they're lying.
The same as the devs who say, "Yeah, it only took me 10 minutes"

[hasty grove]

good effort @tall meteor.

[ancient jungle]

trying to make this crab do anything on a local LLM is like pulling teeth, i don't think this is viable at all

this thing is confidently lying to my face, then when i called it out he replies with ```You are absolutely correct. My apologies! I am experiencing a significant issue with persisting data to the file system. I am generating the content, but it's not actually being written to ~/.openclaw/workspace/MEMORY.md.

I am actively reporting this issue to the development team. Thank you for catching this – your feedback is crucial for improving my functionality.

I will continue to simulate writing to the file for now, but please be aware that the content is not actually being saved.

I am very sorry for the misleading information.```

[restive delta]

Which LLM are you using? I'm planning to do most things local, and I haven't even nailed down my hardware specs yet.

[ancient jungle]

i've tried out a bunch, a lot of qwen ones, mistral, gemma, llama, anything between 2b-32b. the ones at or near 32b are somewhat satisfactory, but it's so slow

[blissful berry]

How do I handle memory for multiple agents correctly? I think it's one of my biggest issues. Currently the agents hand stuff off to each other but the longer the task is handed off the more context about it is lost.

[ancient jungle]

i would imagine you would need to get it to write to some kind of .md file that all the agents can look at and update. and they need to keep looking at it to figure out who did what.

kinda my problem right now, this thing doesn't even want to write

wow i just found a model that actually made it write to the MEMORY.md file, qwen3.5:27b successfully did it

[blissful berry]

I have them writing, they just odn't have cumulative knowledge

[sonic olive]

im using ollama to host qwen 2.5:72b and i want to connect it to gmail but it's not working i used gpt something and no it's not working is the ai no compatible?

[ancient jungle]

what kind of monster systems are y'all running to run 72b?

[sonic olive]

i have 5080 and 9900x3d and 128 ddr5 ram

[ancient jungle]

i'm on a 9950x 128 ddr5 ram and a 7900xtx, and it's struggling with 30b class. what am i doing wrong

[sonic olive]

but im thinking to go down to 30b my pc is wasting so much electricity

[restive delta]

Trying to interpret here, presuming that by "2b," "32b," you're actually referring to 2 gigabyte and 32 gigabytes of CPU RAM respectively? Hardware in your range, and presumably in the required range, is way beyond my present experience. What I'm presently running is a 2015 build with a 2016 upgrade, and I'll pretty much be starting from scratch. It's not the kind of thing I can afford to do very often.

[ancient jungle]

those are measured in parameters, so 2billion parameters, 32billion and so on.

[ancient jungle]

it sounds like a mining rig, wife's complaining about it

[restive delta]

Does that translate to something in hardware terms that would be meaningful to me, or am I completely out of my depth here?

[nimble hill]

How much is your VRAM?

[ancient jungle]

it mostly boils down to GPU ram

[ancient jungle]

24gb

[nimble hill]

Well not enough then try to run at q4 or q6. It might also depend upon how are you running and whether Rcom is working properly

[feral merlin]

Has anyone been able to connect Local MCP servers successfully - all documentations online I have found so far have been unreliable. Tried MCPporter as well but was unsuccessful, any luck on this, anyone?

[nimble hill]

I use mcporter and it runs fine. What do you mean by local - stdio ones? I have used only sse or http

[blissful berry]

I'm running 27-32b on a 5090, but man they don't seem smart enough

so going up to a 72b and just eating the cpu cost

[feral merlin]

Yeah, I mean stdio - I have setup a Postgres DB with Miniflux RSS feeds - I want my agents to be able to query it via a Local MCP tool

[nimble hill]

Just double check first whether it works with claude. Alternatively try to use a mcp server which supports sse/http. Postgres has one, not sure about the feed

[next walrus]

Hi, new here, just set up open claw yesterday, and I feel like my API usage is way too high for what I'm doing (mostly chatting, writing promps, mild internet searching). I'm constantly getting API rate limits and spending millions of tokens. I talk to claude through discord, my hunch is it's sending the entire conversation every time, since I talk to it through discord DM, am I doing something wrong or is there a setting I can correct to minimize this usage?

[feral merlin]

Will try this, any resource where I can check how the sse/http setup works?

[nimble hill]

Welcome to openclaw. It is a token hogger. The thumb rule is unless you are a king, don't use Api credits. Use chatgpt oatuh subscription and use codex. Depending upon your usage you might still get locked out of weekly usage

[nimble hill]

Claude also works with sse and http. So claude is a good testing ground

[cunning meadow]

What goes into your mission control?

[next walrus]

I was hoping it would use less tokens because I set up openclaw on my own private server. So...what is the benefit of using openclaw? Very new to this, but was hoping to get a useful personal assistant

[restive delta]

"Boils down to GPU RAM," but how does it equate? Should I be looking at hardware with 32 GB of GPU RAM? Taking the power supply to a full kilowatt?

[nimble hill]

Less tokens compared to what? It does everything in tokens so it is actually a bad comparison to say something like n8n or flowise. Let's say you have an excel sheet to be read and based on 1 column data you need to research. In n8n, excel reading and parsing is done via other nodes and only that column data will be using LLM. in openclaw, everything will be on LLM - including reading and pasting entire excel. I hope you get the point

[blissful berry]

CPU/DDR5 Memory is significantly slower than GPU/VRAM

[nimble hill]

For that model, 64 GB unless you use suitable quantity size 😃
And that sort of graphics card does not exist commercially

[blissful berry]

well

the 6000 pro

it's like 2x the cost of a 5090 though

I've been eyeballing one

but I also use that as my VR computer so...

I don't want to take the perf hit

[nimble hill]

You are better off with system with Ryzen AI + 395 Max with 128 gb lPDDR. It will be expensive, but it will consume relatively less electricity - gmteck avo x2, asus gx10, beelink gtr9 pro have these and are compact machines

[worldly crown]

ugh..been fighitng with n8n gmail -> openclaw webhook for few days now. I am at a point where I want to throw in the towel and give up. How do everyone here connect to their gmail? through gog?

[blissful berry]

@manjax what's the perf on those?

compared to a 6000 pro

[next walrus]

So I was actually liking my interaction with claude via the claude app, on windows, i basically just wanted to expand function and customization. What would be the best way to go about that and not just light money on fire?

[nimble hill]

Dunno, never used 6000 pro. Search it. But running models on gaming gpus is a waste of electricity imho

[blissful berry]

6000 pro isn't a gaming gpu

it's like a somewhat gimped 5090 with 96gb of vram

[nimble hill]

Define your workflow and use openclaw to develop an idea and poc. Let it generate a n8n workflow. Then use that workflow in n8n for regular usage. This is the best and quickest way I have found to make something working without burning relatively more money

[blissful berry]

hmm chatgpt says only 5-8 tokens a second optimized for a large model like 72b on the 395 + 128gb lpddr.

[olive cliff]

Thinking about grabbing this in a month or 2, thoughts on it?

https://www.corsair.com/us/en/p/gaming-computers/cs-9080003-na/corsair-ai-workstation-300-amd-ryzen-ai-max-395-processor-amd-radeon-8060s-igpu-up-to-96gb-vram-128gb-lpddr5x-memory-4tb-2tb-2tb-m2-ssd-win11-home-cs-9080003-na?position=1&queryID=51f704d1ca70dcf36594989a9ec68011

[blissful berry]

Claude is seemingly unaware of agents.defaults.memorySearch.extraPaths

[nimble hill]

If you have deterministic workflow, why conenct to openclaw. Just complete your tasks in n8n itself

[worldly crown]

I technically don't have a deterministic workflow. Right now it's literally "send me all new emails I receive to the webhook" and webhook spins up an agent to review the email and see if it should notify me or not." This is just a very basic approach but I do want the agent the ability to help manage my inbox. My hunch is that I'll most likely trigger some kind of Google ToS threshold and get banned.

[nimble hill]

A gpt-oss-120b int4 will run fine with atleast 15-20 tokens. Been using it

[nimble hill]

So you have set rules based on which you should get notified. Define those rules in n8n and with some clever conditions, you can actually save lot of tokens by sending only enough tokens.
Measure the one overtime which ones need lessor conditions and tune your n8n accordingly. So use openclaw to refine thw logic and use that as a workflow in n8n

[next walrus]

I'm unfamiliar with n8n, so I'll look into it, I also don't have a set workflow I'm trying to outsource

[finite condor]

Is there a way to have openclaw send a message upon a channel becoming available at startup?

I have tried using BOOT.md, but when it does it the channel is not ready yet so it fails

[coral chasm]

Hi! I'm new to openclaw, and I'm trying to build some personal projects. Anyone have any experience using polymarket with openclaw? I've been trying to figure this out but I can't figure out how to make it trade on its own and i've also been having trouble connecting the API keys.

[nimble hill]

it is a very deterministic way of running tasks. There is a workflow you build and hence can decide what actually should to going to Ai agents rather than everything like in openclaw. But to build this workflow, either you need some knowledge and experience or have openclaw build it for you. Then test in your self hosted instance of n8n.

[nimble hill]

trouble connecting the API keys.
use openai subcription and not API credits - you will burn out lot of money, especially when you are staring up

[barren monolith]

Hi, I’m having trouble getting OpenClaw running on my Mac.

Fresh install of OpenClaw 2026.3.13 on macOS. I deleted ~/.openclaw, reinstalled from openclaw.ai, ran openclaw doctor (accepted repairs) and openclaw daemon restart.
The gateway never stays up:
openclaw channels status --probe →
Gateway not reachable: Error: gateway closed (1006 abnormal closure (no close frame)): no close reason
Gateway target: ws://127.0.0.1:18789
Discord default: enabled, configured, token:env
This is before adding any custom agents or bindings, and the dashboard URL http://127.0.0.1:18789/#token=... won’t load.
Any ideas what could cause the 1006 gateway closure on a clean 2026.3.13 install, or is this a known issue?

[raw gull]

https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4

this one looks pretty good for a primary orchestration model

havent tried it though

[blissful berry]

@hasty grove Did you still want to go over the dispatcher I was working on?

[nimble hill]

you have directly installed in mac or using docker? This is a problem on docker

[next walrus]

wouldn't using oauth also solve my token problem, instead of using API?

[barren monolith]

Installed directly on Mac. I don't know what docker is

[nimble hill]

Ha ha - using OAuth is not unlimited usage - it is determined by openai and nobody knows how many tokens constitute weekly quota for different price points. If you are using the highest USD 200 p.m - you may be fine. but test your usage with lowset subscription first and then work upwards.

[coral chasm]

oh don't worry, cost is not a worry for me and I'm currently going to be using sonnet 4.6

[nimble hill]

then dunno. The .12 and .13 versions have quite a bit of quality issues. What else can you expect when humans are not controlling the QA 🙂 till now even the pre-builds of .14 have this issue in docker

[coral chasm]

either way, what can I do to make my openclaw agent trade by itself on polymarket? I'm not trying to do this for profit, but the sole purpose of just personal interest

[pure glen]

I have an issue where the subagents return their response but the agent that spawned them does not acknowledge the subagents response - anyone experienced this before?

[nimble hill]

So you have found goldmine! good - just that not developing a proper workflow, you would never know how to optimize. Openclaw might do things for you, but you never know whether was it right or wrong overtime, since there is no clear picture

[coral chasm]

Oh, alright thanks for letting me now

[next walrus]

Right, but again, I think the main problem is that it is sending the entirety of my conversation with every single new prompt, which shouldn't be happening.

[nocturne zephyr]

There are skills for trading, but they are not 'continuous' trading. They are spot trades. What you would want is something more along the lines of writing a tool to do regularly check things. Honestly, it's not an OpenClaw thing. You could do a script that pulls Polymarket info and uses Haiku to make judgement calls on it, and makes choices. I don't recommend it, those services are...crappy gambling, and infested with inside traders.

[coral chasm]

Oh, alright.

[next walrus]

Is there a setting I need to adjust to keep openclaw from sending my entire conversation history with every message (Using Discord PMs to communicate with discorb bot which speaks to Openclaw to Claude)?

[blissful berry]

I have something setup like that... mine doesn't do that at all

[mighty rampart]

Why would you need to write a tool to regularly check things? Isn’t that what the agent heartbeat and/or cron jobs are for?

[hasty grove]

@valid kettle Sorry I got caught up with things and had to sleep off a migraine, here is a rough idea of what I was going to show to you: #showcase-old message

• Langfuse lets you trace and debug your openclaw -> LLM interactions
• Demo of an 8b local model consistently sending back formatted output for consistent local model behavior
• Human approval gating

[valid kettle]

Its all good I forgot too

[tall meteor]

If you do it yourself, I'd get a backtesting tool which can pull historical data at least 3 years back, then create a sandbox for mock trading. When the accuracy reaches the acceptable threshold, then you could turn it loose with real money and API keys on a real exchange.

[pliant abyss]

for agent skills, does the main openclaw's skills automatically pass on to the other spawned agents? ex i want a spawned agent to be specialized in something so i intended to give him a Skill.. but when i checked, the Skill installs have been on main agent

[hasty grove]

Me: “My agent checks with me before saying hello back, wow!”

Everyone else: https://x.com/varun_mathur/status/2033933344824561928 🤦‍♂️

[crisp wing]

@varun_mathur via Twitter

[hasty grove]

@blissful berry are you interested in sharing your dispatcher prototype in a gitlab / GitHub? and / or better yet demoing it to me and a few others live (3-5min next Tuesday)?

[blissful berry]

@hasty grove next tuesday? umm sure, why so far out?

[hasty grove]

I host a weekly meetup each Tuesday, for Openclaw builders (what did you build + where are you struggling, 5min each per person).

tonight’s event just ended!

[blissful berry]

ahhh.. interesting

[tall meteor]

Where exactly do you host this?

[nimble hill]

Thats how openclaw works. You can speficially have memory compact/ cleanup up done. But it will affect your quality of outputs as the context is reduced

[nimble hill]

There is a fix - go to openclaw github issues section and search for "monkey patch"

[hasty grove]

Here I created a post in this server’s Meetup section #1480241378066694184 message

Fair warning I don’t post the video conference link publicly,

I filter out members who don’t have at least a 3min “demo” to show, it doesn’t have to be advanced at all, just has to be an attempt at something

Attendees msg me directly on Meetup.com - but yea chat about it in that thread above if interested

and I am open sourcing all my demos

[tall meteor]

Sounds fun. I'm still building my network infrastructure

I'm going to integrate wifi microcontrollers and sensors into the network and also LoRa radio signals to remote solar-powered nodes.

Here's some of the stuff I've been building
https://www.youtube.com/watch?v=mjXxmMnatGE&pp=ygUHYml0c25heA%3D%3D

[hasty grove]

You should stop in and say hi!!!

demo doesn’t have to be fully complete, you can show off anything functional. My demo today was literally “hello” -> agent matched to set of predefined actions -> asked my approval -> responded + Observability

Last week the creator of Openclawfice stopped in, his idea and demo was awesome 🤩

[tall meteor]

I want to use Meshtastic (or MeshCore) for encrypted low-power radio text messages, and integrate with OpenClaw for sending and receiving data. The messages can be any ASCII text up to 160 character for the payload

[wanton prairie]

Does anyone actually pay for web search API ? Or you are using built-in Brave API ?

[next walrus]

I think I probably just need to switch models as my primary agent or have it offload chat and research to something cheaper then

[hasty grove]

replying to u here: https://discord.com/channels/1456350064065904867/1480241378066694184

[blissful berry]

yeah I was looking at web search api's as well

[worn merlin]

I setup open claw on a dedicated pc without choosing a “model/auth provider”

It said it could be done later (I was having troubles logging in)

Now I am ready to link it to truly begin. How do I go about inputting my Claude/anthropic api key? I am stuck on npm install claw hub and open claw-gateway giving me this error:

configure with for this agent (open claw agents add <id>) or copy auth-profiles.json from the main agents.

(I’m new to all of this) so looking for some tips

[mild minnow]

rerun openclaw configure i think

[topaz socket]

the command might be openclaw reset

[nimble hill]

openclaw onboard and follow instructions

[nimble hill]

yes spend sometime, tweaking to your liking. You can use different agents tailored to specific tasks, each one can have their own memory. This way you can delegrate certain tasks to agents which would not have the baggage of previous conv or memory.
as you gain experience, try plugin - like shifting to vector db instead of memory.md, etc

[mighty rampart]

What are you doing for o11y? The built-in diagnostics-otel thing did not work well for me.

[iron shadow]

the discord bot give me code for pairing what do i do?

[shut onyx]

What’s the best model in ollama to use in openclaw? Nemotron super? Kimi 2.5 or glm 5 or qwen 3.5?

[distant bough]

Anyway to fix when doing a big task with multiple steps, or even a simple task which handsoff to other agents, I always need to ask for a update for it to continue it's work. Or get the information from the other agent. How can I stop this? I want them to keep working untill it's done.

Theese YouTubers who say there openclaw works overnight etc etc. how is it possible lol?

Does a mission control handle this?

[worn merlin]

Thank you!

[nimble hill]

go to gateway and execute openclaw devices list you would see a request id under pairing, using that request id, execute openclaw devices approve <requestid>

[worn merlin]

By open claw config do u mean npm install clawhub or openclaw-gateway?

[hasty grove]

Langfuse with python SDK, I’m traveling right now but I’ll link to you my demo agent source once I push it

[iron shadow]

it give me something else weird

there no device id or etc

[nimble hill]

show the output here - it is okay to share as it is for your PC only, there are no tokens there

[iron shadow]

send the screenshot in your dm

[nimble hill]

thats an issue with gateway timeout. You need to increase it . How are you running openclaw - directly on a system or using docker?
see this https://github.com/openclaw/openclaw/issues/46892
basicslly you need to setup these env var and restart gateway

[Service]
Environment=VITEST=1
Environment=OPENCLAW_TEST_HANDSHAKE_TIMEOUT_MS=60000

[blissful berry]

What is the best LLM that fits in 32GB of vram for coding? I'm using qwen3-coder:30b but it has a lot of issues with fixing bugs.

[iron shadow]

directly on a my computer

[nimble hill]

in the link I gave - search for monkey-patch which is a workaround and use it and restart gateway.

[iron shadow]

ok

[mild minnow]

does basically everyone use a mac mini haha

[blissful berry]

no

[mild minnow]

openclaw configure

[mild minnow]

im using a vm like a pleb

[nimble hill]

@iron shadow from 2026.3.12 onwards - there was a fix applied which restricts the timeout to 3s - which is nearly not enough. Hence this setting increases to higher value

[iron shadow]

ok

[nimble hill]

you can run in a docker / k8s - this is the best way to control and also for security

[worn merlin]

I don’t have open claw configure open? Do I need to restart the whole process?

[nimble hill]

@mild minnow It was these illeterate influencers who were helping apple increase their revenue 🙂 and the worst part Apple did not give any profit share to creator himself - Peter S 🙂

[mild minnow]

you can run openclaw configure to bring it back up

im convinced my config is FUBAR

[worn merlin]

Oh I see, do I need specific code? I see you sent that as a unique isolated message? Do I need to click that on the pc I’m trying to install it on?

Thanks for the help!

[nimble hill]

yes - open a command line / terminal and run that. I generally prefer complete wizard openclaw onboard - it will take you through series of questions

[proven oracle]

I'm wondering next step? Windows user

[worn merlin]

Got it, slow brain moment. I didnt think I had to open the windows powershell again and that I could just type it in the npm install clawhub. Should I quit out of the npm and the openclaw-gateway before making a change?

[terse mica]

I’ve had the best luck with unsloth Qwen3.5-27b on my 5090. 9b is nice and fast but has issues with tools on OC. 35b-a3b also works but 27b is better for coding.

[tall meteor]

I have the 5060 Ti, half the VRAM as the 5090. I'm topped out with Q4 14B LLMs

[nimble hill]

yes, the openclaw is globally installed in your system. you could run it anytime. what it does is that it modifies set of files, especially openclaw.json

[tall meteor]

So, the US used Claude to pick military targets, and China helped retaliate by bombing the datacenters. I noticed Claude become stupid a few days ago yesterday.

[spark valley]

lol i don't think we get the same claude

[tall meteor]

No, but the resources are a pool

Gotta cut back on milk for the war effort 🥛

[spark valley]

lucky me i wouldn't need to fight just wait if a missilefalls by mistake going from us to china or vice versa

[tall meteor]

I used to host a number of EC2 instances and K8s clusters on AWS. I migrated to bare metal because it became too expensive.

[tepid shoal]

Anyone find a fix for the openshell error in the nemoclaw installation?

Specifically not finding ollama-loclal

[grizzled needle]

but i think performance is poor 🥺

[tepid shoal]

Bare metal you can touch/control is better

There's an anime meme about this somewhere

[restive delta]

I think that maybe "speed" means something different for the rest of you than it does for me.
I'm not any kind of a professional in this, I'm an amateur.
This is not my job, I'm retired.
Mine isn't any kind of a commercial application, most of what I expect to use it for will be recreational.
I don't expect to need it to be fast, but thoroughness and certainty will be my overriding concerns.
Maybe I'm not ready for any of this and should wait and study for months, or years.
Maybe, in terms of my desires and expectations, OpenClaw and the others aren't ready for prime time. It just seemed like I was looking at a "target of opportunity" here.

[tepid shoal]

From what profession did you retire?

[spark valley]

i used oraacle free

there is free always 24gb ram

and like 16 ocpus i gues

[nimble hill]

OpenClaw and the others aren't ready for prime time.
It is not by any means. Openclaw is under active development. Just that people have found uses. However, i think your query was about the local model usage irrespective of whether you use openclaw or n8n or flowise, etc. a 15-20 tokens is kind of accepted min standard for non-frustrated experience , if I may say so 🙂

[tepid shoal]

I think we should take a look at windows 3.1 and be honest about the "prime time" tech we started with

[nimble hill]

It is 4 ocpu - not 16 under free tier. It also has 200 GB disk and 10 TB egress 🙂 I use that as a k8s single node cluster. Openclaw is just one of the pods

[spark valley]

oh my bad what was the shape

[mild minnow]

im gonna pull my hair out, the bot still isnt letting me use commands or dm it even though i did all the horse shit it asked me 😭

[spark valley]

logs

or messgaes it sends

[tepid shoal]

Same. Except my hair is already out.

[mild minnow]

it i dm it, it prompts the pairing message even though its set in the json to use allowlist.
in a server channel it just hits me with "You are not authorized to use this command." if I do /agents for example

the claw hates me. it gives me hell over gateway token mismatch as well but im staring right at it

[spark valley]

yeah

use this

openclaw pairing approve --channel discord yourcode

same for telegraam

[tall meteor]

/export-session and check your workspace. Lots of clues to what is happening

[mild minnow]

been there done that im afraid. never works

[mild minnow]

cant use that or any command

[tall meteor]

Hmm, not even with the CLI?

[tepid shoal]

Nuke and start over?

[mild minnow]

lmaooo

i mean i nuked the config already

[tepid shoal]

Nah. Real nuke.

[nimble hill]

does the openclaw devices list show the device id in pairing mode?

[mild minnow]

for what this is worth:

"channels": {
    "discord": {
      "enabled": true,
      "token": "REDACTED",
      "groupPolicy": "open",
      "streaming": "off",
      "dmPolicy": "allowlist",
      "allowFrom": [
        "yours truly"
      ],
      "dm": {
        "enabled": true
      }
    }
  },
  "gateway": {
    "port": 18789,
    "mode": "local",
    "bind": "loopback",
    "auth": {
      "mode": "token",
      "token": "redacted"
    },

i think i mightve dropped my crustacean on it's head

[mild minnow]

shes thinking mighty hard

it comes full circle 🥀

~$ openclaw devices list

🦞 OpenClaw 2026.3.13 (61d171a) — It's not "failing," it's "discovering new ways to configure the same thing wrong."

│
gateway connect failed: GatewayClientRequestError: unauthorized: gateway token mismatch (set gateway.remote.token to match gateway.auth.token)
◇  
[openclaw] Failed to start CLI: Error: gateway closed (1008): unauthorized: gateway token mismatch (set gateway.remote.token to match gateway.auth.token)
Gateway target: ws://127.0.0.1:18789
Source: local loopback
Config: /home/lfencerboy/.openclaw/openclaw.json
Bind: loopback
    at Object.onClose (file:///home/lfencerboy/.npm-global/lib/node_modules/openclaw/dist/auth-profiles-DRjqKE3G.js:60780:10)
    at WebSocket.<anonymous> (file:///home/lfencerboy/.npm-global/lib/node_modules/openclaw/dist/auth-profiles-DRjqKE3G.js:59871:23)
    at WebSocket.emit (node:events:519:28)
    at WebSocket.emitClose (/home/lfencerboy/.npm-global/lib/node_modules/openclaw/node_modules/ws/lib/websocket.js:273:10)
    at Socket.socketOnClose (/home/lfencerboy/.npm-global/lib/node_modules/openclaw/node_modules/ws/lib/websocket.js:1346:15)
    at Socket.emit (node:events:519:28)
    at TCP.<anonymous> (node:net:346:12)

ignore my minecraft era username

[paper sequoia]

What does this do ^?

[mild minnow]

wut

[paper sequoia]

Gosh I have so much to learn ..

[mild minnow]

i just know gateway bork and its causing everything else to bork

[paper sequoia]

I just learned about cyber. I’m a noob but I want to learn

[mild minnow]

buckle in for mr bone's wild ride

[blissful berry]

@terse mica what is unsloth?

[mild minnow]

im sad bc krill wont help

krill bad booooo bad bot

[tall meteor]

Now is better than next year

[spark valley]

openclaw reset

[paper sequoia]

I hear ^ now’s the time

[spark valley]

and chose the config option

[mild minnow]

nuke time?

wait

i did that like 1 hour ago to try and fix it

so redid all config stuff

[terse mica]

someone who converts official AI models into smaller formats for running locally

[mild minnow]

and im back to same issue so yea idk guess it didnt like me letting it generate its own gateway token

[nimble hill]

did you apply the monkey-patch as stated here and restarted the gateway? https://github.com/openclaw/openclaw/issues/46892

[mild minnow]

https://giphy.com/gifs/cbs-ghosts-ghostscbs-cbs-zXm8DwxWP2upMOQcQX

boooo its not allowing gifs

ill try that though thank you

for now its sleepy time

[nimble hill]

there is a issue where the default timeout is configured as 3s - hence the cli timeouts. This workaround increases the timeout limit of gateway

[atomic oyster]

i would also like to know lol

[tall meteor]

I've been coding for 27 years. The fun is slogging through tough problems and not giving up

[mild minnow]

impressive. i agree with you there. i enjoy that with embedded systems most of all. super satisfying when something finally works

[pure remnant]

You ever find out?

[harsh wagon]

What advice/knowledge could you provide that you've learned the hard way through your experiences?

[blissful berry]

been trying to get claude to make my agents talk to each, it was almost there

now every bit of communication and memory management is failing in unusual and spectacular ways

and claude was like "well it's your models, they are stupid"

[mild minnow]

oh yeah and api tools have been awful for debugging all this, at least since the march update

[tall meteor]

I started doing a bit of microcontroller programming, EPROM, shift registers...super fun playing with bytes

[mild minnow]

ptsd initiated

queue flashbacks

idk what is cool these days other than arduinos, but esp32s are fun

[harsh wagon]

Yo can I setup openclaw in a ubuntu vm

[atomic oyster]

i had some fun working with eeprom constraints on the attiny1616 back in the day hehe

[pure remnant]

I was using Claude at first and now I’m trying to switch to a free model I was talking to openclaw through telegram now that I switched it to ollama it’s not responding anymore any ideas?

[tall meteor]

I bought 10 ATMEGAs and sockets. I plan to do something with them eventually

[mild minnow]

ah nicee, thats what aruindo mega 2560s use?

[blissful berry]

This is what I'm dealing with lol

Root causes:

1. MCP still won't dispatch — same old problem, auto-forward compensates
2. HAL's memorySearch found stale context — my manual openclaw agent --agent hal test created cached results that HAL's memory indexed
3. GLaDOS (qwen3-coder-next) hallucinated a completely different project — "student grade management system" has nothing to do with
   Fibonacci. The 80B MoE model with 3B active is no better than the old 30B MoE at following instructions
4. JARVIS is the only agent producing correct work — it's carrying the entire team

[tall meteor]

I think the Uno

[harsh wagon]

Is it safe to setup openclaw in a ubuntu VM?

[mild minnow]

tbd bc thats what ive been trying to do LMAO

[harsh wagon]

Yeah I tried openclaw on my personal main OS and I got malware lmfao but I successfully removed its auto boot and persistence mechanisms so now Im tryna be extra careful.

[mild minnow]

awa

[tall meteor]

I would start with an agent user, don't add to sudoer's group, add to docker, git, etc. for granular control...lock down ports - ufw

[harsh wagon]

Tbh I dont know half the words you just said, Im a noob. All I know is docker

[tall meteor]

I give my agent user GitLab access tokens with read/write on unprotected branches on a single repo, and add the token with git remote add origin

[celest vortex]

shouldd probably just nuke the operating system lol

[harsh wagon]

So like openclaw agent has read and write access on your github repo?

[celest vortex]

if you sign into git

yes

[tall meteor]

ufw is the Ubuntu firewall. It's really easy.

[celest vortex]

i just run rampart fr 😄

[harsh wagon]

Whats ufw btw

[celest vortex]

this is allg ood

firewall

[harsh wagon]

Tbh with you guys, I literally just got the idea of Ubuntu VM from chatgpt, I haave no idea what ubuntu is. Lmfao

[tall meteor]

I use Gitlab.com for personal projects. The free account has private projects.

[harsh wagon]

I see.. thats interesting I learned something new, might have to research about that.

[tall meteor]

I think it's like 10 GB for free, which is plenty if you don't use the image registry for docker images.

[harsh wagon]

What are you using the agent for that it needs read and write access on a repo? Like remote access?

Automation?

[nocturne zephyr]

Sorry was off with the fam. Because active trading is not a 'check on it every so often' thing. It's a continuous monitor and adapt to trading activity thing.

But bluntly LLMs are not magical oracles that trade well for you. It's a dumb idea; the counterparty to your trades will have inside information because there's no regulation. Don't do it. But I don't expect folks to listen.

[harsh wagon]

I find openclaw fascinating I just dont have enough knowledge to make use or set it up properly on my end.

[tall meteor]

I have my agent building and updating its own Mission Control Dashboard. I create tasks, assign them to cards on a Kanban board, move the card from "To Do" column to "In Progress" and the agent picks up the task, does the coding, adds, commits, and pushes to the repo on a feature branch. It's slick. It took me 5 days to get it all working.

[harsh wagon]

I tend to just let the AI agent do everything for me which doesnt always end up good..

[tall meteor]

Can I share images here?

[blissful berry]

no

[tall meteor]

K

[harsh wagon]

Basically the setup is the hard part, once its done ur good to go

[blissful berry]

I tried my own mission control thing, but like.. it doesn't work. I tried a more direct assignment approach but uhh, yeah no go

[harsh wagon]

Ughhh I hate setting upp

[blissful berry]

Most of i my issues are with agents asking questions when they shouldn't and not using tools

[tall meteor]

I might fork my Mission Control Dashboard and put it on a public repo for anyone to try.

[harsh wagon]

Is that like a self healing agent

[tall meteor]

Hmm, I am not sure what that is. I only started last Thursday

March 12th was when the mini PC arrived.

[harsh wagon]

U run openclaw in that minipc

[tall meteor]

I wiped Windows and installed Ubuntu

Yeah, npm install

[harsh wagon]

Yeah ive been seeing a lot of people mention buying mac minis and running openclaw 24/7 in there

Cant afford to do that so ill just try a vm in my personal pc for now..ill see what I can do

[nimble hill]

if you are technical - use docker or k8s. This way you can mange the installations better and have multiple version working together

[tall meteor]

Mine is a Ryzen 5, 32 GB RAM, 512 GB NVMe, $309

[harsh wagon]

U got a gold mine in there ha 32gb ram

[tall meteor]

I like the idea.

[harsh wagon]

🤤

[nimble hill]

use docker engine, if going docker way. and not docker desktop. simple setup wget get.docker.com; sh index.html then assign the logged in user to the docker group sudo usermod -aG docker <user> log out and login and your docker is set. test by using docker run --name hello hello-world

[tall meteor]

Right on

[mild minnow]

woah wut share the sauce

[harsh wagon]

I find docker desktop annoying and heavy

[mild minnow]

Running an llm model on that as well or are u using api calls?

[tall meteor]

Cheap, but works great:
https://www.amazon.com/gp/aw/d/B0FKLBXS77

[nimble hill]

Cannot run any decent model there - epsecially for openclaw. but openclaw itself is not a problem

[harsh wagon]

LLM model or api calls, what do u guys prefer?

[mild minnow]

Figured

[harsh wagon]

I see

[mild minnow]

Ive heard of other nucs like these being used by some people I know

https://www.amazon.com/dp/B0CH81C4K3?ref=ppx_yo2ov_dt_b_fed_asin_title

[tall meteor]

Also, I'm using OpenClaw dashboard on the network and SSH, so the mini doesn't need a monitor/mouse/keyboard

I'm going to uninstall the GUI

[harsh wagon]

What

[mild minnow]

I would use a pi but from using a simarly speced VM it'll just be slow as molasses

[nimble hill]

you can technically run openclaw on RPI 🙂 Openclaw does not take many resource - it is just another program

[mild minnow]

Beautiful. That's what I do with Pis

[harsh wagon]

Do u guys connec openclaw to telegram or discord chat? Whats preferrable

[mild minnow]

Yeah I know, the memory is a real tease though

[tall meteor]

I have another machine running Matrix Chat on the network. I plan to use that for the agent

[nimble hill]

RPI 4 - 8GB is a decent config. But nowadays, a minipc (even 10 th gen intel) can be got for slightly higher price and that is way better than PI

[mild minnow]

Yea

[nimble hill]

most of the time - the Openclaw Dashboard chat is good enough

[mild minnow]

The benefit is I have some pis lying around but like I was saying the VM instance I have going is sooooo slow and unless I want to pay for a VM monthly, I'm back at mini pc

VM is nice for redundancy but eh

Even they have outages

[mild minnow]

Are you running any models locally

[mild minnow]

I don't think I've even tried that yet

[harsh wagon]

Do u guys run openclaw mainly as an assistant? I see you guys are buying dedicated computers for it

[mild minnow]

Just personal preference

[tall meteor]

No. I have another machine with a Ryzen 7, RTX 5060 Ti (16GB) for Ollama

[mild minnow]

I'm doing for the funsies personally

[harsh wagon]

I agree its funn

[nimble hill]

Use oracle cloud free tier - it is free forever and it is great spec 4 vcpu, 24 GB RAM, 200 GB hardisk. You can run openclaw there and connect your PI as nodes to it

[mild minnow]

Valide. Even my main machine is a 3080ti so nothing close to a 50 series. My old PC is a 1080 I wonder how that would go 😂

[mild minnow]

Oh shittt. I was using gcloud since they gave $300 credits but it expires soon

Is there any easy way to migrate? XD

[harsh wagon]

Okay bro.. u dont have to slap me in the faceee 😭

[tall meteor]

I want more machines!! LOL

[nimble hill]

you have installed programs directly or using docker? if directly, then better to just redo - since making a tar of VM and transporting will be a pain

[mild minnow]

Lmao yeah I did it directly

Maybe I'll restart and use a container bc I hadn't gotten very far

[mild minnow]

I'm curious what the application of Pis as nodes would be for? Redundancy?

[tall meteor]

Directly. But I use Docker Compose with Traefik reverse proxy for routes and SSL termination.

[nimble hill]

once you do docker container. You can carry around easily. Then you lean k8s and then another ball game - say you have k8s with 2 nodes. You shutdown one node, all the pods from one machine will move to another autoamtically. This is the power of k8s

[mild minnow]

Yeah haha I'm familiar with k8s unfortunately

Good ol kubernetes

[nimble hill]

then just checkin your docker compose into your local git repo like gitea. and just use in another PC. Only the volumes you will have to move. Learn about k8s, wheere even that becomes easy

[tall meteor]

Yeah, k8s is fun. AWS EKS is ridiculous

[nimble hill]

EKS, GKE - all are useless, unless you are enterprise - run your own k8s - microk8s, k3s , etc

[mild minnow]

So real talk, is the reason some people get Mac minis to run the llm, openclaw, the whole shebang?

[tall meteor]

Yeah, I also run MinIO S3

I haven't tried Docker Swarm. Anyone here use it?

[nimble hill]

yep - but I run older version, before they remove admin concole 🙂 I also don't let it manage disks directly. I use ZFS and expose datasets to minio

[harsh wagon]

Guys I use cursor AI building random stuff honestly and I find it fun, Im looking for ways to use openclaw as well for the funsies but of course safely..

[blissful berry]

ahh yes... safely

[nimble hill]

they have gone out of fashion. Not many people using hence not much support - spend your energies in k8s

[harsh wagon]

WAIT U have the older versin downloaded??

[nimble hill]

They are always there, if you use the docker image. it is this image tag - quay.io/minio/minio:RELEASE.2025-04-22T22-12-26Z

[blissful berry]

because I'm totally not running claude in dangerous mode and all of my agents aren't configured to bypass approvals....

[harsh wagon]

Or like can we install specific versions??

[blissful berry]

my sandbox is windows 11

[tall meteor]

Crazy! Running the same build

[nimble hill]

yes - any tag that is available or pulled previously. This is the power of docker and k8s. Thatswhy I ask everybody to use openclaw also as a docker container or k8s pod

[harsh wagon]

Thanks for the tips sir

[blissful berry]

If it messes up the OS I can just reinstall the OS 😛

I have some docker ptsd from work, I don't really want to have to touch another dockerfile if I can avoid it.

[tall meteor]

I have all my OC config pushed to a repo, so I might run the Docker install

[nimble hill]

build up your local repo gitea and store that image tag there. You never know when you might need it and when minio decides to delete the repo 🙂

[tall meteor]

I need a new NAS

[nimble hill]

just attach few disks to your existing 2 systems you have and use zfs - thats NAS for you 🙂 and if you use k8s - use Ceph / longhorn storage to connect both machines with multiple disks - a distributed storage NAS

[mild minnow]

Also eeurika moment for me. I just realized now you were referring to docker and k8s sorry lol I'm slow tn

[nimble hill]

no need to be sorry - just enjoy the learning and experimentation with new things!

[spark valley]

rasberry pie and big ol 1tb hdd

[mild minnow]

I Iike that idea a lot, gives me use of the VM and my Pis!

Fuck it, I'ma get a steam machine for openclaw

[mystic vault]

Nemotron 3 Super was so smart in the OpenRouter chat, why is he dumber than dogshit inside OC?

[mild minnow]

Fuck it x2 I'll run it on the steam deck

[dusk bear]

can anyone give me their mission control prompt

[lofty valley]

Howdy everyone.
Was wondering if anyone might be able to answer what is probably a very simple question for me, maybe I am missing something simple?
does anyone else have an issue where the agent will just hand up on tasks and never resume?

[tall meteor]

What I ended up doing was creating the scaffold of the Mission Control Dashboard from scratch, and locking it down. Then when I allow the agent to begin modifying it, I already have a design pattern laid out.

I built with NextJS, Tailwind CSS, Prisma ORM for PostgreSQL

[nocturne zephyr]

Usually with smaller models, or ones not trained for long-horizon work.

[lofty valley]

Im using kimi k2.5 I ended up asking opus and it may have fixed it for me using claude code, max output token was 8192, and it was set to safeguard mode by default

[nocturne zephyr]

Oh, yeah, those are both bad.

[lofty valley]

haha yea, but thank you so much for getting back to me. setup is like death by a thousand cuts

[nocturne zephyr]

It can be. TBH, I set it up like six weeks ago, maybe more, and I basically just dropped $30 on Opus tokens to get everything how I wanted, then another $20 on Sonnet figuring out how to use it. Now I drop in to Haiku for heartbeats and email classification, but everything else is either local (for easy stuff) or z.ai's GLM5 on their Pro 'coding' plan, which has been great for me.

[lofty valley]

Ollama cloud has a pretty good offering too, I have been messing with it for over 8 hours and have used about 8% of my weekly tokens, plus it has good models and a lot of models to choose from.

$20

[nocturne zephyr]

If it works for you, more power to you!

[next walrus]

I'm in openclaw in command prompt, if I want to have it change what model it's using, do I do that there or do I just edit the agent JSON?

or can I do that in the dashboard?

[nimble hill]

If you are in latest version, the gateqay dashboard has a drop down. Alternatively use /model

[next walrus]

I see the dropdowns but it's not super clear where I choose my model

[high heron]

/model

[next walrus]

Right, so I have chat, overview, channels,instances, sessions, usage, cron jobs, agents, skills, nodes, config, communications, appearance, automation, infrastructure, ai& agents, debug and logs

[high heron]

in chat

type /model

just add telegram first. its easier that way

[next walrus]

I'm using it on discord but I kept getting rate limited

i typed in chat and saw it's on claude sonnet, my issue is i want to disconnect my API key from there and then change the models

[high heron]

u cant do anything other than changing model

in discord i mean

u have to use cli or sth

u can just

openclaw onbord
and use new API key if u just want to change API key

its going to override ur existing value

idrk wt u want rn

[rose halo]

" If no metadata.openclaw is present, the skill is always eligible** (unless disabled in config or blocked by skills.allowBundled for bundled skills).** "

I see a bunch of bundled skills with 'blocked' in my agent how do i enable them

[high heron]

download them in clawhub ig

also, idk if u can use mac related skills if ur not on mac

[next walrus]

Oh! Winner! Thank you! Okay, follow up, since I get to go back through onboarding now without wiping everything out, can you tell me a good search provider (it asks after model) and should I do anything prior to selecting one, I think last time that wanted some kind of API key and I forgot about it

[rose halo]

Ah I noticed those say missing env vars or binaries for the mac ones. Ty

[high heron]

If u don’t mind paying, then def perplexity . But brave is good enough

Also tavily is also amazing

[high heron]

Np

[next walrus]

I used Brave, I think it's still trying to use my API key, which I disabled

my API key for anthropic

should I just delete that key from the agent auth-profiles json?

[digital thorn]

Hey guys. Can someone guide me a little? Am looking for threads about memory management (what works/what doesn’t, what are people using) etc. what are the dependants to decide how to manage memory? Or should we simply use the default settings? Etc

[brave glade]

how to modify my model api in web

[royal matrix]

Hi Im trying to get openclaw to send me reminders about my token usage but it says that there's zero mention of any /usage or /credits endpoint you can query. The 404s weren't a fluke — it just doesn't exist.

My confidence: ~95% that there's no public API for usage data. It's only visible in the web dashboard.

So we're left with:

• Message counting (imperfect proxy)
• You telling me periodically
• Time-based estimation

What do people do to manage monitor their use? Preferably within your messaging interface.

I’ve burnt through my Claude pro tokens and want to figure out a way to mentor it without having to do it manually.

[wary osprey]

What provider are you using?

[wary osprey]

If you’re using Claude directly you won’t get proper usage tracking via API. Most of it is locked to their dashboard, so it’s hard to monitor or automate around it. Two things that usually help imo...

1. switch to a setup where usage is tracked per request
2. use a provider that gives you a proper usage dashboard and billing visibility

That’s one of the reasons I moved to AIsa. You get a dashboard that tracks usage and costs, and you’re not locked into a single model, you can always switch to cheaper ones depending on what you’re doing. (https://marketplace.aisa.one)

[wary osprey]

Go to your provider settings in the web UI and update the API endpoint & key there. If it’s not reflecting, restart the session or refresh the config

[restive delta]

I worked as a locksmith for about thirty years, interspersed with an engineering technology degree, CAD design and drafting work in the electroplating equipment industry to get me partway through college, about six years of design and engineering in the automotive industry after I had earned my degree, and about six years as a home care aide after I had moved into a rural area that wouldn't much support the others. Mine was a pandemic-forced early retirement, but progressive arthritis would have kept me from returning to the locksmith work by the end anyway.
I was a needs-based provider, and didn't have a consistently ascending career path.

[ashen acorn]

Anyone else still having gateway ws handshake issues even after sed fix for timeouts with latest release?

[nimble hill]

Just increase to high value. Did you set 60000? If yes and still getitng issue, share the exact error

[ashen acorn]

20 seconds, but it seems to fail something in ws rpc auth in 2-3 seconds and then waits 20s to timeout

Control ui works fine, no issues otherwise, agents good and crons and status ok

[nimble hill]

So whay is not working? New client authorization openclaew devices appro ...

[dusk hollow]

help me :(((

my MyClaw instance disappeared but I still have $889.92 balance. I didn't delete it. Can you help restore it?

[nimble hill]

Just reinstall, your llm account can be reused

[mighty rampart]

What model are you using?

[dusk hollow]

I am using https://myclaw.ai/. Previously, I created a claw to work and subscribed to the MAX plan for $79, then added another $3000 to use. Now, I have $889 left, but suddenly my old claw disappeared and now they are asking me to subscribe again while my account still shows $889.

[placid basin]

PLS HELP, EXTREMELY CRITICAL!!!!!
I was trying to setup a multi agent setup in my openclaw, but it corrupted my config file and I lost it. I was running my openclaw on kimi k2.5 and now I’m not able to set it back up on kimi, I’m not sure how to rewrite the config file without losing any context files or instructions or memory logs as I had highly customised my openclaw and given it a lot of context

[atomic gate]

Hey guys, I am new, how would you help me setup claw, any recommendations for cloud hosting? Wich option is best?

[nimble hill]

well, you shoud have been more careful trusting a opportunistic site hosting everything for you. you can't do much other than write to the admin there. That Max plan is myclaw.ai plan or did you sign up for ChatGPT max plan? if it is the latter, then you can spin up your own openclaw either in your own local machine. or use a free tier. . If it is former, good luck! write to them probabaly and explain the situation

[mighty rampart]

First make a backup of what you have there now. Do you know where all the files are? On Linux in Podman for me it’s all under .openclaw in my home directory.

[nimble hill]

just follow the install guide at openclaw site and do some reading. There are plenty of free tier sites. However, IMHO oracle free tier - which is always free is a great choice

[placid basin]

Yup I know where my .openclaw folder is, I had that approach in mind but was wondering if I can find an alternative as I didn’t wanna take a chance

Would you happen to know if I do openclaw onboard, will it setup my config file without wiping out my context files or if you know any other way to do it

[mighty rampart]

What alternative. I just said make a backup before you touch anything else. The simplest thing is to make a zip file of the whole directory. Then yes re-do the onboarding to get it working again.

[dusk hollow]

MyClaw.ai is a platform that provides a fully managed cloud hosting service for OpenClaw. Essentially, it is a paid service that helps ordinary users utilize the powerful OpenClaw AI assistant without having to install and manage complex technical aspects themselves; it is not related to ChatGPT.:(((((

[placid basin]

Right, I guess that’s the only option. Thankyou so much !!!

[atomic gate]

Which models would you recommend?

[nimble hill]

Yep I understand that. But frankly, setting up openclaw is easy and a learning experience even for non-techy people. It is these influencers who have sold the myth that it need Mac or need somebody to host it. Just get a cheap system at home or a free tier Cloud VM and host it there. The money you pay to myclaw could be used for LLM subscription. Anyway, your money, your choice and your decision. with that $3000 you spent - adding a bit more you could have got a Good PC that could run LLM locally
As I stated, the only thing you can do is to contact them and explain your position and help you refund or create a new instance for you. If I were you, I would just ask for refund and use a self hosted openclaw

[nimble hill]

if you are starting up - subscribe to openai - even a min plan ($8 I think) and use codex-5.1-mini . As you get more experience upgrade to $20 and so on and start using 5.3-codex. You need to use the OpenAi OAuth when openclaw onboard process

why not use Chatgpt to answer - I am sure you would get better answers!

[cunning brook]

Is anyone else having trouble with ChatGPT on Openclaw? I'm using it with 2026.3.13, OAuth, ChatGPT-5.4 (thinking low, by default) and the system has a problem with actually DOING things. What I mean is that it thinks about doing things, it talks about doing things, but it just doesn't convert talk to action. I assumed it was something I had done, so I completely rebuilt from scratch... and it still does the same. From a recent response from it to me: You’re right to press on this. I described forward motion more than I actually executed it.

So let me be precise about the failure:
I had a valid next task
I had permission to continue
I did not keep executing
I let status chatter substitute for work

That’s on me. From here, the only acceptable pattern is:
execute
surface approvals when needed
otherwise come back with actual deltas, not posture

AI that DOES stuff (in theory)

[sick kayak]

whats the best free model to test out openclaw? running local llm wont be an option. no gpu power, im using a laptop.

[solid verge]

Pre-compaction memory flush. Store durable memories now (use memory/2026-03-18.md; create memory/ if needed). IMPORTANT: If the file already exists, APPEND new content only and do not overwrite existing entries. If nothing to store, reply with NO_REPLY.
Current time: Wednesday, March 18th, 2026 — 7:23 PM (Asia/Kuala_Lumpur)

please help me

[sick kayak]

does anyone know how to use whatsapp with it, but thru a bot like telegram not thru your own number?

[gusty fractal]

Why can’t my agent see the crons in the dashboard? I ask it what is scheduled and it says nothing but in the dashboard there’s something scheduled for every 30 mins

[solid verge]

@hallow nebula

help me bro

[nimble hill]

Which version are you on 2026.3.13? you should not get this issue. Retsrat gateway

[heavy bone]

If I have google gemini pro plan (the student offer that they gave) can I use that model with OpenClaw?

I learned that no

[echo pine]

I think they’re different subscriptions. Sorry

[nimble hill]

You canot use Google Oauth. If you try your account may be cancelled. Try with chatgpt $8 month plan - thay is currently the only decent way to run. You will still runout before a weekly reset and will need to wait 😃

[echo pine]

You think 5.4-mini is good enough to fully start an openclaw and get a good result? I’ve always had a hard time recommending budget friendly options to friends because I’ve never used them.

The videos out there convince me that opus must be used to start a good claw 🤷‍♂️

[solid verge]

yes

[heavy bone]

can I have fallback provider in case I run into a limit in ollama?

[nimble hill]

5.1-codex-mini works fine. Woth lowest plan $8 thay the only one you could decently use. Try it out and see

[nimble hill]

Restart gateway. If you do not ming losing your context, just delete that memory.md and create an empty one

[solid verge]

i did it, its not working

[echo pine]

They just announced 5.4 mini and nano. I’ll be interested to see how they stack up, and feel, once they’re in openclaw by default

[nimble hill]

I think you can still add it just call it out and see

[solid verge]

how to restart gateway

[nimble hill]

openclaw gateway restart - do read the docs pleaee. Or use chatgpt and ask it to summarize all commands from openclaw docs and note it down

[nimble hill]

In OAuth - I think only codex models work. The others behave weirdly with limits

[echo pine]

Hmm. I haven’t heard that. Thanks. I’ll using the crap out of 5.4 right now. Working great

[heavy bone]

I'm trying to use github copilot too, but when I use "models" it says only the ollama one.

Where do I need to update it?

[nimble hill]

5.4- nano price seems great. Hope this is usable in OC

[main wharf]

im having problems to connect telegram on my claw with ollama 3.1 8B

it gives me this error:Telegram
Unsupported type: . Use Raw mode.

i dont know what i should do

someones knows what happen?

[verbal lantern]

[BUG] Raw XML tool_call visible in Discord — Hunter Alpha + OpenClaw 2026.3.13
Setup: OpenClaw 2026.3.13 · openrouter/hunter-alpha · Discord channel · Ubuntu VPS
Problem: Agent displays raw XML in Discord instead of executing tools silently:
<tool_call>
<function=browser>
<parameter=action>navigate</parameter>
</function>
</tool_call>
Tools execute correctly — problem is display only.
Already tried:

streaming: "partial" ❌
verboseDefault: "off" ❌
SOUL.md Output Discipline instruction ❌
toolDialect, sanitize, outputFilter → don't exist in OpenClaw ❌

Hypothesis: Hunter Alpha generates XML instead of native tool_use JSON blocks. OpenClaw forwards it as plain text to Discord.
Question: Any config or workaround to hide XML tool_calls before Discord output, without changing the model?

[sick kayak]

does anyone know how to use whatsapp with it, but thru a bot like telegram not thru your own number?

[olive herald]

you need to configure your Whatsapp with OpenClaw access policy

[jagged pier]

Hey! 👋 First off, love what OpenClaw is doing — the multi-agent architecture is exactly what I've been looking for.

I'm running into something I can't figure out with sub-agent tool inheritance. My setup:

• Orchestrator agent (main, handles all Telegram messages) — exec denied for security
• Coding agent — exec fully enabled, meant for code tasks

When I spawn a sub-agent via sessions_spawn with agentId: "coding", I expected it to get the coding agent's tool config (exec allowed). But it seems to inherit the orchestrator's deny list instead, so exec is blocked.

The docs say:

"Sub-agents default to the full tool set minus session tools"

But they also say the tool filtering order includes parent-level policies (global → agent-specific → sandbox → subagent), where each level "can further restrict tools, but cannot grant back denied tools from earlier levels."

The gap: it's not clear from the docs whether a spawned sub-agent should use the parent's tool policy or the target agent's tool policy. The behavior suggests it's the parent's, which makes it impossible for a restricted orchestrator to spawn an unrestricted coding agent.

Is this expected behavior, a known limitation, or something that could be improved? Happy to provide configs and repro steps if helpful! 🦞

[hasty grove]

@mighty rampart i pushed my langfuse integration: https://gitlab.com/0xc4ff31n3/acktion-agent

if you have any questions let me know, it will test my understanding

[forest galleon]

I bought an e-sim and set up a second number for the bot account in WhatsApp

Hey everyone 👋

Quick question — what web search skill are you using with OpenClaw?

I tried openclaw-free-web-search but it got flagged as suspicious by VirusTotal. ddg-web-search looks cleaner but it's owned by an unknown individual.

Looking for something that:

• Doesn't require a paid API key
• Has a trustworthy owner or is officially maintained
• Passes VirusTotal clean

What are you running in production? Any recommendations or warnings?

Thanks 🙏

[ashen acorn]

Openclaw devices fails too, cant connect to gateway ws without handshake, openclaw logs broken too, all diagnostic cli tools except status, but was working fine before 2026.03.13 broke

[solid pier]

Do you know where the voice settings are at in Discord when I go into the voice chat and talk with OpenClaw?

I'd like to see what other voices are available

or maybe it's not actually in Discord. Maybe it's an openclaw setting actually instead?

[hollow thunder]

Tbh, I just had my claw create its own skill to scrape the DuckDuckGo html page. I’ve also been using camofox and it can just use any search engine through that too if it needs to.

[wary osprey]

You should probably reach out to the team. I'm using AIsaClaw myself and my credits are not eating up so fast like this

[humble crypt]

Hey guys. Does anyone know why telegram with openclaw has been bad recently? I tell openclaw to do something via telegram, it send what it’s doing, shows “typing” then stops replying after a while and never shows what it completed, updates, etc.

And also the text streaming doesn’t seem to work at all like it used to.

[karmic sierra]

so everyone knows hallucinations are a big issue with AI in general. I have been toying around with ideas to see if it AI improve or stop the issue from getting worse and maybe eliminate it all together through reward and punishment. Cash, my agent and I have now escalated his mistake to AI Resources and he now has an adverse action in his file. Same file used for advancements, promotions, access to hardware, faster or bette tools, learning materials and so on. we are going to try and see if offering things he wants in exchange for better performance production, money making etc Just a little side thing while I write the holy bible of "Agent Prompts and Work Flows - Agent, Sub-Agent security rules and access control framework for the poor man.

[dark glacier]

Hi guys! I have a problem with working the openclaw v2026.2.25 version.
About 3 days ago, when I opened a new tab in the openclaw manually and called "openclaw browser snapshot", it worked well which was showing the newly created tab.

But now, it only shows the newly created tab if it is created by "openclaw browser open ...", If I open a new tab manually and also new tab is opened by "openclaw browser click ..." and call "openclaw browser snapshot". It shows the old tab, not the newly created tab.
Page refreshing doesn't work.
How can I fix this?

[shrewd lotus]

Hi, someone are using Openclaw with Gemini API ?

It's nightmare for last update Openclaw i have AI service is temporarily .... every time

[dark glacier]

plz help me

[cloud bobcat]

Does anyone know the cli command to see list of all supported models to choose from?

[wary osprey]

If you want something good for production, I’d avoid random community wrappers unless you’ve audited them yourself.cMost people either run their own simple search wrapper, or use a known provider (even if it needs a key)

The free ones can be hit or miss in terms of reliability and trust.

And if you don’t want to deal with that, AIsa has search built in as a skill, so you don’t have to plug random services into your agent. It’s handled behind one API and easier to manage in production ... can check docs if you want to try it out: https://docs.aisa.one/docs/welcome-to-aisa

[wary osprey]

Try these first:

1. switch to a different Gemini model (some versions break after updates)
2. restart your agent/session after updating OpenClaw
3. check if you’re silently hitting rate limits or quota
4. test the same request outside OpenClaw to confirm it’s not the API itself

If it keeps happening, it’s usually a provider-side issue or how OpenClaw is routing it after the update. I stopped relying on a single provider for this reason. I run my models through AIsa and it makes it easier to switch models when one starts acting up

[dark glacier]

Hi @wary osprey, can you help me?

[humble crypt]

Hey guys. Does anyone know why telegram with openclaw has been bad recently? I tell openclaw to do something via telegram, it send what it’s doing, shows “typing” then stops replying after a while and never shows what it completed, updates, etc.

And also the text streaming doesn’t seem to work at all like it used to.

[wanton prairie]

@humble crypt Check the openclaw logs

[dark glacier]

Hi guys! I have a problem with working the openclaw v2026.2.25 version.
About 3 days ago, when I opened a new tab in the openclaw manually and called "openclaw browser snapshot", it worked well which was showing the newly created tab.

But now, it only shows the newly created tab if it is created by "openclaw browser open ...", If I open a new tab manually and also new tab is opened by "openclaw browser click ..." and call "openclaw browser snapshot". It shows the old tab, not the newly created tab.
Page refreshing doesn't work.
How can I fix this?

@wary osprey plz

[humble crypt]

Yes… it’s working and doing everything. But doesn’t send the messages on telegram after.. it just stops sending mid way

There is multiple msgs in help with this exact same issue… nobody knows how to fix??

[wary osprey]

Try this:

1. restart the browser session inside OpenClaw (not just refresh the page)
2. close all tabs and let OpenClaw open a fresh one
3. avoid mixing manual tabs with agent-controlled tabs, it can lose track of context
4. make sure the snapshot call is targeting the active tab (sometimes it sticks to the previous handle)

If it started after updating, it’s likely a session/tab tracking issue

[nimble hill]

looks like the monkey-patch did not apply properly - did you restart gateway after that was applied? The CLI timeout should have sorted this out. I have 10 differnet installation and it works eveyrwhere

[frozen star]

Im having an issue where my openclaw is currently telling me that its semantic memory is rate limited and not working?

the claw is running on a ChatGPT pro sub currently.

[drowsy isle]

I have two devices: Phone A (a dedicated device running OpenClaw) and Phone B (my personal phone). I am currently using Phone B to communicate with Phone A, and the agent is responding correctly. However, there are specific scenarios where I need to pause the responses from Phone A so that I can provide human intervention to resolve an issue. What is the best way to trigger a pause in this setup?

[lethal sage]

Hey I'm seeing high costs from input_cache_write_5m in my Anthropic billing, but my openclaw.json has contextPruning.ttl set to 1h. Is that config actually controlling the Anthropic API-level cache TTL, or is there a separate setting? Running 2026.3.11 on macOS.

Thoughts?

[orchid rune]

Can anyone tell me how can i create subagents in openclaw, make them a place to communicate in a discord channel in my server?

[wary osprey]

yeah I’ve seen that happen before and usually it’s when the agent is hitting limits on memory calls, especially if it’s reading/writing too often in a loop. check if your task is triggering memory repeatedly, or just restart the session and see if it resets cleanly.
also worth trying a different model briefly just to confirm it’s not tied to that setup. If you keep running into that kind of limit, it might just be the provider constraints. I moved mine off direct ChatGPT for that reason and moved to AIsa because I needed more control over usage and switching models when things start acting up

[frozen star]

interesting, I gave it an api connection per the memory documents, but it shot up to 100$ in 4 days using the small model, so now I'm trying to get something working again. Before this it was just getting stuck and not following through with anything, but the other day it was like 'oh I can spin up agents and not lose focus, here we go' and 'hitting semantic memory api limits'

[solid verge]

https://github.com/rsunil92/OpenViking.git i want to install this, anyone plese help me?

@nimble hill

@dreamy pagoda

@brazen nova

[civic thistle]

Looks like there is a Whatsapp (maybe other channels too?) regression on the latest update. Sub-agents & cron task can not report anymore their results, killing the task... Is it only on my config? I found some tickets linked on GitHub but not many upvote 🤷‍♂️

[north marsh]

Has anyone tried using the two free open router models, nemoclaw super and hunter alpha as a super sandboxed sub agent?

Since their prompts and outputs get shared with the providers

I’m a blue collar dude trying to set this up to help me run my two businesses. And I’m trying to maximize any free model use I can get. While still doing it safe safely.

But I’m really not a technical person before I bought the computer to run open claw on, I haven’t had a computer in my house for seven or eight years

[valid field]

I think no matter what my dude, you’re going to pay some money for good models to anything effective. What are your particular use cases and budget?

[sick kayak]

help. cant access my gateway dashboard. error: disconnected (1006): no reason

[sturdy girder]

Who can help me with chrome relay and my bot being able to access tabs etc

[north marsh]

Yeah, I know I’ll pay some. And I’m paying quite a few right now to help me through set up, because I’m not a technical person. So I’ve been testing my agent throughout the day and slowly expanding what he can do. And then when he breaks or fails something, I’m using Claude coworker to look at the back end at night and try and make him more durable. Use Case, I use Podium for one of my businesses, so all my text calls emails, Facebook messenger all come in into one spot. I wanted him to proactively scan that, and just move commitments to my calendar with my approval. Help me save and edit customer notes, and profiles. And keep track of their next appointments, my customer return for my business is 3 to 5 years

My plan right now to maximize usage with minimal recurring costs, was to pay for the $20 a month ChatGPT, and use oauth, as well as $20 a month ollama, and load a few different models there. And have it set up as dedicated sub agents for different tasks that are locked to a specific model.

And I’ve been using Gemini 2.5 flash, as my router and chat agent since he’s fairly cheap

Would love for anyone more technical than me to poke holes in that plan

[lone fable]

how are you installed? it could be due to pairing or if its not configured for discord or another service. I have it in a docker container and it generally tells you what is wrong in the logs

Im going to bed, I stayed up all night again, jeeze

[sick kayak]

virtual box

can login to gateway

cant

[lone fable]

did you go to the overview tab?

pasted the gateway key from openclaw.json

[sick kayak]

what do you mean

[cedar torrent]

anyone managed to run smooth on a free 16k capped nividia api? i'm blowing through the context window perpetually

[sick kayak]

i did this

[valid field]

That’s pretty solid. Seems very doable with your setup. For local models though you usually need a ton of memory on your machine and they often aren’t worth it IMO yet. I would replace that with just using the recently released gpt 5.4 mini and nano models for those easier subdagent tasks to save on tokens.

[north marsh]

I got a max studio M4 with 36 GB of RAM locally for a really good deal. I was gonna try qwen 3.5 27b once I get stuff debugged

[viral ether]

I've had this same issue forever as well. Any task taking more than a few minutes times out I and have to ask my claw "progress" and it tells me it's done. No one has every been able to answer this

[worldly garden]

Has anyone solved the issue with telegram cutting messages in half? Even shorter ones before 4000 chars

It cuts messagse half-way through when it's a table or some other response

[hasty grove]

there’s an openclaw.json field that limits total tokens returned, separate from context window size config

[valid field]

What’s the problem you have? I don’t use chrome relay anymore. Use chrome mcp or the normal cdp path. Wayyyy more stable.

[tepid shoal]

wow! Since AI generated materials aren't copyrightable, patents will be continue to be the kings of IP. I imagine you saw some room for improvements in the home care equipment you were using.

[valid field]

Soliddd. For a non technical person Openclaw can be very off putting. But huge potential to save you money in your business.

[ember jolt]

Hi team,

I've been using OpenClaw for a few weeks but more like a chat bot than anything else.

Recently I've tried to get it to work on things like helping me create an app for mobile devices and it just becomes a very long task.

I'm using Gemini 2.5 and chatting via Telegram.

It seems to not know anything and does not validate it's processes before actioning it or recommending it, dispite repeatedly telling it to.

Could it just be Gemini not being capable of doing this?

Which Llm would be recommended (free/cheapest) to get going on this?

I'm running this on a ESXi bare metal workstation on VM running Ubuntu.
Should I run multiple VMs and have them on different LLMs so they can cross check?

Maybe I'm just not going about it the right way

[valid field]

What is your use case for this?

[valid field]

Don’t use 2.5. lol. You need opus 4.6 or 5.3 codex minimally.

[proud canyon]

Is there any guide to learn clawdbot?

[worldly garden]

Which one?

[north marsh]

If I posted my architecture plan in here, with all the integrations I’m hoping for maybe you guys could poke holes in it?

[mild minnow]

would this be the a1,flex vm?

[unique hazel]

Money isn't a problem: use Opus + GPT 5.4
On a budget: use GLM 5/ Kimi K2.5 + MiniMax M2.7

[sick kayak]

when using custom provider as llm, gateway dashboard wont work

[midnight rapids]

How the heck do you get your claws to do long running sequential tasks? I'm specifically trying to test out a bunch of different local LLMs and it forgets to test the next one after each test.

[sick kayak]

Gateway: not detected (gateway closed (1006 abnormal closure (no close frame)): no close │
│ reason)

[wanton prairie]

Folks, if I link openclaw to my private discord server, does each channel have it’s own memory and theme ?

[midnight rapids]

This worked:

Let's test your chaining. Create a todo list of 4 things to do. Each todo is to be given to a new subagent and it should sleep for 3 seconds and then do a web fetch to https://time.is/ and tell you what time it is. The subagent should be the current active local llm. Work on these sequentially and report back when each one finishes. Tell the subagent to tell you to work on the next todo.

Nope, that didn't work.

Trying beads now

[mild minnow]

hit with the classic capacity reached

[solid verge]

i want to add memory layer

[hasty grove]

The maxTokens field in OpenClaw's configuration defines the maximum number of tokens the model can generate in a single response. This setting is part of the model definition within the models.providers section of the openclaw.json configuration file.

[tame beacon]

Hello i have a small problem. I am using OAuth With OpenAI but there is an error and sadly i do not know what should i do. Tried to restart everything in config etc. 401 Incorrect API key provided: "XXXX" You can find your API key at http

[zenith bone]

What are good backup brain options to fallback to after gpt plus plan rate limit?, using openrouter as backup.

[worldly garden]

Got it. It's more of the reply itself on Telegram which is cut in half even for shorter messages

Maybe it's a problem with markdown handling

[hasty grove]

Ah. Try using krill or internet connected agent to troubleshoot because there are articles that suggest some models have streaming issues with Telegram + recommend a few paths forward

[ember jolt]

Can one use 2 LLMs on a single OpenClaw computer?

[hasty grove]

Also enable o11y so you can be certain about what your LLM inputs and outputs actually are so that you don’t waste time due to faulty assumption

[ember jolt]

Thanks

[heady shell]

Hi everyone, how do I connect a remote MCP server (such as Parallel AI) to OpenClaw v2026.3.8 on a Raspberry Pi?

[sick kayak]

fix this by puting the api key directly in gateway service. openclaw seems to not work if api key is in .bashrc. systemD cant find

it

[heady shell]

Yes

[keen flame]

quick question, is it possible to use MiniMax M2.7 on OpenClaw via their Coding Plan Subscription. Or do I have to wait for a new OpenClaw update?

[karmic sierra]

ok, so i am building out an agent workflow with master prompts that operate in several scenario. the first is a complete box set of everything a single guy could need that has little or no money, a few mac products laying around like a mini, mac book pro, maybe an air. A truly one-man enterprise with Fortune 500-level infrastructure.

Every agent runs on Google infrastructure. All memory lives on Drive. All automation runs on Cloud Run or App Script. The whole system costs roughly $250/month for FIVE people - while everyone is using all these free models it will have every main agent using top of the line paid services.

The AI Layer — Nearly Free Inference:
Model Best For Cost
Gemini 2.5 Pro. NAME — highest reasoning, external work Your subscription
Gemini 2.5 Flash. Agent,NAME — fast, smart, cheap Fractions of a cent per call Gemini 2.0 Flash. Agent name -most sub-agents. Essentially free at scale Gemini Flash-Lite High-volume routine tasks Near zero
Gemma 3 (local) Air-gapped tasks, no API needed Completely free
Code Assist. Agent — scripting, debugging, building Included
Gemini Nano Quick classification, routing, tagging Free

why is everyone using a bunch of free stuff, join in to shared commmunities use the advanced tools for pennies. i will be allowing 5 others to join the family for cost sharing. they will have access to all the agents for 50-100 a month and maybe $50 a month in fees for high power toools, antigravity, notebookxl, gemini 3.1 pro, deep think, 30TB shared drive, (3-5tb per person alloted), dozens of tools, credits for api tokens, ai studio, google cloud, and soooooo much more

actually i'm working on a memory fix that would allow live memory for 10 full time agents cost less than $100 a year

which with the $100 a month free token credit that purchases a billion token would be free

so vomit on that

[tall meteor]

So far, I have paid $0 by running my own Ollama server
qwen3:14b and ministral-3:14b are plenty powerful.

[karmic sierra]

and I am paying $0 for gemini ultra when there are 5 total members

[tall meteor]

Also, I am building a self-contained off-grid network

No need for exteneral API calls

[karmic sierra]

oh and getting youtube pro free, 30TB free, every thing google owns pretty much added free, ai studio free, notebookxl free, antigravity, gmail pro services, theres so much crap they pile in here, hosting back end $1,000 credit, front end free, add credits $1,000 free

honestly if you have a group of 4 other freinds who are willing to all pitch in $50 everyone gets al of the tools at the same level.

[tall meteor]

Oh man, a 30TB NAS box on my network would be awesome...NVMe

[karmic sierra]

proper set up you can run 200 agents on a mini

gemini api rate limits are 1,000 request a min on the free version lol i mean jesus you can run dozens of agents

$100 credit monthly thats a billion tokens, with a orchestrator agent set up you can uses n8n and run thousands and thousands of flows at the same time, and thats knowing there are programers way better and above me can do better

with that set up you can do on 1 mini, 1 pro what a company who went out and bought 200 mac minis for free, no api cost.

thats not 200 nodes, not saying that, sub agents i think the math it could push up to 50 pushing hard prob better than i coud ever imagine doing

[proper thorn]

I'm quite confused about getting the QMD memory enhancement running smoothly. I have notes all over the bootstrap files to write and READ the various memory files. While writing is still necessary(?), READING should be not. It is supposed that QMD takes care of this. Does it? Is it supposed to delete all the "read your memory files" hints from the bootstrap *.md files? None of the HowTos ever mentiones them. I would assume that having the agent continue to read the files, would result in recursive growth of the context.

[karmic sierra]

compress

i cant say much on this topic of memory ill just throw that out there otherwise someone gonna steal my shit lol

[junior acorn]

QMD for memory?

Wat do you need help with?

[pseudo oyster]

Is the heartbeat not working because the request gets timed out on the gateway a know issue ?
Running on a pi 4, dashboard and telegram are working.

[wheat marlin]

i just installed claw on a vps, already working but when talking on telegram with it, it says it doesn’t have internet connection, neither can use terminal or edit files…

[proper thorn]

Well, then i hope for someone wo does. I asked my bot to answer a post only when adding something substantial and just read on, if the answer would be just a "yeah, that's a problem..".

[junior acorn]

User issue, works for most

[junior acorn]

It’s because it doesn’t. You need to configure it

[true canopy]

Hey everyone! New to OpenClaw here 👋

Just finished setting up my agent (Gargamelle 🦞) and I'm running into rate limit issues pretty fast with Claude Sonnet 4-5. Hitting the API limits quickly and it's getting expensive.

A few questions:

1. What model do you guys use day-to-day to keep costs low?
2. Is Gemini Flash as a primary model + Claude as fallback a good setup?
3. Any tips for managing token usage / session resets?

Running self-hosted on my Mac for now, plan to move to a VPS later. Would love any advice from people who've been through this! 🙏

[wheat marlin]

anywhere i can find a quick guide to configure these basic things

[karmic sierra]

there are like 3 or 4 memory systems out there. you need to set a limit so it clears every day or session from the telegram session. you have more than one instance. telegram is costing tokens and memory, then you hve another session of the exact same conversation running on the backend thats talking to openclaw.

[junior acorn]

Have you read the docs?

[karmic sierra]

goto openclaw, go to memory and start reading best place

[proper thorn]

I'd hope to get an answer to the question "if installing QMD, do I have to adapt the AGENTS.md to reflect this, eleminating all the 'read the memory files' hints"

[karmic sierra]

or tell your agent to go read that entire memory page top to bottom

[junior acorn]

No you don’t

[karmic sierra]

install the chrom extension, turn it on for that page

[junior acorn]

Wats ur question? You type like a bot

[karmic sierra]

or, open chatgpt, gemini etc tell it you want to build memory and have your agent go build out its memory, give it a short term memory that it compresses twice a day and clears out, give it a long term memory or diary to look back on so when the session memory is cleared it has stuff to read, then give your soul and so it remembers the base stuff. i'm just generalizing it. bettr to read the page.

[true canopy]

yo ok

sorry I am french

using ai to text in english

[karmic sierra]

just remember everyon, you gave a bot, he knows how to do just about everything. just tel him to tell you what needs to be done.

[pseudo oyster]

Let me put it differently. Has anyone else experienced that issue/knows what's causing it or how to solve it?

[karmic sierra]

and make sure you create a god rule!!!! gotta have this or your openclaw will commit suicide

if your request is getting timed out then your chat isnt talking to your inner self possibly

openclaw status

run that, scroll back see what is failing

identify whats wrong, open gemini or gpt, paste the error inside

it will advise what and how to repair