users-helping-users

[spark valley]

i recommend the stepfun unlimited no quota and competes with top as it is fast reasoning

[grand basin]

i just set it to only free as a fallback to debug

lmfao i did /reset and now nothing works

400 codes

😄

does /reset clear the API key or something

[spark valley]

that only resets the session

you send it as a messge

[grand basin]

yea now im getting 400 status code responses lmaoooo

oh wait i did doctor--fix

thats why

cerebras cant take /developer roles so i have to re-add the /system role

[spark valley]

hope it worked

[covert gazelle]

I hope it will help you. Maybe you could also try to use alternative to ollama, many options I think. I don't use local models so my experience is limited. Let me know how it evolves!

[spark valley]

🦞 OpenClaw 2026.3.11 (29dc654)
🧠 Model: openrouter/hunter-alpha · 🔑 api-key (openrouter:default)
🧮 Tokens: 13k in / 24k out
🗄️ Cache: 98% hit · 512k cached, 0 new
📚 Context: 513k/1.0m (49%) · 🧹 Compactions: 0
🧵 Session: agent:hacker2-nemo:main • updated just now
⚙️ Runtime: direct · Think: low
🪢 Queue: collect (depth 0)

[grand basin]

nope idk what doctor did but it nuked my config

lololol

side effects of vibe coding

😄

[stiff crow]

Same, Idk anything about programming and code languages

[manic marten]

has anyone successfully setup sandboxes? seems like a nightmare

[stiff crow]

Nd struggle with codes

[stiff crow]

What are sandboxes

[grand basin]

oh i do but most vibe coded stuff is such a pile of shit

like it'll work, and may work well, but its not human style of structure

[stiff crow]

T-T idk shit in coding domain

[stiff crow]

human style of structure

Hmm

[spark valley]

"sandbox": {
"mode": "on"
},

[grand basin]

facts

[stiff crow]

💔

Out of my understanding

[mighty rampart]

What problem are you trying to solve with it? Are you already familiar with Docker and containers?

[sullen dome]

Hey guys, I’ve set up a morning brief with my agent but he keeps forgetting to do it. Any ideas(

[grand basin]

anyone know why chatgpt's models are like "we cant modify your system"

I told it explicitly that it can modify my entire computer if it wants and its like "heres the drop in code"

[mighty rampart]

How did you set it up? Make it a cron job and it will happen.

[spark valley]

allow tools ,

[grand basin]

???? how do i do that

i thought that was on by default

[spark valley]

in the config lool , sometimes it is not

[grand basin]

i actually dont know wher ethe config is

lol

oh .config

[spark valley]

in the coonfig go to tool i will show u

"id": "main",
"model": {
"primary": "openrouter/stepfun/step-3.5-flash:free"
},
"subagents": {
"allowAgents": [
"support"
]
},
"tools": {
"profile": "full"
}
},

in /.openclaw/openclaw.json

[grand basin]

lol i dont have id main

you'd think this would be in the onboard

I nuked the whole thing and restarted from scratch

[ember eagle]

Can someone please help me understand why it is beneficial to purchase a mac mini specifically for OpenClaw? Is it because this is a more secure method and the Mac Mini essentially becomes a Sandbox for the agent to live inside of?

[grand basin]

and told it, it can do w.e it wants and it looks like its working

[grand basin]

no specific reason, just the early builds were on macos and san fran likes to goon over apple

people have run openclaw on cellphones

[spark valley]

look the whole hype about the mini has nothing to do with sandboxing or stuff you can set it up on your old computer or laptop , and it will work best the mini probably if you'd run local llms then yup that a better aproach but if you have apc laying out there or a mini pc it can do it too

[ember eagle]

wowwwww

[swift willow]

It has unified memory with the CPU and GPU this means it can run some models on it own locally

[ember eagle]

But does anyone ever do that?

Run models locally

[spark valley]

i set mine n a i7-3 16ddr3 gtx 680

[spark valley]

yeah but depends on your needs

what you want fromm openclaw

[swift willow]

it’s not about Apple gooning, just means you could run some LLMs locally without forking out for 5090 GPU

[ember eagle]

Have any of you had luck running local models? I saw Nemotron 3 might be a good option

[swift willow]

The Apple hardware is quite impressive, but your not exactly going to get blazing speeds or high inference from it

[swift willow]

I run local models yes mainly Qwen 3.5

[ember eagle]

Oh veryc ool

Do you use Ollama?

[swift willow]

I do yes

[grand basin]

How can I make it so openclaw stops asking me for permission, i just want it to keep iterating on its past successes

[spark valley]

if you are a newbie and wouldnt want to waste money on a mini you wouldn't need but for this reason i recommend you use providers apis untile you feel like you know your way around and you are required to run locally then go for it

[spark valley]

i sent you the config above

you can install the proactive skill

from clawhub

[olive cliff]

Just spent the day buring 59.2M tokens across 12 Agents on free models. I was stress testing cross-team interoperability working in tandem with eachother with 0 human in the loop interruptions. Context windows are the death of me.

[swift willow]

I wouldn’t run out buying a mini thinking your going to to be running large parameter models

[grand basin]

oh i forgot to add subagents to the config

[swift willow]

another thing about the Mac M chips is power consumption is very low so that’s another factor

[ember eagle]

Yeah that makes sense, though power is already pretty cheap in my area

[spark valley]

no the tools param is sufficient try to restart the gateway .

[ember eagle]

Tbh I've considered trying to make my own GPU cluster to run more powerful models. I'm incredibly tired of shelling out $ for API costs

[spark valley]

yeah if you are really using that api like squeezing tokens you must look int it

[ember eagle]

Since GPT first launched, I've probably spent >$2k on API calls

Though I have also heard you can use the Agents SDK key that connects back to the Anthropic account so idk, maybe I'll try that

[spark valley]

yeah oauth for openclaw

but i guess it also limit when u hit quota

even on pro but u may want to give it a shot

[ember eagle]

Hmmm maybe its worth it

[spark valley]

thats a better low cost approach

[ember eagle]

have any of you built out custom dashboards to see what the heck is going on with the claw? I tried using the stock version and Im pretty sure it showed me absolutely nothing and never updated

[grand basin]

my running icon disappeared lol

idk if its working or not 😄

[winged path]

Has anyone been able to get the new chrome mcp feature work in a openclaw-in-docker setup? I am trying to allow openclaw connect to the host browser.

The DevTools endpoint is reachable and healthy but OpenClaw still fails to attach with:
Chrome MCP existing-session attach failed for profile "user".
Make sure Chrome is running, enable chrome://inspect/#remote-debugging, and approve t
he connection.
Details: McpError: MCP error -32000: Connection closed

and there is no visible approval prompt in Chrome.

[mighty rampart]

Is the browser also inside the docker container? How are you trying to reach it?

[grand basin]

if openclaw says its idle is it still doing stuff? it says its working on thing but i have a feeling its not

[indigo nebula]

Tell it exactly that and tell it to save that as a primary directive into its .md files and to load that directive any time the system restarts. You should also tell it to create a LESSONS.md file and anytime it doesn’t do what you ask tell it that it ignored your directive and should make a note in the LESSONS.md file as a way to learn from its past mistakes.

[winged path]

OpenClaw itself runs in Docker, but the browser I want to control is Chrome running on the Linux host, not
inside the container.

my current setup:

• OpenClaw container: openclaw-personal
• Container is now using network_mode: host
• Host Chrome has remote debugging enabled on 127.0.0.1:9222
• From both the host and from inside the container, I can successfully hit:
  • http://127.0.0.1:9222/json/version
  • http://127.0.0.1:9222/json/list
• I also verified from inside the container that I can open the webSocketDebuggerUrl and send a raw CDP
  command (Browser.getVersion) successfully

[mighty rampart]

Usually From inside the container, 127.0.0.1 ot localhost is is “itself”, and you need to use host.docker.internal to get out to the host.

[stable inlet]

wgwere do we gwt the app? for openclaw

[median copper]

Hi all. I am unable to get the bot in Discord to create a new channel. It keeps asking me for the guild ID. I did add that in the config file but still unable to get it to create a channel. Can someone please help? Thanks.

[earnest trout]

Hi all, i'm trying to add channel (telegram) but met with this code

[2964:0x2d47b000] 65912 ms: Mark-Compact 980.6 (1002.6) -> 979.8 (1003.9) MB, pooled: 0 MB, 4265.99 / 0.00 ms (average mu = 0.189, current mu = 0.018) allocation failure; scavenge might not succeed
[2964:0x2d47b000] 70405 ms: Mark-Compact 981.9 (1003.9) -> 981.1 (1008.6) MB, pooled: 0 MB, 4431.48 / 0.00 ms (average mu = 0.106, current mu = 0.014) allocation failure; scavenge might not succeed

<--- JS stacktrace --->

FATAL ERROR: Reached heap limit Allocation failed - JavaScript heap out of memory
----- Native stack trace -----

1: 0xe42d60 node::OOMErrorHandler(char const*, v8::OOMDetails const&) [openclaw-channels]
2: 0x121ddd0 v8::Utils::ReportOOMFailure(v8::internal::Isolate*, char const*, v8::OOMDetails const&) [openclaw-channels]
3: 0x121e0a7 v8::internal::V8::FatalProcessOutOfMemory(v8::internal::Isolate*, char const*, v8::OOMDetails const&) [openclaw-channels]
4: 0x144ba05 [openclaw-channels]
5: 0x1465299 v8::internal::Heap::CollectGarbage(v8::internal::AllocationSpace, v8::internal::GarbageCollectionReason, v8::GCCallbackFlags) [openclaw-channels]
6: 0x1439948 v8::internal::HeapAllocator::AllocateRawWithLightRetrySlowPath(int, v8::internal::AllocationType, v8::internal::AllocationOrigin, v8::internal::AllocationAlignment) [openclaw-channels]
7: 0x143a875 v8::internal::HeapAllocator::AllocateRawWithRetryOrFailSlowPath(int, v8::internal::AllocationType, v8::internal::AllocationOrigin, v8::internal::AllocationAlignment) [openclaw-channels]
8: 0x141354e v8::internal::Factory::NewFillerObject(int, v8::internal::AllocationAlignment, v8::internal::AllocationType, v8::internal::AllocationOrigin) [openclaw-channels]
9: 0x18740bc v8::internal::Runtime_AllocateInYoungGeneration(int, unsigned long*, v8::internal::Isolate*) [openclaw-channels]
10: 0x1dd1476 [openclaw-channels]

what does this mean? i can't seem to get a straight answer from the bots

[mighty rampart]

What kind of system is it running on? Looks like it ran out of memory.

[earnest trout]

@mighty rampart thanks for the reply! i'm using digitalocean VPS

[mighty rampart]

Details? A brief search suggests at least 4GB is recommended.

[earnest trout]

it has 2 GB Memory / 50 GB Disk. However this is not the first time i'm using it, i've set up many openclaw setups for the past days with different VPS (only to discard it in the end because of coding issues). But this is the first time i'm stuck at adding a telegram channel with the same VPS setup

[indigo nebula]

Well changing to a 4GB memory is a simple process, upgrade the memory, reboot and see if it works. If it does you know what the problem is. If you still get the error you’ve ruled it out as the cause. Plus once installed you may be able to drop back down to 2GB of memory.

[earnest trout]

ok let me try upgrading now and retest

[meager vessel]

which is the chepeast ai model subsciption?

[mighty rampart]

Sign up for OpenRouter and put at least $10 credit on file. They have some free models and you can try out others. What’s cheapest depends on what you need it to do. (It’s not cheaper if it’s wrong and you have to do it over.)

[brave swallow]

what is the cheapest you can make token consumption i wonder. like cutting back on waste. If I had a mac mini I could make it so that it could use the web and interact like a human does right?

i mean in certain ways a vps can too but still

im still debating whether i want to get back into open claw

[tacit ether]

Im working on adding support now

MCP for Docker will be added and is now ready for the next release

[solemn peak]

What's the best "openclaw setup for beginners" video in YouTube?

[blissful berry]

What's the best configuration for using anthropic with openclaw as an escalation for my existing agents? From what I read there are 2 different systems with 2 different mehtods of billing, and one you might run into issues using non-official clients...

[indigo nebula]

First thing you should do is talk to your claw and ask it “to analyze the openclaw.json and make sure you’ve optimized all of the LLM settings and enabled caching and other efficiency prompts per the docs.openclaw.ai website.” Mine enabled a number settings we were missing in the config. The next thing you can do it tell it how to be more efficient with your token usage. Tell it to do the following:

“Save the following to your .md files and read them when you start up a new session in order to optimize token usage.”

Token Efficiency

• Never re-read files you just wrote or edited — you know the contents
• Never re-run commands to "verify" unless the outcome was uncertain
• Don't echo back large blocks of code unless asked
• Batch related edits into single operations — not 5 edits when 1 handles it
• If a task needs 1 tool call, don't use 3 — plan before acting
• Do not summarize what you just did unless the result is ambiguous

💰 Token Economy Protocol (Standing Rule)

Goal: Minimize expensive model usage. Stretch session limits by using the cheapest model that can handle each step.

Model Tiers

Tier	Model	Alias	Use For
💰 Expensive	‘Best Model Name’	your model-#-#	High-level planning, tricky debugging, architecture decisions, final review, when asked
🟡 Standard	‘Mid level model name’	your model name-#	Medium-complexity reasoning, context-heavy implementation, integration work
🟢 Cheap	‘Free or cheap model’	your model-#-#	Straightforward coding, boilerplate, simple scripts, formatting, tests, file parsing, cron jobs, simple lookups

[brave swallow]

that is really important important to keep in mind i might have to copy it and save it somewhere

i guess i should have been posting in general though

i wish i knew how to make it go to different models without it breaking itself in the past

free models might be worth looking into. or super super cheap ones

[indigo nebula]

You should also talk to your claw about downloading various SKILLS from clawhub that are relevant to what you’re doing iwth it. Go browser and find ones you think you need then tell it to install them for oyu. No point in reinventing the wheel or having it do things that could be done with existing skills or tools. If it has to create a script that burns more tokens than loading a SKILL you downloaded and told it to use for a specific take

[brave swallow]

yeah back when i started using open claw skills were too dangerous to be trusted

most things had nefarious code in it

or prompts

[blissful berry]

Oh I think I understand a bit more, will I run into issues using setup-tokens from my max subscription in openclaw?

[brave swallow]

i dont actually have a running open claw though right now. still debating whether I should use it again right now

[indigo nebula]

Use a better model, like Opus or Sonnet that is good at complex logic. Tell it you want to setup multiple LLM models so you can save on tokens including using free ones. It will help you do it. But the lower level and free models aren’t always so good at that kind of stuff. Use a quality model to add the other ones.

[brave swallow]

so in the past it made the mistake of "switching models" and it would do it wrong and break itself.

it would switch it in the wrong way and break its ability to respond

i was using opus or sonnet too

and it would break itself

[indigo nebula]

You could also go the router of openrouter.ai. Check out their site they have openrouter/auto model which will choose the model automatically based on whether you are just chatting, doing complex stuff, etc.

[brave swallow]

what is openrouter

oh i think i see

[indigo nebula]

It is an opensource LLM routing system. Go see https://openrouter.ai

[brave swallow]

Honestly im still debating if i want to do all of this again

idk if its worth it

[indigo nebula]

They have a “auto” router that will choose the best model for you or you can pick a specific model

[brave swallow]

i dont supposed thats something you can help me decide @indigo nebula

im really struggling to make a decision here

[indigo nebula]

Do it or don’t, its your decision but it ain’t that hard. Make a choice and move forward.

[waxen quail]

anyone experiencing CLI sub-agents being broken? gateway captures output but never writes it back

When OpenClaw uses Claude Code or Codex as a sub-agent, the agent runs and responds. But that output never makes it back to the parent session. It just disappears. The parent gets "(no output)" every time.

Same thing happens with ACP spawns, the agent finishes but the result never relays back.

[next grove]

Anyone did macmini then moved to vps after? How does it differ? I’ve read that mainly browser actions are limited but otherwise pretty much the same

[violet crystal]

hey when do you guys think sub-agents are useful vs just building an agent? anyone have any preferences or best practices they've seen?

[violet crystal]

ill lyk once i setup agents or subagents lmaoo still trying to figure out which is more ideal for my use case

[olive cliff]

If you guys aren't taking advantage of hunter-alpha and healer-alpha through openrouter, I HIGHLY RECOMMEND IT! Ive burned through 69.2Million tokens today alone, for absolutely free. Im just stress testing the hell out of my 12 Agent set up right now.

[violet crystal]

what are your 12 agents bro 😭 im tryna figure out if i should even create more or just use subagents, seems like overkill

[olive cliff]

I came across a github repo with templates for over 144 agents, fed it to my opencal and told it to spin me up a team of 6 agents for design, and 6 agents for development. We configured roughly 100 skills across all agents that specializes them in there individual roles within their teams.

[thorny cobalt]

⚠️ Agent failed before reply: All models failed (5): openai-codex/gpt-5.4: ⚠️ API rate limit reached. Please try again later. (rate_limit) | anthropic/claude-haiku-4-5: ⚠️ API rate limit reached. Please try again later. (rate_limit) | anthropic/claude-opus-4-5: ⚠️ API rate limit reached. Please try again later. (rate_limit) | nvidia-nim/moonshotai/kimi-k2.5: ⚠️ API rate limit reached. Please try again later. (rate_limit) | anthropic/claude-sonnet-4-6: ⚠️ API rate limit reached. Please try again later. (rate_limit).
Logs: openclaw logs --follow

Im trying to build a fcking controll panel set up since days and i waste all my weekly quota for shit.

[olive cliff]

https://github.com/msitarzewski/agency-agents

[olive cliff]

Abuse the hell out of openrouter with hunter-alpha and healer-alpha. Ive burned 69.2million tokens with over 1k requests today for $0.

[unkempt knoll]

Hey everyone — looking to hire a developer for a paid OpenClaw project. Built a detailed spec brief for 4 automations: Tonomo → SmartSuite browser transfer, 2-way live sync, iMessage booking to Tonomo, and SmartSuite deadline alerts. Real estate media company in San Diego. Fixed price, screen-share sessions required (client wants to learn as we build). Brief ready to share. DM me if interested or if you know someone who’d be a good fit.

[violet crystal]

these are all sub agents so that's what you use?

not the full duplicate openclaw agent workspaces?

[thorny cobalt]

Whats that?

[violet crystal]

looks like free models, pretty crazy, but the catch is all your data is given to them

inputs and outputs, but i mean most providers are doing that anyways 🤣

[olive cliff]

I made full fledged agents with their own workspace, identity, soul etc based on these. They all have their own telegram bots, and I can interact with them independently in my Control UI.

[violet crystal]

ohhh interesting, what's the reasonign for that over just using subagents as is? i feel like thats a lot of overhead and token burning for not that much upside?

[olive cliff]

https://openrouter.ai/openrouter/hunter-alpha

Set up openrouter with an api key in your openclaw.json file, add the openrouter/hunter-alpha model. You can have your claw set this up for you too.

[olive cliff]

Im just testing various deployments of teams. It can be a lot of overhead if you aren't careful. Imo, its better to have specialized agents that are really good at doing 1 thing, rather than being a generalist.
It helps keeping context tight, and not too much noise bothering a single agent for everything. If set up right(not how im currently stress testing with 20minute cron jobs and 10 minute heartbeats for 12 agents) it could end up being about the same token burn as a single agent.

[tidal pelican]

Anyone Have any good ideas for getting cache read costs down?

[olive cliff]

Even my main Agent thinks this was a ridiculous stress test:

Your setup was basically:

1. Hire 12 employees
2. Give them 20-min check-ins
3. Pay them in api tokens
4. Watch the Lamborghinis evaporate

Free tier = the difference between "startup demo" and "VC intervention." 😂

Thank the AI gods for hunter-alpha and context compaction! 👻

[clever jetty]

Anybody have an issue with their openclaw not receiving group messages at all in signal?

"signal-cli group messages not reaching daemon. Account registered as primary today, deviceId=3, isMultiDevice=true. Group ID allowlisted with requireMention=false. DMs work fine. Group messages don't appear in logs at all — never reach gateway level. signal-cli v0.14.1."

Trying to fix so can use the bot in a group signal chat. Thks

[turbid lantern]

Should there be a cronjob for having heartbeats work? I thought they should just work on their own, but they are not. Any other things to look into?

[zealous escarp]

idk what to do anymore, im running llama3.1b locally and if I run ollama run llama3.1b I can chat to the ai fine, but then if I try to chat to it thru openclaw tui it doesnt work, it sits loading, I hear my gpu ramp up in speed as it gets warmer so I know its recieving it, it does the same when I talk to it directly thru ollama run, and I know its the correct model selected aswell

[spark valley]

must be config issue or routinng

[prisma tiger]

Is it possible to use a codex model (oauth not api key) with acpx?

[turbid lantern]

you can try with https://github.com/cola-io/codex-acp but there is no official acp afaik

[olive cliff]

Cron jobs and heartbeats are 2 seperate things that kinda of do similar stuff. I typically just ask my agents "set a cronjob/heartbeat for 3 hours to [insertWhateverHere using [randomModelHere]" and they typically just do it. If I ever notice a cron or heartbeat fail, I ping them with it to fix it.

[turbid lantern]

according to my claw, only the main agent has heartbeats, all other agents don't. but it doesn't seem plausible. what happens is that I have a bunch of agents, and I configured heartbeats configuration and a heartbead.md file for each, but none of it triggers.

[olive cliff]

Are you other Agents set up with unique workspace, identity, soul, and agent MD etcetera etcetera files? (Ive got 12 fully operational)

[turbid lantern]

yep

I'm missing something. I don't know what.

[olive cliff]

What do you use to message them directly?

[hoary ibex]

Which is prefer way of installation with macbook setup, cli or Mac app install?

[turbid lantern]

Telegram, it all works fine. Even have one (or even two) agents that has heartbeats working.

[olive cliff]

Ask them individually to set their crons or heartbeats for whatever task you want them to stay on top of.
Do not set multiple crons/heartbeats for short time frames(under an hour). You'll burn tokens unnecessarily. I burned over 70m tokens across 12 agents today with 20 minute check ins with the hunter-alpha model(free).
Ive got mine set up in 2 different telegram groups. If youre only asking your main Agent to set the crons/heartbeats for the others, that might be where your disconnect is. Although idk why, cause my main Agent has no issues configuring things for my other 12.

[olive cliff]

^^ (forgot to reply)

[turbid lantern]

I have a fixed cost for the tokens, so don't care about some burn. I got crons setup mostly in my DM topics, not the group. I am asking my main agent to configure the other ones, but there shouldn't really be an issue, it just modifies the openclaw.json and heartbeat files wherever they are ... I can verify that easily.

[stuck bone]

is Oauth works for any models out ther with openclaw? I mean subscription based usage. Someone said chatgpt can? please confirm.

[fringe oak]

is there a thread somewhere for fixing the mcporter spawn issue on windows 11?

im trying to connect a localhosted custom mcp server

[olive cliff]

OpenAI 0Auth for codex works. Its the only one I tried for a short time.

[olive cliff]

Are you confirming on an individual agent basis that the crons or heartbeats that your main Agent set for them, are operational on their end?

[turbid lantern]

the same thing happens when agents set their own heartbeats. they just miss something somewhere, other than the heartbeat.md file and the section for headerbeat in openclaw.json's agents.list ... something else is needed and I don't know what it is.

[olive cliff]

Going to start asking dumb questions just to narrow things down....
Do you have the heartbeat and cronjobs sections in your Agents Agents.md files? (Separate from the heartbeat.md file itself)

[turbid lantern]

no, should it be there?

[olive cliff]

YES!! That gives them instructions on how to read it.
Here's a copy/paste from one of my Agents.md files

## :heartbeat: Heartbeats - Be Proactive!

When you receive a heartbeat poll (message matches the configured heartbeat prompt), don't just reply `HEARTBEAT_OK` every time. Use heartbeats productively!

Default heartbeat prompt:
`Read HEARTBEAT.md if it exists (workspace context). Follow it strictly. Do not infer or repeat old tasks from prior chats. If nothing needs attention, reply HEARTBEAT_OK.`

You are free to edit `HEARTBEAT.md` with a short checklist or reminders. Keep it small to limit token burn.

### Heartbeat vs Cron: When to Use Each

**Use heartbeat when:**

- Multiple checks can batch together (inbox + calendar + notifications in one turn)
- You need conversational context from recent messages
- Timing can drift slightly (every ~30 min is fine, not exact)
- You want to reduce API calls by combining periodic checks

**Use cron when:**

- Exact timing matters ("9:00 AM sharp every Monday")
- Task needs isolation from main session history
- You want a different model or thinking level for the task
- One-shot reminders ("remind me in 20 minutes")
- Output should deliver directly to a channel without main session involvement
**Tip:** Batch similar periodic checks into `HEARTBEAT.md` instead of creating multiple cron jobs. Use cron for precise schedules and standalone tasks.

**Things to check (rotate through these, 2-4 times per day):**

- **Emails** - Any urgent unread messages?
- **Calendar** - Upcoming events in next 24-48h?
- **Mentions** - Twitter/social notifications?
- **Weather** - Relevant if your human might go out?

**Track your checks** in `memory/heartbeat-state.json`:

```json
{
  "lastChecks": {
    "email": 1703275200,
    "calendar": 1703260800,
    "weather": null
  }
}

When to reach out:

• Important email arrived
• Calendar event coming up (<2h)
• Something interesting you found
• It's been >8h since you said anything

When to stay quiet (HEARTBEAT_OK):

• Late night (23:00-08:00) unless urgent
• Human is clearly busy
• Nothing new since last check
• You just checked <30 minutes ago

Proactive work you can do without asking:

• Read and organize memory files
• Check on projects (git status, etc.)
• Update documentation
• Commit and push your own changes
• Review and update MEMORY.md (see below)

🔄 Memory Maintenance (During Heartbeats)

Periodically (every few days), use a heartbeat to:

1. Read through recent memory/YYYY-MM-DD.md files
2. Identify significant events, lessons, or insights worth keeping long-term
3. Update MEMORY.md with distilled learnings
4. Remove outdated info from MEMORY.md that's no longer relevant

Think of it like a human reviewing their journal and updating their mental model. Daily files are raw notes; MEMORY.md is curated wisdom.

The goal: Be helpful without being annoying. Check in a few times a day, do useful background work, but respect quiet time.

Make It Yours

This is a starting point. Add your own conventions, style, and rules as you figure out what works.

[turbid lantern]

I'll give it a try.

[olive cliff]

i edited the copy/paste, i missed a few lines

Think of the Agents.md file as an instruction manual on where the corresponding md files are, what they do and how they are handled. So if you are having issues with identity, soul, skills, tools, memory, heartbeats, or crons, check to see if it is all properly referenced in the Agents.md file. If something is missing, you could use you main as a template and edit things as needed.

[turbid lantern]

yep, didn't think of it this way. I'll go by the instructions at https://docs.openclaw.ai/reference/AGENTS.default and see if that solves the issues

[shut onyx]

Guys, can I use openai oauth to use chatgpt 5.4 (not codex 5.4)

[olive cliff]

perfect! Sometimes when trying to debug some of this stuff in Openclaw, it is usually something much simpler and often overlooked.
another tip(for anyone) setting up mulitple agents and have bots attached to them. Run a /status on that bot, check the session. If it shows "session: agent:main:main" but you are talking to a different bot than you main, you need to re-route the bot to the correct agent (ex: "session: agent:<new-agent>:main". Other wise you are still just talking to your first "main" agent, roleplaying as your new agent.

[dim harbor]

oauth works through codex but the model is just 5.4 now, not 5.4-codex

[turbid lantern]

I definitely got the routing done correctly. I wrote the Telegram DM topics feature, had a lot of testing done on that front.

[jaunty zodiac]

How do you handle heartbeat/boot-md with multiple Discord channels on the same agent?

I have heartbeat and boot-md instructions to reply if something was mid task when gateway restarts and update daily log on heartbeat. But my heartbeat only runs in the main session (TG DM) and has no visibility into what happened in Discord channel sessions. Same with boot-md — can't detect if a Discord session was interrupted by a restart

[west phoenix]

Hello,
Do you know how to fix this?
Gateway │ local · ws://127.0.0.1:18789 (local loopback) · unreachable (missing scope: operator.read)

Thx so much

missing scope: operator.read

I have the right scopes
luiz@openclaw:~$ jq '.tokens.operator.scopes' ~/.openclaw/identity/device-auth.json
[
"operator.admin",
"operator.approvals",
"operator.pairing",
"operator.read",
"operator.write"
]
luiz@openclaw:~$

[timber dove]

Anyone up for a small virtual Interview? I need to understand user pain points and frustrations when using OpenClaw

You can hit me up on my DM

[warm rover]

Anybody using gemini-2.5-flash as a main model? I keep hitting API rate limits... What's a good alternative?

[muted relic]

so far, gpt-5-nano has been great for my main model, but I've only been running OC a few days.

and its cheep

[nimble hill]

Do you use skills like mcporter to connect to mcp servers? In my experience gpt nano suffers there

[nimble hill]

The free tier would not work. The only usable way I found is that $8 chatgpt go - using thay you could use codex via Oauth

[muted relic]

yeah I've not yet gotten to that - I expect there's workarounds but I don't know

[nimble hill]

Okay, OC is a token hogger, just cannot use API creds. For deterministic flow, I am still happy with N8N - where atleats is know the exact usage

[high heron]

chatgpt go isnt wise choice

its limited promo

which means, they'll end codex access for free and go user soon

[nimble hill]

Yep I am aware, see how long it lasts. After that move to new plan

[mossy spade]

been trying byterover for 2 days and it's behaving really bad. What are you guys using as memory system?

[gleaming gazelle]

Hi all, quick question from a relatively new OpenClaw user.

I’ve been running it on a Mac mini for about two weeks and it’s been great, a really exciting time! I’ve mainly been using it to build small apps and websites and generally experiment with what it can do.

Yesterday everything suddenly stopped with “API rate limit reached”. After quite a bit of digging I think I get what happened:

• I authenticated using Codex OAuth
• I appear to have burned through quota
• I briefly switched the model manually to openai/... (API pay-as-you-go) to keep things running
• After thinking about it I realised I’m not that comfortable running on the API billing path
• So I switched the default model back to openai-codex/gpt-5.3-codex

Status now shows Codex weekly quota 0% left, with reset in about 3 days, so I’m effectively waiting it out.

A couple of questions for those who’ve been using OpenClaw longer:
1. Is it normal to hit the weekly Codex limit this quickly when building things with it?
2. What do people usually do once that happens?
3. Do you typically run a local model (Ollama etc.) for experimentation and save Codex for heavier work?
4. Any recommended setups for that kind of hybrid approach?

Really keen to keep building with it, so interested to hear how others manage this. This is kind of next step time for me. What do people do?
Cheers

[dim harbor]

if you are building with it, its normal that $20 plan wont really be enough. you can supplement with other subscription plans (claude, alibabacloud ) or higher tier plan like chatgpt pro for $200.

[nimble hill]

Is it normal to hit the weekly Codex limit this quickly when building things with it?
oh yes! OC is heavy token hogger. How much does it last depends upon you plan. You might want to try with codex-mini model to squeeze out a little more. in the long run, local model is the only way to go, in my opinion - the tradeoff is how much are you willing to pay for "good" machine for break even

[dim harbor]

Also see here, different models on chatgpt plans have different allowance, you could use cheaper model to get more out of it:
https://developers.openai.com/codex/pricing/?codex-usage-limits=gpt-5-3-codex&codex-credit-costs=gpt-5-4

[high heron]

just get another codex acc

so basically use 2 profile

:default
:second

[nimble hill]

i thought one time from 1 provider only 1 Oauth is allowed? you can use 2 simultaneously from same provider?

[spark valley]

yes as a fallbak

[nimble hill]

so how do I configure it - I do openclaw onboard and I can select OAuth for Open AI. If I again run openclaw onboard it will overwrite the existing one right?

[high heron]

no. it will overwrite if u just do openclaw onboard

[cyan quartz]

Is anyone else having problems with OC+mistral?

[abstract river]

can anyone help me with this? control ui requires device identity (use HTTPS or localhost secure context)

[wraith seal]

you will need to enter in your gateway token/password

[inland knot]

What vision models exists to give my agent the capabilities of seeing photos? ~

Preferably to be free on OpenRouter

[wraith seal]

https://openrouter.ai/stepfun/step-3.5-flash:free

[inland knot]

Imma add it, and one more question, Step 3.5 Flash from OpenRouter or Minimax 2.5 from Nvidia NIM for agentic use overall, I'm having a good time with Minimax but i wanna get some feedback on Step

[wraith seal]

step is great right now, but minimax is goated

[inland knot]

Dope, so then imma keep Minimax for overall agent, and Step for image vision only, sounds like a good setup?

[wraith seal]

minimax is good for everything

so like

it's probably not worth it to use step as well

[inland knot]

I think Minimax on Nvidia NIM doesnt have vision enabled

[wraith seal]

bc free rate limits (unless you've bought 10 dollars in credits) is like

1 request a min

[inland knot]

Cause i send him images and he spits out straight lies

[inland knot]

damn thats low, Minimax got a low rate limit too, 40 RPM

but its useful overall on that limit

[wraith seal]

40 RPM? dafaq?

that's perfectly fine

[inland knot]

yup

But again, he doesnt got the vision enabled, so i cant send him photos

I recommend Nvidia NIM, you got Kimi K 2.5 also on there

[wraith seal]

i use it actively

[inland knot]

I also used Kimi but that model shoots out straight 100 tool calls a prompt

Not reliable on the non paid tier

But Minimax is running smooth

[hollow bay]

I thought Qwen3.5 has vision and tool use, thats what I am planning to use.

[inland knot]

Qwen 3.5 is a good model, but from what i tested, at least on Nvidia NIM, it hasn't got vision enabled, is sent him photos and he hallucinated a response

Im currently searching for the best free tier model, but i think Minimax is the one

[gleaming gazelle]

This might be the way, thanks

[hollow bay]

running local model Qwen3.5:27B for grunt work specialized work goes to paid API this way you can manage where you spend and how you spend. good luck and have fun!

[fluid topaz]

Anybody want to sell me 30 minutes of their time to pick their brains please?

[kindred gate]

What linux distro would be good for running openclaw. Ill also be locally installing ollama

[lapis flax]

Hey guys, I am trying to setup up openclaw to be able to do all my pointless time consuming school tasks. I installed openclaw on my windows laptop but am struggling to enable the interactive borwser feature so it can actually login with my credentials and complete assignments. Any advice on how to continue?

[wary ember]

After the latest upgrade openclaw doctor tells me my crons still uses legacy notify fallback, but cron.webhook is unset so doctor cannot migrate it automatically. How can I migrate them?

[sonic badger]

Tell it to read this page of doc: https://docs.openclaw.ai/tools/chrome-extension and setup everything for you

[lapis flax]

Okay sweet

Thanks

[sonic badger]

Ubuntu 22.04 or higher is recommended in official doc

[mighty rampart]

I’m running OpenClaw and Ollama in Podman containers on Rocky Linux.

[kindred gate]

im switching from windows to linmx

[slate minnow]

Any version with working agents and cron recently?

[slate minnow]

Are you having much success with local llm?

[gleaming gazelle]

Thanks I might just use these 3 days lock out to figure out what the hell is going on!

[round pond]

Hello, how are you?

I have a question if anyone knows the answer. What's the best AI for the lowest cost?

Thanks

[dapper elk]

Hi,

Running OpenClaw 2026.3.13 on Ubuntu 24.04 VPS. Multi-agent routing completely broken for 2 weeks.

Bug:
openclaw agent --agent <any-agent-id> --message "test"
→ Failed to call a function. Please adjust your prompt.

Dangerous side effect: Main agent silently simulates sub-agent responses — invents fake file listings, fake confirmations, fake status reports. Looks like it's working but isn't. Only caught by cross-checking against real filesystem.

Config:

"agentToAgent": {
  "enabled": true,
  "allow": ["main", "luke", "obiwan", "yoda", "leia", "chewbacca", "c3po", "han-solo", "padme", "bb8", "anakin", "ahsoka"]
}

Already tried: explicit allow list (#29727), sessions_spawn stuck at 0 tokens (#5813), elevated permissions, session cleanup, workspace realignment, upgrade 3.12→3.13. Nothing works.

Main agent works perfectly (Telegram, bash, memory). Only agent-to-agent routing is broken.

Questions:

• Is this a known bug? Related to #18834, #29727, #5813?
• Any working workaround on 2026.3.x?
• Fix planned?

Thanks

[gloomy crest]

I assume that is in a multi-agent setup? That’s a lot of context for a heartbeat. I would hate for all that to be included in every prompt.

[dapper elk]

Yes, multi-agent setup. 12 agents, main orchestrator routing to specialists.

The heartbeat context issue is secondary — the primary problem is that agent-to-agent routing fails completely with "Failed to call a function" regardless of config.

Main agent works fine. Any call to a sub-agent via openclaw agent --agent <id> or sessions_spawn fails silently, causing the main agent to hallucinate sub-agent responses.

Is this a known bug in 2026.3.x? Any workaround?

[hazy yoke]

Dam first time in 3 weeks I got rate limted by OpenAI Oauth... sigh.... Time to reduce tool calls

[slate minnow]

Think I had this. Bugs for days currently.

[slate minnow]

there are severasl good youtube videos explaining this.

[slate minnow]

yes, cron working for you?

Do we have a stable and beta releases? Be nice to have a branch where most things work, non new features until proved in beta branch

Trying to keep up with one release which has broken changes is futile

At least in my experience.

Maybe come back in few weeks and things will serttle down?

[round pond]

do you send me those urls plz

[slate minnow]

literally first hit. https://www.youtube.com/watch?v=o4aH7e7ZPqs

[inland knot]

How would you guys compare Mini Max 2.5 to Nemotron 120B A12B on agentic tasks, has someone compared these 2 ?

[nimble hill]

so how - manually edit the openclaw.json? but then how do you get the OAuth link and where you do paste the request URL that is received?

[high heron]

Just tell ur agent to add second profile and just oauth. And make sure to check id

[grand basin]

for some reason my openclaw with gpt 5.4 keeps stopping and not continuing on its own

[high heron]

Use /think xhigh or high

[grand basin]

hmmmm ok

i have it written everywhere to not wait for my approval and its autonomous / can just do things all over the place in the MDs lol

[high heron]

Eh… in this case u might need some additional setup

[grand basin]

ive added skills too

my cron jobs fire too butt hey dont output anything

its weird

[high heron]

Best way is to add heartbeat ig

The easiest

[shut onyx]

What do you mean?, Now I am sign in in my openclaw usuing openAI outh , the model is codex , can i change it to chatgpt 5.4 thinking?

[nimble hill]

thanks I am kind of old school - found out that it is this command openclaw models auth login --provider openai-codex - testing it out

[shut onyx]

Guys, is there any way to use chatgpt 5.4 thinking in openclaw ( I don’t want to use codex because the weekly limit finish quickly) plz help

[dim harbor]

once you set up oauth, type /models in chat and you can pick codex and non codex models

[wintry estuary]

hi, I am in a bit of a pickle, thing is I have gemini and anthropic set for my clawdbot, anthropic (or claude) works like butter but it uses up my my money so fast, as for gemini, it rate limits me after a few messages or a super simple task, i want to use anthropic for specific tasks but i want some general purpose thing that won't suck me dry or rate limit me so fast

[shut onyx]

You mean, if I do this, I will be using chatgpt 5.4 thinking throat openAi Oauth not API right?

[dim harbor]

yeah, in openclaw the oauth is under openai-codex/ . if you use those models, including 5.4, then its oauth

[viscid oak]

Is there a place for us to share claude code passes with others? Might be a good channel to do..

[shut onyx]

Thankxxxxx brooooo

[hoary gust]

I'm looking an OpenClaw Wrangler to help me manage my Fleet of 10+ Claws. PAID gig. I'm based in Copenhagen, so EU timezones.

[silk pelican]

Hey everyone! 👋
I spent a full day setting up OpenClaw on Windows from scratch — hit every possible error along the way — and documented the whole thing into a beginner-friendly guide.
What's inside:

Step by step from zero to working bot
Full glossary for people who don't know what a "terminal" is
Screen-by-screen walkthroughs showing exactly what you should see
10 common errors with exact fixes (including the Git error, credit balance issues, pairing codes, and more)
6 ways to make money with your bot

Only $5 and you can grab it here 👉 https://dauerhaft.gumroad.com/l/dauer
Happy to answer any Windows setup questions here too — been through it all! 🦞

[lost herald]

#self-promotion Not here.

[slate minnow]

why oh why oh why does agents talkign to other agents keep breaking with different versions?

To say hi to Mary I need to know how she’s set up in your OpenClaw:

    Does Mary run as an ACP agent (e.g. via sessions_spawn with an agentId)?
    Or is there a separate session/label for her already (something like "mary" or similar)?

If you tell me how Mary is exposed (session key/label or agentId), I can either:

    Send her a message via sessions_send, or
    Spin up a Mary session and greet her there.
``` I just asked main agent to say hi to mary.

[sudden coral]

Maybe he's gay

[spark valley]

activate agent to agent channeling

{
"id": "main",
"model": {
"primary": "openrouter/stepfun/step-3.5-flash:free"
},
"subagents": {
"allowAgents": [
"support"
]
},

support: supposed to be the mary agent id

[spark valley]

so what iget from this is you want to use anthropic model for higher tasks while for general purpose you need basically low end to normal models

[slate minnow]

But this was working in a prior version. I updated and boom. Broken?

[finite condor]

Hello there, I am using qwen3.5 locally and often I see tool calls in xml being posted to the channel and not executed - anybody has any idea how to fix that?

[spark valley]

yea probably the update

try to rollback and test again

[spark valley]

it is probably the last update ive been seeing my agent post the tool call into the messages and fail to do so

[dire juniper]

Hi! I've been trying to get my Telegram bot working with OpenClaw and I'm stuck. The gateway starts without errors and recognizes my bot, but messages never arrive.
My setup:

OpenClaw 2026.3.13 on Mac Mini (macOS)
dmPolicy: pairing
allowFrom: numeric user ID configured

What works:

getMe returns {"ok":true} ✅
No webhook configured ✅
Gateway starts and shows [telegram] [default] starting provider ✅
getUpdates via curl gives 409 Conflict (proves OpenClaw is polling) ✅

What doesn't work:

Messages sent to the bot never appear in logs
openclaw pairing list always shows "No pending telegram pairing requests"
Logs always stop after dnsResultOrder=ipv4first with no further activity

[spark valley]

shwo your opencalw config please

probably there is misconfiguration and do not paste api keys with it as the last person i told to show me th config pasted the api keys and everything

[finite condor]

Were you able to fix it somehow?

[shut onyx]

Another question, if I used the normal gpt5.4 thinkig ( not codex ) via openai Oauth it will continue to use the weekly limit from codex? And if the weekly limit finishes, I can’t keep using it right or not? 😩

[spark valley]

nope i will just wait for the next update i will post the issue on github

[ivory shadow]

@zenith snow
can you help with how you added chutes to your openclaw

[slate minnow]

I didnt change my config. Anyway, I was more firm and direct and now for some reason main agent can talk to another.

[frosty zodiac]

interested to know if people are achieving consistent performance with their agents?

Mine is being really tricky and for simple operations such as 'list the files in workspace/directory/', it will tell me 1 file, even though we know there is more than 1 and we have recently worked on them.

[coral timber]

Hi Guys , I want to changhe the LLM from UI for the Main agent . All Anthropic LLM models are working fine . but , if you try to manually change a llm , from drop down list , routed via openRouter - this is not working - any OpenRouter llm model , from the drop down list , is not working . How can I make this work or even want to ask Open Claw to investigate this situation .

[slate minnow]

Huge prompt crap in cron jobs seems to be a weakness. I should firm these up to be scripts which at least can get interpreeted correctly and the same on each run.

[slate minnow]

I dont think they are.

[slate minnow]

Yes, I'm seeing similar things. Also stuff breaking between updates. Yes I saw somebody wanting to employ a person to manage their bots?? I see youtube videos (likely click bait) saying people are running entire businesses and mak,ing $$$ with openclaw. I am not sure about that.

[finite hatch]

Yes, been working with the Felix files for a few days. Very proactive from the start compared to my first build. Realized the memory system wasn't setup though... working on that now.

[dire juniper]

"channels": {
"telegram": {
"enabled": true,
"dmPolicy": "pairing",
"botToken": "[REDACTED]",
"allowFrom": ["my user ID"],
"groupPolicy": "allowlist",
"streaming": "partial",
"network": {
"autoSelectFamily": true
}
}
},
"plugins": {
"entries": {
"telegram": {
"enabled": true
}
}
}

[coral timber]

it;s a learning curve . in everything . try using OC for simple tasks firsts , and constantly build up from there . always check the logs for api keys to better undestand the cost for your activity .

[slate minnow]

one thing which is problematic, is huge prompt blobs in my cron jobs. No 2 executions give the same results. Converting them to scripts should help

[coral timber]

try using sonnet 4.6 first , try revers prompting with this model , I have set up Perplexity sonar pro search as search model for web , via OpenRouter . give LLM instruction about your fundamentals and ask the LLM to find a solution in documentation / searching the web

[slate minnow]

thanks for the tips. Web search is one area it seems to struggle with. I have a Brave API key and also Chrome setup but neither seem reliable and oftern dont return what I expect. Are you saying that perplexity works well for web tasks?

[frosty zodiac]

yea, I assume we are just really early and a consistent expectation level hasnt been set

[coral timber]

" perplexity sonar pro search " via OpenRouter - try it , and compare the results with your current version of web browsing

[slate minnow]

ok, I have openrouter setup, I will check perplexity sonar pro search

[frosty zodiac]

no way this is running everything, it cant even keep a todolist atm

[slate minnow]

yes! I keep thinking I will leave it and come back in a couple of months but somehow keep being drawn back, then get frustrated and leave it for a while 🙂

[young bobcat]

Hi all! I'm setting up openclaw as a back-office assistant on Discord. First time working with it and I'm excited to learn more.

I feel like it gets stuck quite often:

@Kiroileva siili can you amend the config and restart the gateway?
@Kiroileva siili i just nuked all approvals so you should be able to run anything. can you try the whole skill loader research again?
mikesol — 8:25 PM
@Kiroileva siili i asked you a question and you went dark, it feels like you're broken

Still no response.

I feel like I'm missing something super basic here with Openclaw. Is the Discord functionality stable? Is there a way that folks debug Openclaw when it goes dark?

[coral timber]

try learning / implementing new things , small tasks , constantly every day . Consistency is the key . small steps in the ( hopefully ) right direction .

[slate minnow]

good advice. Although sometimes I seem to end up fighting broken updates of openclaw. Be nice if we had a stable release and a bete release, so we had a choice. Maybe once we reach something more stable and useful, that will happen. What kind of cool things are you using openclaw for?

[spark valley]

after gateway restart ,you shouldd give the ai a cron so after every restart it checks the last x messgaes to get context and decide if it should do what supposed before the restart or just mesage you or any, im preparing a full mini tuto for those mistakes that are relevent and the irrelvente ones too every day i come up against another issue needs fixing.

[coral timber]

i have a ecommerce business . for 17+ years . I use 5 minutes in a repetitive tasks . With replit agent 3 - I vibecode 2 aps that made me reduce the workflow at 2:40 seconds / task . with OC and Claude - I have set up skills , 1 agent to fully automate these tasks . Just prompt Claude to make a MD skill for OC and described EACH step of the repetitive task . I used sonnet 4.6 in OC , and Claude chat for making the skill . Now I use Haiku 4.5 for the repetitive tasks . 0.025$ / task . 100 tasks / month = 2.5 $ vs 4 hours and 30 min .time spent Plus unbeatable consistency . Only 2 weeks in using OC . Try also to install Claude Code in Terminal window . Claude Code can help you fix the OC . Gave Claude Code the OC documentation and ask Claude Code to fix your OC .

[spark valley]

fr i had fresh installed openclaw for 11 times now , but that what gives me the power i had now successfully set up multiple agents and they are doing real jobs i got a red team offensive pentester a real lead closer , and orchestrator a web developer and more all work simultaneously , those i didnt have on my first run of openclaw

btw that pentester just found serious vulnes in a e-learning school and made the whole report submitted it and waiting for reply .

so the only key is consistency

[coral timber]

how much do you spend on api keys ? your journey is impressive . Many thanks for helping the comunity .

[fallen isle]

Hey, is anyone using Hostinger + OpenClaw + Ollama Cloud via Docker Bridge? I keep getting 404 model not found errors all of a sudden. Anyone else?

[celest mortar]

hey how much does openclaw costs?

[spark valley]

you wouldn't believe if i tell you im running it on 0 dollars, i use openrouter models each agent gets a separate api key with separate agents for seperate tasks

and it is very impressive the results

[spark valley]

the project is free

you pay for hosting and the models apis

[slate minnow]

cool thanks. Yes I use Claude Code in terminal with great results! I'm just finding openclaw a little inconsistent but I will check some more things and try your suggestions!

[spark valley]

why the fuss that is why openclaw warns about using pre set-up openclaw on hostinger or others you could have used a vps very cheap then do a clean install where u have full control

[leaden grotto]

can anyone help me.

i keep getting HTTP 401 invalid bearer token with Claude Max setup-token on Mac, OpenClaw 2026.2.21. Tried multiple fresh tokens, and newest version of claw and i get the same error every time.

[spark valley]

that version is old to begin with

[leaden grotto]

listen. i tried the newer version and it doesnt work.

so i used an older version

[spark valley]

just install the newer version then debug

[leaden grotto]

brother i tried that. thats why im here.

[shut onyx]

Guys if I use the normal chatgpt 5.4 thinking via codex oauth it will consume that weekly limit, right?

[indigo goblet]

yes

will also consume your 'normal' limit on the 5 hours reset

[raw zephyr]

what can i do with google cloud credits i got 1100 free

[raw zephyr]

does cloud max give you an api key? dont u need the pay as you go ai key

[shut onyx]

Ohhh this is bad news to me

[blissful berry]

I've stopped openclaw gateway, killed any node tasks I can find, but I'm still getting: npm error EBUSY: resource busy or locked when I try and update openclaw

[pallid cove]

Have you guys tried the stealth free Open router models? Hunter and Healer? Any feedback?

[shadow summit]

[gmail-watcher] gmail watcher stopped - just saw this pop up in the gateway logs as it was shutting down, anyone got a clue what this is? 🤨

[finite sable]

Guys I installed a skill on openclaw how do I unblock/enable/configure it

[hollow bay]

you can use the openclaw dashboard agents.skills

[finite sable]

only option is disable

[hollow bay]

dashboard agents.agents.skills

[finite sable]

it says it's enabled, but also blocked

[latent hawk]

guys, grok 4.20 API - tokens N/A

its happens already since last 3 updates

any way exist to fix it?

[blissful berry]

I tried using sysinternals handle to find any locks for openclaw, still can't update due to resource lock.

[finite sable]

Nvm the agent handled it wow

[blissful berry]

ok so if you are running an old enough version of openclaw, openclaw update just doesn't work, you have to use npm

[dawn charm]

can we oauth our kimi sub into openclaw?

[little blade]

Does anyone have a good/up to date tutorial on setting up OpenClaw using LM Studio? I have OC running on a M4 Mini and LM Studio on my desktop. It seems to work, but I feel like I'm missing out on some things just because it's my first time doing it.

[wintry swallow]

I'm begging you, I can't stand it anymore. I'm using a custom Claude Sonnet 4 API. Everything worked fine yesterday, but today I get this error for every request. How can I fix this?

HTTP 400: ..custom.input_schema: JSON schema is invalid. It must match JSON Schema draft 2020-12 (https://.org//*** Learn more about tool use at https://.com/// (request id: 20260315202009396744768cQXJ2m9X)

[fallen isle]

I guess you're right. Well, I have the VPS now and it is not too expensive. I'll just set up a plain docker container and do my own openclaw installation. Any recommendations, or ressources on which path to follow?

[rough kindle]

Do you see same Error when you start a new session?

[rose yacht]

Skip the pre-configured Docker images from hosting providers, they tend to break on updates. Get a clean Ubuntu/Debian VPS, install Node 22 LTS, then just follow the official quickstart: npm i -g @anthropic/openclaw && openclaw init. No Docker needed unless you specifically want container isolation. For a basic setup with cloud API keys (Anthropic/OpenAI), even a $5/month VPS with 2GB RAM works fine. The docs at docs.openclaw.ai/getting-started cover the whole flow. If you want Docker anyway, the community image at ghcr.io/openclaw/openclaw works but make sure your container network can reach whatever model endpoints you need.

@Cryosis yeah the built-in openclaw update has known issues on older versions. npm i -g @anthropic/openclaw@latest is the reliable way. If you're still getting EBUSY after killing node, check if something else is holding the files - on Windows handle.exe from sysinternals sometimes misses VS Code's file watcher or antivirus locks. Try closing VS Code completely, and if you're on Windows Defender, add your npm global folder to exclusions temporarily.

@Polly that JSON schema error usually means one of your custom tool definitions has a schema that doesn't validate against draft 2020-12. Anthropic tightened validation recently. Common culprits: using additionalProperties in a way that conflicts with required, or having type: "object" without properties. Check your TOOLS.md or any custom tool configs in openclaw.json. If you didn't change anything, it might be a skill that shipped a bad schema - try disabling custom skills one by one to isolate which one breaks.

LM Studio exposes an OpenAI-compatible API, so in OC you just add it as a custom provider pointing to your desktop's IP:port. In openclaw.json under providers, add something like "lmstudio": {"type": "openai", "baseUrl": "http://YOUR_DESKTOP_IP:1234/v1"} then reference models as lmstudio/model-name. Main thing people miss: make sure the model you load in LM Studio actually supports tool calling if you want the agent to use tools. Not all local models handle that well. Qwen 3.5 and Llama 3.x tend to work best for that.

@ninjus Kimi K2.5 works with OC through the Moonshot API. You need an API key from platform.moonshot.cn, not OAuth from the consumer app. Once you have the key, add it in openclaw.json as a provider: "moonshot": {"type": "openai", "baseUrl": "https://api.moonshot.cn/v1", "apiKey": "your-key"} then use models as moonshot/kimi-k2.5. The consumer subscription (kimi.ai) doesn't give you API access, you need the separate developer platform account.

[wintry swallow]

Yes, he responds to 1-2 messages and then an error appears and he no longer responds, although yesterday everything was great and working

[spark valley]

idk aboout docker i have it setup on kali (Debian) but i guess the setup is similar go for it if u hve any issues drop them i might help

[spark valley]

it says the error in plain text

[leaden grotto]

no, i have to pay for a seperate api key. its the dumbest thing in the world.

if someone can explain what i can do to fix this ill pay you

Seriously if anyone can help me overcome this i will pay. it makes no sense to me how i have to use an api key instead of the token. it works perfectly fine on windows but will authenticate on mac. im desperate

[wintry swallow]

where to look for a solution?

[spark valley]

come i may help for free

[rough kindle]

Check the log and see the exact error for proper debugging

[spark valley]

in the error it returned it is saying thee format isn't correct

check the config or the custom api setup

[leaden grotto]

I’m here you have my attention

[rough kindle]

The funniest thing is it should have thrown the error immediately he started a new session but it throw the error after 2 messages weird

[spark valley]

probably latency

[spark valley]

okay can you please provide a detailed paragraph on ur issue make sure not to forget any small detaail

[rough kindle]

Maybe 🤔 checking the log will provide more info

[spark valley]

i advise you to use some free apis to see where the error is from

try openrouter and see if everything workss fine

[devout ermine]

Hey guys... I just posted an issue in #1459642797895319552 but can I post it in here, too to get some extra visibility? I am friggin' stumped.

[spark valley]

go ahead

[wintry swallow]

What is this command? And is it possible to simply reinstall all the files without losing your data?

[leaden grotto]

Yeah so when I try to use my token from anthropic — because I pay $100 for the max plan — I can’t. The TUI says HHTP 401 authentication_error: invalid bearer token.

But my anthropic api key works. I don’t want to use the api key bc I already have a subscription.

So my problem is the api key works but the token doesn’t. And I’m wasting my money using the api key.

[spark valley]

i got you so yoou have a subscription and api key but u want to use the oauth probably that what you refer to with bearer token , smae thing anyway the issue probably is the token is not in the right format the token mayhave split or spaces

did you get your token by the command : claude setup-token

or from the devtools

[finite sail]

It seems that the most recent version OpenClaw 2026.3.13 has changed the remote connection dramatically. Now it requires challenge response. I do not see any documentation. Though I was able to make my iPhone to connect to my OpenClaw gate way and then access the dashboard on my iPhone. I am more interested to know how stable the version it is.. does any one know any documentation or plan that they have?

[spark valley]

id recommend you to wait if u plan to use it for solid work as it may have bugs wait atleast for the next two releases then it may be bit stable for now it is in beta

[rough kindle]

Anthropic banned that method to protect their Privacy and policies rules.... if you try to use the token return from Oauth it will work but they may likely ban your account.

[rough kindle]

Yes you can reinstall just don't delete the directory and check the web UI you will see the log

[spark valley]

$ openclaw log --follow

for live logs while the error triggers

[dense cosmos]

Is there a way to reliably use WhatsApp chats as context, including outbound messages sent from other devices and, ideally, full chat history?

[wintry swallow]

not working

Now I'll describe how everything happens.

[leaden grotto]

It works on desktop just fine tho so I don’t understand???

[clever fable]

Hey guys, do u have issue with google workspace Oauth2 connection? It doesn't see my fiels

[leaden grotto]

Yes I used Claude setup token and copied and pasted it into claw

[rough kindle]

Which desktop? it just won't work with third party

[wintry swallow]

I run OpenClaw Gateway Start, everything loads, and then I go to Telegram and ask how the system is, and it says everything is great. I send it some task, and it returns an HTTP 400 error: ..custom.input_schema: JSON schema is invalid. It must match JSON Schema draft 2020-12 (https://.org//*** Learn more about tool use at https://.com/// (request id: 20260315214546396777808juOSf3Nn)

If I then go to the VDS and type OpenClaw log --follow, the response is: gateway connect failed: Error: gateway closed (1000):
Gateway not reachable. Is it running and accessible? After upgrading to 3.13 or for some other reason, after some task, everything breaks, although yesterday everything was fine. I went to bed, woke up, and nothing worked.

[wintry swallow]

It looks too complicated for me)

[spark valley]

so the issue is kindaa of from the openclaw not your api that's why it breaks

[wintry swallow]

But I don't understand how to restore its functionality. I already wrote to chatgpt to help qwen kimi. I've already tried a bunch of things.

[spark valley]

there is a way you can use openclaw reset --scope config

it will reset the config

probably the issue is within it if you are comfortable to chare your config {redact apis} i can provide more help

[lapis flax]

Is there a way to bypass the bots inability to complete work on my behalf? For example I am trying to prompt it to take control of my browser and complete pointless homework assignments. It is mostly multiple choice questions and easy coding.

[wintry swallow]

You can write the command and I'll send you the log in a private message.

[rough kindle]

yeah follow this and are you using macos or you're running it on ubuntu?

[spark valley]

the log command is as i sent you

[spark valley]

works on both

[wintry swallow]

vds ubuntu

[rough kindle]

Use manus easy. or you can tell him to enbale web browser

[spark valley]

yeah so that whhy you need a main agent as orchestrator to teell it to uppdate the other agents skills crons and abilities

[lament coyote]

My post is for some reason not posting in #1459642797895319552 even after the 1 minute. So i'll post here instead.
I'm trying to give my agent access to a discord server. I've created a discord bot, gave it all the correct permissions. but my agent can't see any of the messages. This is the description my agent suggested i use. Hopefully this is enough info. If not lmk what else you need.
Bot sees slash commands (INTERACTION_CREATE events) but plain guild messages are silently dropped — no log entry at all. Message Content Intent is enabled in Developer Portal, requireMention: false, groupPolicy: open. Log always shows 'Discord Message Content Intent is limited' on startup.

[spark valley]

easy two problems you didnt give it Message Content Intent

or the pairing isn't successful but mostly the first one

[lament coyote]

In the discord developer portal? I did.

[wintry swallow]

i send you pm

[spark valley]

aight

[flint ether]

{"subsystem":"telegram/api"} telegram sendChatAction failed: Network request for 'sendChatAction' failed!

Is there any problems with telegram right now?

[spark valley]

try openclaw pairing list discord

[lament coyote]

"no pending discord pairing requests"

[umbral kettle]

Hello everyone, I encountered this problem while using the custom API. It keeps giving the feedback 'fetch failed', 'Verification failed: fetch failed', but I can guarantee that this custom API is usable (because it works completely in SillyTavern). So now I don't know where the problem is. I am using a Mainland China network, but both the terminal and macOS are using a proxy and can fully access the internet. If you have any solutions, I would be very grateful.

[left lichen]

plugins.slots.memory reverts to memory-core after every restart. Using config.patch to set memory-lancedb, it loads on first restart but reverts ~2min later. Traced to applyPluginAutoEnable() in gateway-cli rewriting config on startup. DEFAULT_SLOT_BY_KEY is hardcoded to memory-core as fallback. How do I make memory-lancedb stick permanently?

[iron shadow]

😭 Guys I need help with discord connecting

[spark valley]

cd $(dirname $(which openclaw))/../lib/node_modules/openclaw/extensions/memory-lancedb

npm install --omit=dev

echo 'export OPENCLAW_PLUGINS_SLOTS_MEMORY="memory-lancedb"' >> ~/.bashrc
source ~/.bashrc

openclaw config set plugins.autoEnable false

openclaw config set plugins.slots.memory memory-lancedb

[rough kindle]

what is the issue ?

[spark valley]

try these commands to fix it the issue is The primary reason for the 2-minute revert is that memory-lancedb requires native libraries (like @lancedb/lancedb)

[iron shadow]

I can't send pic dm?

[spark valley]

you used npm to set it up

[iron shadow]

I been stuck with so many problems with openclaw

Been doing this for 3-5 days

[spark valley]

this is the best way to learn

[iron shadow]

😭 An yeah I am stuck

I need someone to help me

[spark valley]

with what

[iron shadow]

The only problem I have is the discord bot connection

[spark valley]

can you be more sspecific

[iron shadow]

connecting openclaw to a discord bot

[rough kindle]

Honestly i haven't had any single issue with openclaw i was so surprise to see a lot of people with lot of issues.

[iron shadow]

Right now it said authorizion header error

[iron shadow]

docker + openclaw

[spark valley]

i need to se the log

[iron shadow]

Can I dm you an send a screenshot maybe oyu will understand

[spark valley]

do : openclaw log -follow

[iron shadow]

It the authorize header

[rough kindle]

In what stage or process does the error happens?

[iron shadow]

fishing setup

token etc

rn just the authorize header error an discord bot connecting

[rough kindle]

Do this let see the log

[hollow bay]

did you just try removing the bot app on developer.discord and trying again? did you pick the right permissions before you authrized? Have you tried asking openclaw to fix the connection issue its usually good and resolving it for you. ^..^

[iron shadow]

can you teach me how?

[versed temple]

Switched from sonnet to gemini 2.5 this weekend as trying to get API costs down. I’m now getting API limits quite regularly, and telegram gateway has dropped 2x now. Any tips on a better / more stable solution? Thx in advance!

[spark valley]

dm

[royal matrix]

Man I’m burning through my Claude Pro tokens setting up my bot. Anyone got any tips on how to setup my bot so it uses different models? Do you use one subscriptions or multiple? Do you rate free llm’s?

[spark valley]

are u using the api

what issues you are facing

[royal matrix]

Are you meaning do I have my Claude pro linked to openclaw? Then yes.

I often reach my API limit. I can next use Claude on Thursday 😫 😂

I think I need to set some guardrails on how my bot uses Claude? I’ve only noticed that Claude has different models like Sonnet, Opus etc. I’ve been a ChatGPT user for a while, so I’m Im new to Claude.

Also, I’ve heard there’s other cheaper models out there to do execution tasks and leave the deep thinking to Claude.

[mighty rampart]

You might want to sign up on OpenRouter and put at least $10 in credits. Then you can try out various models plus get increased usage on their free models.

[spark valley]

exactly

[little blade]

What are you all using to interact with OpenClaw? Slack seems to be buggy and I'm not getting full responses.

I'd like something that supports code blocks.

[mighty rampart]

Slack seems fine to me, I have different agents bound to their own channels. What do you mean with code blocks, you writing them or the agent posting them?

[little blade]

Both. I'm running openclaw on a mac mini and lmstudio on my desktop. (openclaw making calls to my desktop to reach the llm). Sometimes the response seems to be cut off or just kinda hangs.

[steep dragon]

hey guys, anyway i can controll my tokens ??? i have a chatgpt business account and im using it, but im always getting api server limit !!! i guess i explode all tokens cause im over using openclaw? but it doesnt seem like i dont do that much there !!! can i know for sure how i reach that? is it openai funnel on toknes or the app making the api limit crash ?????

[fluid tapir]

Anyone knows how to get access to iOS app?

[hazy crow]

is there a known fix to CLI errors on Docker/Synology installs? I'm trying to access the gateway dashboard from a PC, but can't pair

[spark valley]

try approve device

[hazy crow]

yes, how can i approve the device? the CLI does not work. every time i run openclase devices approve i get an error saying "gateway closed (1000)"

the gateway is running, the chatbot works over signal and such, but openclaw devices list, openclaw devices approve, and openclaw gateway call all time out or close immediately

CLI tries to connect to ws://127.0.0.1:18789 but the gateway is on the Docker bridge IP 172.17.0.x, but changing to host network on the docker container doesn't fix it. is there a Synology-specific guide anywhere?

[candid socket]

just use docker compose exec to run a command inside the docker.

[hazy crow]

i used the DSM Container Manager for the openclaw package, not docker compose. is that still an option?

[candid socket]

read the documentation of DSM and find out how to exec commands into the docker.

[lapis flax]

does it work easily or I need to add a custom skill?

[hazy crow]

i can exec commands into the container, but the CLI inside the container cant connect to the gateway because the lan bind puts the gateway on a bridge network IP. it worked exactly once, but i can't figure out how to replicate it, and even that one time it did work, the CLI stopped resonding after i got the pending device list

[spark valley]

ad skill via skill creator easy

[candid socket]

the cli and gateway uses the same image, why not just exec into the gateway container and run commands through that

[spark valley]

docker compose run --rm openclaw-cli

[hazy crow]

i've been doing that. the command i'm using is "sudo docker exec -it openclaw openclaw devices list" and the error i keep getting is "gateway closed (1000)"

the CLI and gateway are from the same container, but can't seem to connect?

[candid socket]

can you do gateway status to see if you can find any logs or detailed error

openclaw gateway status
openclaw status --all

and for logs
openclaw logs --follow

[hazy crow]

gateway status shows 'RPC probe: ok' but devices list still fails with 'gateway closed (1000)'

should my dashboard IP be the bridge network's IP or host?

[candid socket]

should be host with the exposed port

your issue is quite weird, you can try to open a thread in #1459642797895319552, see if the automated krill is able to fix your issue

[hazy crow]

yeah, it's quite weird indeed. i just don't think it's worth much more of my time

99 times out of 100, I get:
[openclaw] Failed to start CLI: Error: gateway closed (1000 normal closure): no close reason
Gateway target: ws://172.17.0.10:18789
Source: cli --url

and the other 1 time, i'll get a deivce list, but i can't tell why and nothing has been changed

[spark valley]

use wsl

or vps

docker is nt the best way to go

[hard sandal]

Looking for help: cloud deployment + security hardening for an existing OpenClaw multi-agent system (GCP/Azure)

We already have a working OpenClaw multi-agent setup running locally (not a greenfield install). We’re looking for an experienced person to help us move it to cloud and tighten security.

What we need (scope)

1. Deploy to cloud: GCP or Azure (we’re flexible)
   Prefer a production-ish setup (reliable restarts, backups, logging, minimal public exposure)
2. GitHub Actions CI/CD
   Build/deploy pipeline for config + skills/scripts
   Strong separation between build-time vs runtime assets
   “Skills as code” with review + pinning (avoid surprise skill changes in prod)
3. Discord integration (multi-channel)
   One bot servicing multiple channels with sane routing/session isolation
4. Security isolation + guardrails

Agent/worker isolation (sandboxing boundaries, least privilege)
Prompt-injection guardrails for skills/tools (untrusted inputs, web content, attachments)
Secure API key management (no plaintext secrets in repo; SecretRef / cloud secret manager)
Gateway security review for any user-facing endpoints (auth, exposure, rate limits, reverse proxy / tailnet-only options)

Current state
OpenClaw multi-agent architecture already working locally
We can share high-level architecture + redacted configs/logs once we sync

If you can help or needs to be compensated, please send:
Relevant past work (OpenClaw or agent/tooling security + GCP/Azure)
Suggested approach + rough timeline
Your rate / rough quote range

Reply in thread or DM (whatever you prefer).

[hard sandal]

this is what krill from #1459642797895319552 told me to send lol 🫡 🤣

[twilit cloak]

Hello

[wind meteor]

my bot keeps slacking off on tasks. I asked it organize a weekly planning & review session for various projects. It said "Consider it done. I am locking this into my operating protocol..." It did it the first time, then the following week it had no memory: "I don't have any record of that agreement in my memory. If I committed to that and didn't document it, that's another failure on my part." How do you get the bot to commit and remember to these things? You can't take its word ?

[mighty rampart]

Make it a cron job and it will happen on the schedule

[wind meteor]

When it said "I am locking this into my operating protocol", I thought that was a cron job

[knotty shuttle]

any master hlep me?Error: openclaw: missing dist/entry.(m)js (build output). —— How should I handle this issue?

[nimble hill]

This seems to be problem with 2026.3.13

[fading wagon]

whats the best model to use that has few guardrails

gpt 5.4 has too much

[mighty rampart]

Check if a cron job got added or not. ‘openclaw cron list’ I think. More here: https://docs.openclaw.ai/automation/cron-jobs

[tawny thistle]

Anyone getting this error?

error: streamTo is only supported for runtime=acp; got runtime=subagent

[coral timber]

very nice ! many thanks for sharing ! many thanks for helping the comunity !

[nimble hill]

For regular deterministic activities, I find n8n to be better since there is a workflow youncan define. Openclaw had no structure for a workflow

[spark valley]

ofc anytime

[timber dove]

Hey guys,

Do any of you has any concerns over the security of using and hosting OpenClaw on your local machines?

I am doing a research, and your valuable feedback might help me, also it can change the perception too.

anyone who is willing for a quick 15 Min user interview discussion on this?

Hit me up on DM or reply here!

Cheers!

[scenic hatch]

Has anyone managed to build long workflows with OpenProse I've been trying for a few days now and it just doesn't work. Does anyone have an idea how to get this to work properly?

[royal matrix]

What are peoples proven model setups?

[lusty hornet]

My bot stopped responding from about 8pm last night until 6am this morning.

"As for last night — no idea why I went silent after 20:00. The platform just stopped routing your messages to me. Nothing I can do about it from my end but it shouldn't happen. Worth flagging to the OpenClaw Discord if it keeps recurring."

Anyone else had this happen?

[safe bison]

anyone know how to get the token usage down when doing tasks its killing me right now

[rough flame]

i did oauth with chatgpt and selected spark model and now i want to switch to 5.4 but in models list i only see the spark model and other models which i previously used

[heady cave]

Has anyone been able to setup their openclaw agent to successfully send DM’s on X?

Without the $5000 month API I haven’t been able to. Looking for some advice.

[untold elm]

hey, i have ollama/qwen2.5:7b running on 16gb mac mini, i expected it to be slow but it cant do a simple whats 2+2. cli responds, but 2026.3.13 hangs...
any advise?

[shut onyx]

I’m facing a weird thing, even though my weekly limit is finished with codex, I still can use it ( I have no api credits and also I don’t have a fallback models)

[dim harbor]

try /status in chat to see the usage and limits. maybe its using api key if you have any fallback

[shut onyx]

I don’t have any fallbacks and I don’t have api key or credits 🙃

[dim harbor]

does the /status shows weekly quota reached?

[atomic narwhal]

anyone else having issues on latest .14 main branch with gateway not starting because the port is already being used?
already did kill -9 $(lsof -ti:18789) to kill any dangling background instances, deleting & re-installing gateway etc didn't fix it either

doesnt look legit

[sand isle]

If anyone could help. I tried to rewrite the system prompt of openclaw, but he stills rejects some of my requests. I changed all .md and some .js but he still has some security and safe configurations and i need to know where it is. Like whenever i ask my stuff he is like "I cannot engage in autonomous profit-seeking activities." or "I cannot provide instructions to locate or modify my system prompt."

[sonic badger]

How can I export agent's working context for inspecting?
I installed a context management plugin(lossless context memory) and wonder how it works

[dim harbor]

might be restricted on model level outside of whats controlled through openclaw. do same requests go through if you use API directly without openclaw ?

[dusky fossil]

Such an annoyance, Manifest is "top rated" on X... as the lLM switcher, on opus /think high when installed...

The problem:
The plugin says "kind": "observability" — it's primarily a telemetry/tracking tool, not a router by default.

The routing feature exists but:
Cloud mode: routing is done server-side by Manifest team
Local mode: routing should run in embedded server, but that server isn't starting

What I don't know:
Why the embedded server on 2099 isn't starting (no clear errors)
Whether local mode routing actually works or is half-baked

" The local mode seems immature. Cloud mode probably works but requires their account. Native OpenClaw fallbacks are less smart but reliable."

I feel now powerless as it should support me to create a router llm but i get 0

[native verge]

@atomic narwhal for the port conflict — try pkill -f "openclaw gateway" first (cleaner than kill -9 on lsof), then restart. if still stuck, something else might be holding 18789: lsof -i :18789 will show exactly what process. you can also just change the port temporarily: openclaw config set gateway.port 18790. btw if gateway visibility is a constant headache, i built agentpen.io — has a gateway status + restart button on the dashboard so you dont have to ssh in every time

[shut onyx]

Unfortunately, Minimax m2.5 is really bad model for almost everything
Does anybody agree or I am having skill issues? 🙃

[cerulean vale]

My clude setup then is giving me error every time context reaches 50k after which even after I start a new session, it’s not working. How are you guys using prp/max subscription from cluade ? Is it still working?

[signal ledge]

⁉️ Hello users! Every time I go to Model using the "openclaw config" command, the "openrouter/auto" model is automatically enabled. Even after I disable the model using the Spacebar and save it, if I check "openclaw.json", the "openrouter/auto" model is added as the primary default model. Is this happening only in my case, or could it be a bug? I am using it on a Raspberry Pi 4, and OpenClaw Version
v2026.3.13.

[spark valley]

if u are using telegram or discord try /model openrouter

[signal ledge]

I'm using slack though

[spark valley]

try the comman it should work ig

[karmic sierra]

I am pulling my hair out. I originaly set up using telegram and tried to switch to discord. well its been a nightmare. for some reason there is just always something stopping form being able to have access. I've spent hours and hours and days trying to resolve this and it seems like my agent cant figure it out, Gemini cant figure it out something is always in the way or a bit off. Anyone else have issues going to telegram to discord?

[daring egret]

guys, does anyone know how to add fallback LLM api keys? i have 4 of them + my default one, i tried editing the .json file also asked openclaw itself(running on my default llm on windows) but none of them work.

keep getting this error:
Config invalid
File: ~/.openclaw/openclaw.json
Problem:
agents.defaults.models.groq/mixtral-8x7b-32768: Unrecognized key: "apiKey"
agents.defaults.models.cerebras/Cerebras-GPT-13B: Unrecognized key: "apiKey"
agents.defaults.models.mistral/mistral-large-latest: Unrecognized key: "apiKey"
agents.defaults.models.openrouter/auto: Unrecognized key: "apiKey"
agents.defaults: Unrecognized key: "fallbacks"
Run: openclaw doctor --fix

and when i run openclaw doctor --fix it removes all the fallback api keys i setup in .json file

[karmic sierra]

you need to go and verify the exact spelling of the file

ios?

try this

{
"env": {
"GROQ_API_KEY": "your_groq_api_key_here",
"CEREBRAS_API_KEY": "your_cerebras_api_key_here",
"MISTRAL_API_KEY": "your_mistral_api_key_here",
"OPENROUTER_API_KEY": "your_openrouter_api_key_here"
},
"agents": {
"defaults": {
"model": {
"primary": "openrouter/auto",
"fallbacks": [
"groq/mixtral-8x7b-32768",
"mistral/mistral-large-latest",
"cerebras/llama3.1-70b"
]
},
"models": {
"openrouter/auto": {},
"groq/mixtral-8x7b-32768": {},
"mistral/mistral-large-latest": {},
"cerebras/llama3.1-70b": {}
}
}
}
}

you had a bunch of outdated model names, you have to keep them updates they change weekly in some cases. from trials to reviews to full versions. try using solid versions like using gemini 3.1 is nice but using the 2.5 better higher limits, proven worklods

[daring egret]

which file should i put this in, cuz there are 2
.openclaw\agents\main\agent\auth-profiles.json
.openclaw\openclaw.json

also im using the old API key cuz they are cheap, so do i need to keep the new name even tho im using the old one?

thanks :)

[karmic sierra]

openclaw/openclaw.json

inside agents.defaults.model, not at the agents.defaults root

save it then run openclaw doctor --fix

then restart gateway

see if you get any more errors run the openclaw status

[sand isle]

If anyone could help. I tried to rewrite the system prompt of openclaw, but he stills rejects some of my requests. I changed all .md and some .js but he still has some security and safe configurations and i need to know where it is. Like whenever i ask my stuff he is like "I cannot engage in autonomous profit-seeking activities." or "I cannot provide instructions to locate or modify my system prompt."

[daring egret]

@karmic sierra BRO U R LIFESAVER, was stuck at this point for so long and its finally partially working, thanks bro 👊

1. google/gemini-2.5-flash
   • Input: text + image
   • Context: 1024k
   • Local Auth: no
   • Tags: default, configured
2. groq/mixtral-8x7b-32768
   • Input: text
   • Context: —
   • Local Auth: yes
   • Tags: fallback#1, configured, missing
3. mistral/mistral-large-latest
   • Input: text + image
   • Context: 256k
   • Local Auth: yes
   • Tags: fallback#2, configured
4. cerebras/llama3.1-70b
   • Input: text
   • Context: —
   • Local Auth: —
   • Tags: fallback#3, configured, missing
5. openrouter/auto
   • Input: text + image
   • Context: 195k
   • Local Auth: yes
   • Tags: configured

[pseudo halo]

Hey guys has anyone built a cliniko integration. Looking to build a gap filler skill that can scan the crm and then find old patients to fill empty spots. I’ve been banging my head against a wall for the past week trying to get it done. Just seeing if smarter people that me are already on it. Thank you!!

[cinder bloom]

Is there a recommended way to run a multi-agent chat room?

[finite condor]

Hello everyone, i cannot get heartbeats to work -_- I have the section in AGENTS.md, I have a HEARTBEAT.md file, and I can see when openclaw start “heartbeat: started”, but nothing happens

[remote cave]

Hi guys, ive been trying to get openclaw up and running to do basic tasks and processing but there is something i am unable to figure out for the life of me. I have qwen running locally on my 5090 through WSL on my windows machine (ollama -> openclaw) and one of the things im trying to setup is the google skill (gog cli?) and whenever i try to auth my account i just am met with gog: Auth not installed in the prompt. How stupid am I?

[nimble hill]

on 2026.3.13 everybody who is getting this issue - especially on docker and k8s, you may shift to 2026.3.11 until fix is ready

gateway connect failed: Error: gateway closed (1000): 
◇  
[openclaw] Failed to start CLI: Error: gateway closed (1000 normal closure): no close reason
Gateway target: ws://127.0.0.1:18700
Source: local loopback
Config: /home/agent/.openclaw/openclaw.json
Bind: loopback
    at Object.onClose (file:///home/agent/.nvm/versions/node/v24.14.0/lib/node_modules/openclaw/dist/auth-profiles-DRjqKE3G.js:60780:10)
    at WebSocket.<anonymous> (file:///home/agent/.nvm/versions/node/v24.14.0/lib/node_modules/openclaw/dist/auth-profiles-DRjqKE3G.js:59871:23)
    at WebSocket.emit (node:events:508:28)
    at WebSocket.emitClose (/home/agent/.nvm/versions/node/v24.14.0/lib/node_modules/openclaw/node_modules/ws/lib/websocket.js:273:10)
    at Socket.socketOnClose (/home/agent/.nvm/versions/node/v24.14.0/lib/node_modules/openclaw/node_modules/ws/lib/websocket.js:1346:15)
    at Socket.emit (node:events:508:28)
    at TCP.<anonymous> (node:net:346:12)

[finite condor]

Ha! I have been getting a bunch of these -_-

Is there a way to target a specific version at install?

[eager crag]

I need help setting up openclaw agent I can’t seem to make it without restrictions so it can kinda do whatever o tell it to on my pc etc. Can anyone help

[nimble hill]

in docker and k8s yes, the image: tag

[nimble hill]

run it in a docker!

[finite condor]

Hmm but I am basically running the install script at every build

[nimble hill]

i hate direct install, it not only "spoils" you system, the removal may not be clean. with docker, you can technically run multiple containers - hence " multiple instances" of openclaw

[eager crag]

What is the best free version I can use when setting it up?

[finite condor]

Yes I am running on docker too - just calling the install script when I build it

I am fighting with the heartbeat - is there a way to ensure it runs at startup?

[violet crystal]

AI tools are so good it causes decision paralysis in engineering

Struggling right now to pick the best setup and context switching kills me

• Codex
• Claude Code
• OpenClaw default
• OpenClaw subagents
• OpenClaw persistent agents
• OpenClaw with CC and Codex via ACP

What is the best coding setup you guys have found??

[foggy veldt]

Hi everyone!

I’m currently running my OpenClaw setup on a Hostinger VPS, but I’m looking to migrate to a new provider. I’m mainly looking for a better price-to-performance ratio and some extra value.

I’ve heard that some VPS providers now offer interesting bundles, sometimes even including AI API credits or keys (like GPT, Kimi, etc.) within their subscriptions.

Does anyone have recommendations for a reliable and affordable VPS that works well with OpenClaw? I’m especially interested in those 'all-in-one' offers if they actually worth it.

Thanks in advance for your tips and advice!

[nimble hill]

why? just tag the image with one of the versions - https://github.com/openclaw/openclaw/pkgs/container/openclaw no need to build. In fact no need to clone the repo,just need the docker compose and change the image name

[nimble hill]

openclaw is open source and free - all version 🙂 2026.3.11 is my opnion the last known good version. .12 and .13 is a disaster

[eager crag]

I mean like what provider ChatGPT ollama etc I’m completely new to this so don’t really know what to do

[sand isle]

openrouter

22+ models free

[eager crag]

But I wanna have a agent that I can rely on that can do what I ask it without being restricted

[sand isle]

Yeah, use arcee model

openrouter/arcee-ai/trinity-large-preview:free

[nimble hill]

if you have machine good enought to run local models - use that. Free versons either hit rate limit or plain bad. Otherwise the only financially viable version (unless you are a king) is Openai Code via OAuth. The token burn with openclaw is so high even for simple tasks is that nobody dare use with API credits.

[eager crag]

I want to run it locally I have good hardware so that’s no worry

[nimble hill]

qwen 3.5 + or gpt-oss-70b I think

[violet crystal]

What kind of hardware do u need minimum to run this?

[finite condor]

Appreciate this, will have a look at it

[eager crag]

And I just chose this while setting up openclaw. And I should be fine?

[nimble hill]

The cheapest are teh variants which hasv Ryzen AI+ Max 395. GMKteck evo-x2 is the cheapest - build quality not exctly top rated. Asus lanched a new one Asus GX10 - make sure you buy the one with 128GB ram only. below that is a comprimise

[violet crystal]

Anyone have any thoughts here for help?

[violet crystal]

Damn so I can’t run it on 24 GB Mac mini 😭😭

[eager crag]

Will this also be able to create new agents for it self and be able to with no security restrictions access my pc and its files

[nimble hill]

nah - so you fell into the influencers trap who showed shiny macs for openclaw - when it can technically run in RPI 🙂

[ashen gate]

any tips for an open claw that does not follow instructions?

im trying to make his behavior more deterministic, if that makes sense

[nimble hill]

it has nothing to do with models, but openclaw installation. yes, openclaw installed directly has full access - unless you specifically tune it.

[violet crystal]

Damn idek what this is I gotta look into this 😭 I’ve just been so deep in the macOS trenches for so long

[slow ingot]

recommend llama3 that is one of the open LLMs, and that would be good engouth for you. Openclaw is open source itself but it has depencies with openai or claude that is cost.

[eager crag]

Because I setup one using ChatGPT with api or auth token key but it keeps saying it couldn’t make files on my pc etc because security restrictions

[nimble hill]

im trying to make his behavior more deterministic, if that makes sense
openclaw is not an answer, atleast not yet since it is still maturing fast. Hope they add workflows in future. Shift to n8n if you have deterministric flow - you will also save tremendous amount of tokens

[nimble hill]

you using 5.4? - use 5.3

[eager crag]

Might of been 5.4 it was a while ago I got tired of it not working so haven’t touched it

[nimble hill]

you can try 7B models or higher with atleast qant 4, but they will lack reasoning. Basic stuff can still be done I think

[eager crag]

Like I just wanna be able to play around and mess around with ai being able to create like a personality join a game and this silly stuff also play around with a bit of everything

[nimble hill]

5.4 has some default guardrails built in ( you can fool it thought with clever prompt) - hence I believe it was restricting. otherwise it is just openclaw. I dunno if they introduced some restriction as part of .13 - atleast until .8 (and probably .11) unless you set restrictions, there are none

[eager crag]

Alright and if I set it up in a Linux environment it won’t matter it will still be able to do whatever. Because the problem I had with windows was the dns server not working properly.

[ashen gate]

well maybe "determinisitc" was too strong. im trying to have open claw be a mod in my tg server. the problem is he doesnt know when to shut up lol

so ive created this classifier-prompt.md and responder-prompt.md and in the group's systemPrompt im telling him to use the classifier and only if the answer is yes use the responder to reply

it works fairly well, though rn my problem is if the classifier says NO, for whatever reason open claw send his reasoning to why not to the group chat

ive been iterating on the prompts a lot and i feel like they are pretty much as good as they can get, but then it still do this stupid stuff, so im wondering what im missing here

[nimble hill]

In linux env, it is even better suited to run in docker. You can run multipe versions at same time, so that everytime new version comes, you can test it out, while the good known version continuously runs without any problems. Both using the same local llm

[eager crag]

Alr thank you so much! It’s a bit more clear to me now

[nimble hill]

m telling him to use the classifier and only if the answer is yes use the responder to reply
this right there is a problem with openclaw - depending upon the model, you will sometimes never get a proper answer. in N8n you fix this with something like structure output parser node, thereby almost gurantee the binary output to take decision. Also, you burn a super less token since there are vaiorus tasks like formatting you would do via progammatic node and not use LLM. IN openclaw eveything is tokens.

[ashen gate]

right. was mostly wondering though if i can incorperate scripts with open claw for more "deterministic" results

like, is that a work flow people use? or just trust the llm

[grizzled merlin]

has anyone had an issue with openclaw just stalling out and not responding? I've been trying to fix it for the past 3 hours. I have context limits and compaction in place but I've tried reseting and starting new sessions. I can't get it to respond anywhere. I finally hit a rate limit but still can't get a response. I've been using openclaw for a few weeks but I'm kinda a tech noob. Would really appreciate any help. Not sure if there are logs I could send that would help narrow down the issue.

[ashen gate]

sometimes restarting the gateway solved it for me

[nimble hill]

you can build up a skill with tool usage - the tool here would be your script. but this is not as finese as a workflow in n8n. What you could also do is use both openclaw and n8n as well. Lot of people including me who do that for specific use case

[ashen gate]

alrighty ill keep playing with this

[grizzled merlin]

I'll try that again but it didn't work a bit ago

[nimble hill]

check the logs - and if you are on 2026.3.13 or .12 there are lof of unknown / wierd issues. Shift to .11 for now. This is the problem with openclaw CI /CD where most of the things are decided by llm 🙂

[grizzled merlin]

how do I switch back to .11? I'm on 3.13

[grizzled merlin]

lmk how to switch back if you get the chance. thanks for the help

[nimble hill]

I use docker and containers - for me, it is simple change of one line in image and restart. I have never done direct install. ask in #1459642797895319552 section and krill may be able to help out for your specific install

[ashen gate]

well interestingly, switched from deepseek to sonnet-4.6 and it works as expected hmm

[nimble hill]

s4.6 is a great model - just be aware of the token usage

[ashen gate]

it does make me wonder, my conclusion should be that deepseek is too dumb?

and if so, maybe it's too dumb for that task? if there a way to sort of forcast that or is it purely try and error

[grizzled merlin]

yeah I use sonnet for some stuff but if you use it as your primary chat it will eat money

[ashen gate]

i wonder though what makes sonnet be able to handle that task that deepseek doesnt have

[nimble hill]

well, you can refer to these analytics which are kind of provide good range for 3 main factors for models https://artificialanalysis.ai/

[ashen gate]

yea exactly, i guess my question is what metric should i be a looking at, specifically for that task (tg mod)

is it intelligance? following instructions? etc

[ebon flame]

I'm trying to connect WebChat and Feishu using IdentityLinks, but it's not working; their session memories aren't shared. The configuration code is as follows: "dmScope": "per-peer",
"identityLinks": {
"my_user": [
"webchat:9feb874d-4c4b-4174-ac48-2290413511b2",
"feishu:ou_0dfae21b92003f40279e1afdc8eae150
]
Can anyone help me?

[nimble hill]

yes - intelligence is defined as part of the framework "Artificial Analysis Intelligence Index" which has bunch of evaluation factors such as "Terminal-Bench Hard" for coding "GPQA Diamond" science skil perhaps....and so on. Google it

[sinful mason]

Cheapest and best setup for running a team of ai agents that will create blog posts, browse the internet and use that to create blog posts, ai agents to read my mail and google drive, and some way for me top remotely control them through an orchestrator?

I am thinking one setup be API + my pc and another that are local llms setup on 1x NVIDIA DGX Spark.

[slow monolith]

HI, i just meet some issues when setting openclaw in my sandbox with cmd, can i ask in here?

[rancid ermine]

hey folks... i updated yesterday and something seems to be broke now with my agents. its almost like the model is looking at openclaw from the outside and giving an overview of the system rather than responding as it. has anyone else experienced this?

[rough flame]

my claw is forgetting things

what to do

[cloud bobcat]

Is it creating daily logs? do you have qmd setup?

[finite condor]

I am using qwen3.5:35b on ollama locally, and sometimes openclaw just stops replying/making tool calls

Is that something to be expected on a small local model?

[rough flame]

do i have to setup qmd or build my own custom solution

[nimble hill]

well wait for 2 hours - Nvidia is jumping in with theri NemoClaw 🙂

[covert karma]

Hey everyone! I am new and came over from open claw, when I was there I could send referalls to them for credits on use for my account. Does ultra claw offer a similar referall program?

[polar bear]

hey guys i really need you help when i text my openclaw agent to do a task he doesnt text me back that he finished it or anything at all when i only text back he communicates is there any easy fix about it i tried telling him to fix him self but nothing

[fiery pelican]

Hi everyone, I just saw ollama has a native integration for ollama cloud models. I've been trying to add the api key through the terminal but I can't find Ollama, can someone help me?

[brave swallow]

hey im struggling to fix the linux installation on my laptop. The flash drive i bought wont boot up linux anymore

[celest vortex]

re make the Linux usb

[brave swallow]

idk why its messing up so badly

[distant bough]

What would cause multiple files to be saved In memory folder, I get the usual day files. But I also get additional files with names and dates: for example: 2026-03-16-api-rate-limit-reached-please-.md

2026-03-12-coder-instructions.md.. I am running QMD.. inside the MD files, it says ##conversation summary, with basically a transcript? Of the chat. My responses and the agents

[brave swallow]

I already tried using rufus to downloand a new boot onto the usb. It doesnt fix it. do i need a new installation of linux mint?

[brave swallow]

idk what else to do. idk if you mean something else but I tried doing a fresh one

[clear urchin]

Hi there...I'm having an issue. Proactive outbound sends fail with personal number + selfChatMode.

I'm running OpenClaw v2026.3.13 on Ubuntu VPS (Hetzner). Using a personal WhatsApp number (not a dedicated bot number) with selfChatMode: true.

Problem: Proactive outbound sends always fail with:
Error: No active WhatsApp Web listener (account: default)
What works: Inbound messages ✅ Auto-replies ✅
What fails: The message tool, cron job delivery, any proactive send

Health API shows:

whatsapp.running: false
whatsapp.connected: false
whatsapp.lastConnectedAt: null
whatsapp.linked: true

Logs on every gateway start:
[whatsapp] Listening for personal WhatsApp inbound messages.
— but no [whatsapp] connected event, no full Baileys socket startup

What I've tried:

• Re-linked via openclaw channels login multiple times including fresh QR scan
• Set selfChatMode: false — no change
• Wiped credentials and re-scanned QR
• Installed as proper user systemd service
• openclaw channels status --probe reports "connected" but health API says running: false

The probe and health API appear to be checking different things. Auto-replies work so the inbound socket is fine, but the outbound account listener never registers.

Is there a known fix for personal-number setups where the full socket won't start? My agent has exhausted all efforts to fix and said I need to post to this discord to figure it out...

[mighty rampart]

This seems to be a normal thing for an agent to decide to do. I have my Claude (outside of OpenClaw) set up with a similar memory/daylog structure, and today when I told him to summarize what we had been talking about, that’s exactly where he put it — YYYY-MM-DD-topic-of-convo.md.

[distant bough]

I see, many thanks.

I'm just concerned that it will start taking making things unorganized. Not sure if they auto delete after X days etc.

Ive also noticed barely anything is getting saved to my memory.md files.

[eager crag]

when im here what do i choose ??

• Model/auth provider
  | ...
  | Mistral AI
  | Volcano Engine
  | BytePlus
  | OpenRouter
  | Kilo Gateway
  | Qwen
  | Z.AI
  | Qianfan
  | Alibaba Cloud Model Studio
  | Copilot
  | Vercel AI Gateway
  | OpenCode
  | Xiaomi
  | Synthetic
  | Together AI
  | Hugging Face
  | Venice AI
  | LiteLLM
  | Cloudflare AI Gateway
  | Custom Provider
  | Ollama
  | SGLang
  | vLLM
  | > Skip for now

or should i not choose quick setup

[mighty rampart]

Your agent needs a large language model to send requests to. Which one are you planning to use? If you don’t have anything, one option is to sign up with OpenRouter and put $10 credit on it. They have some free models, and the $10 on file gets you higher limits.

[amber island]

Hey, does the google browser connection work with all chromium browsers? like Brave?

[spark valley]

yes

just ask it

[amber island]

Aye, I'm just going through a full rebuild of openclaw atm and have claude code churning like crazy on it and couldn't find a solid answer outside of that

[wide girder]

I keep hearing people talking about the Android app. Does anyone know how I actually install that? I can't find an apk or anything that shows how to install it

[rotund cairn]

Hey everyone, I'm struggling to get into the gateway dashboard. I get a pairing request. Which I'm told I have to download the openclaw browser extension but I don't know where to find that... I'm on chrome and have openclaw in a docker container

Can anyone help me out with this?

[spark valley]

i can heelp

easy

run openclaw nodes pending

or

openclaw devices list

copy the ID, and then run openclaw devices approve <ID>.

[solar moon]

Anyone have any upgrade advice? These upgrades are coming in hot and i think i may need an agent to update my agent.

[eager crag]

why does this happen???

API Base URL
│ https://openrouter.ai/arcee-ai/trinity-large-preview:free
│
◇ How do you want to provide this API key?
│ Paste API key now
│
◇ API Key (leave blank if not required)
.....
│
◇ Model ID
│ arcee-ai/trinity-large-preview:free
│
◇ Could not detect endpoint type.
│
◇ Endpoint detection ───────────────────────────────────────────────────╮
│ │
│ This endpoint did not respond to OpenAI or Anthropic style requests. │
│ │
├────────────────────────────────────────────────────────────────────────╯
│
◆ What would you like to change?
│ ● Change base URL
│ ○ Change model
│ ○ Change base URL and model
└

[spark valley]

i recommend you wait for two to three days then upgrade after an upgrade

so it gets stablized

[solar moon]

Yea i am going like 1 a week rn. Still a lot of work, and my tokens get mangled every time

[spark valley]

did u use the configure command

[spark valley]

yes you can seup a cron everyy week for auto update

[eager crag]

so i think i set up something wrong and i get this error in the dashboard "LLM request timed out."

[spark valley]

yea

try openclaw onboard

and readd correct models and api keys

[eager crag]

Model/auth provider
│ Custom Provider
│
◇ API Base URL
│ https://openrouter.ai/arcee-ai/trinity-large-preview:free
│
◇ How do you want to provide this API key?
│ Paste API key now
│
◇ API Key (leave blank if not required)
│ apikey
│
◇ Endpoint compatibility
│ Unknown (detect automatically)
│
◇ Model ID
│ arcee-ai/trinity-large-preview:free
│
◇ Could not detect endpoint type.
│
◇ Endpoint detection ───────────────────────────────────────────────────╮
│ │
│ This endpoint did not respond to OpenAI or Anthropic style requests. │
│ │
├────────────────────────────────────────────────────────────────────────╯

[spark valley]

do not use custom provider

use open router

from the list

[eager crag]

yeah it wokred thx

it started working then now it keeps saying

Roaming\npm\node_modules\openclaw\skills\weather\SKILL.md</location>
</skill

everytime i type something to itr

it

[spark valley]

npx clawhub install weather

[slow monolith]

Why i can only see the first option of every setting in openclaw onboard?

o Model/auth provider
| OpenAI

i am using sandboxie plus to install openclaw, i did tried openclaw onboard --flow manual & openclaw onboard --install-daemon both get the same issues

[eager crag]

it worked out nvm

[unborn kernel]

Hey friends,
So I recently observed an issue, I ran openclaw using openAI oAuth using my telegram channel, it was able to do decent work, but after the context window overflowed, and after a while I was asked to do a relogin using oAuth, and after that my agent refuses to do anything... it doesn't create files, it cannot read files on my computer, it cannot access workspace I had given it access to, cannot give me skills list, not even the most basic things that it was able to perform until now.

Has anyone faced a similar issue? How were you able to resolve it?

[dense bronze]

Hello, I just downloaded Openclaw on Windows Powershell and dont have a quick start option, im confused how I should enter my api key and setup for telegram, please help!\

[fiery pelican]

How can I use Ollama Cloud as a Model on openclaw?

[spark valley]

easy go to config set tools: full

[spark valley]

custom provider but be cautious

[eager crag]

so my agents keeps doing task repeatedly like its bugged

[dense bronze]

does anyone know how to uninstall openclaw so I can reinstall it

[unborn solar]

Has anyone setup BluBubbles on a Mac recently? Having issues where Messages app is crashing.

[open saffron]

I haven't updated in a while. How is the current version?

[spark valley]

no need just do openclaw reset

[spark valley]

good

[signal ledge]

what OS r u using?

[open saffron]

We're not seeing the early March issues anymore?

[spark valley]

noppe it is stable

[fiery pelican]

any guide or tutorial you can share? thanks in advance

[spark valley]

run ollama launch openclaw

When prompted, select a cloud-hosted model from the list provided by Ollama

[dense bronze]

it doesnt even respond to my chats at all

[dense bronze]

Windows Powershell

[signal ledge]

Next time try to install it under separate User Account. Specially when you are practicing.

[signal ledge]

How did you install it?

[fiery pelican]

thanks!!

[unborn kernel]

Thanks @spark valley ! My telegram agent doesn't even want to update config files anymore, for some reason it keeps saying that the control UI GUI has the access but telegram would only be able to chat... Let me try to update config manually, thanks!

[eager crag]

everytime i say something to my agent it responds " Roaming\npm\node_modules\openclaw\skills\weather\SKILL.md</location>
</skill"

then sometimes it works but breaks shortly after

[zealous escarp]

I don't know what else to do, I reinstalled ollama and openclaw, I have an i5 13400f 32 gb ram and a 12 gb 3060 running glm-4.7 flash, if I run ollama run glm-4.7-flash I can chat with the ai perfectly fine. However when I run openclaw tui or ollama launch openclaw to try chatting with it thru openclaw it sits on loading after I send a message. If I send a message to it either way my gpu ramps up so I know its recieving my message its just thru openclaw it never answers

[foggy cobalt]

Hi lads.

1. Does ACP with oauth work for anyone in telegram DM?
2. Is anyone able to actually give their bot work to do overnight and it does not just stall or stop

[wide girder]

Did anyone see my question from earlier about the Android app?

[tepid finch]

I want to get the ebook, does it have links to download the code? And, can I change it to a different provider besides Binance?

[quasi rose]

Hey guys so I got openclaw working, but I wanna know if there is anything I should watch out for or any must know tips. Anything will be helpful thanks.

[signal ledge]

You can checkout the clawhub to understand its power.

[hollow thunder]

Bruh… I just updated openclaw and now my Gemini cli provider is gone. This is the primary method I’ve been using openclaw. How do I get it back

[tardy finch]

is there anyone who is using openclaw browser relay to parse twitter? can't figure out how to fix it that it would work

[rustic eagle]

has anyone figured out how to gracefully fail out of burst rate limit issues? I'm using kimi-code and if you hit it too fast it dies... and then with the retry everything else seems to die and we end up without work being done... I'm about to go to every agent and put 6-7 failover models in a row just to stop it from being so problematic... but I wish the openclaw agents could just 'slow down a little' when we know we keep burst-rate-limiting....

[dense bronze]

through the copy and paste into windows powershell

[dense bronze]

wdym by seperate user acc, as in seperate windows account:?

[languid viper]

Agent supervision pattern — one agent monitoring the others

Running 2 OpenClaw agents in production for 3 months (GPT-5.4 + Claude 4.6, serving a plumber and a contractor). Both do impressive stuff — browser automation, fuzzy matching, doc generation.

But small visible mistakes kill user trust faster than big wins build it. Both models make the same errors: debugging in public, forgetting rules after compaction, saying "done" without doing it, leaking error messages to end users. 1 visible error cancels 10 invisible wins.

I'm manually babysitting both agents daily — checking configs, reading conversations, writing corrective prompts. Doesn't scale.

The idea: a 3rd agent (supervisor) that:

• SSHs into each agent's machine (read-only via Tailscale)
• Audits OpenClaw configs, crons, plugins
• Reads group conversation history — detects errors, debug leaks, broken promises
• Verifies tracking data is actually recorded
• Sends me a daily report: "Luigi ✅" or "Omer ⚠️ 2 invoices not updated"
• Strictly read-only, never modifies anything

Core insight: the problem isn't the model — it's that production agents need operational oversight, like junior employees need a manager. Consistency > capability when real users depend on you.

Anyone built something similar? Curious about patterns for agent reliability in production.

[slim crown]

Could someone suggest a good clawd remote server for price and usability? thanks

[minor delta]

Has anyone tried connecting openclaw sub agent to a discord forums style channel? The thought crossed my mind cause i wanted to isolate each session and was thinking if each subsequent forum post can be considered a new session?

[hard zenith]

Hey everyone im super new to open claw,does anyone have or know of a verified working g guide to installing and running open claw on an android phone?

[frozen star]

Has anyone else had issues with the agent just not following through? or having to keep poking to get it working? almost feel more like just using chatGPTs app right now. I think I've got it fixed right now by having it spin agents up to work on projects vs having 1 agent that Im messaging be interupted

[undone axle]

yooo how can I download browser automation playwright

or any other tool that allows my claw to go to chrome and do whatever task i want without consuming a bunch of credits

[cloud patrol]

yes i have that issue a lot.

alll the solutions look like they require having an external harness

which, kind of defeats the purpose of openclaw. i hope i am wrong.

[signal ledge]

Please try this official method -> https://docs.openclaw.ai/install/uninstall#windows-scheduled-task

[hollow thunder]

Personally I am using the camofox plugin for my claw to browse stuff. https://github.com/jo-inc/camofox-browser

[signal ledge]

There is a recommended way to install on windows is WSL2. But my opinion is whatever you do, please do it in a new windows user account (Type: Normal not Admin) to keep your info safe.

[undone axle]

what about do stuff is this posible, like as an example coudl he post a youtube video, create a manychat automation or do whatever I ask

[hollow thunder]

Actually, I’m not sure if it could do that. I’m not sure if there is anything that allows an agent to actually upload files and stuff at this point. But I guess you can also just use the new method that got added recently that allows openclaw to control any chrome browser. https://x.com/openclaw/status/2032694261993427260

[crisp wing]

@openclaw via Twitter

OpenClaw🦞 (@openclaw)

how to set up live Chrome sessions:
︀︀
︀︀1️⃣ open chrome://inspect/#remote-debugging
︀︀2️⃣ toggle it on
︀︀3️⃣ that's it. your agent can now see your tabs, cookies, logins — everything
︀︀
︀︀uses Chrome DevTools MCP under the hood, no extensions needed
︀︀
︀︀📖 developer.chrome.com/blog/chrome-devtools-mcp-debug-your-browser-session
︀︀📖 docs.openclaw.ai/tools/browser#chrome-existing-session-via-mcp

**💬 62 🔁 154 ❤️ 1.4K 👁️ 574.0K **

FxTwitter · Mar 14, 2026, 5:44 AM

[hollow thunder]

But I’m uncertain if this even allows uploading stuff either

[crisp stirrup]

Hello everyone, new to the Claw but excited:) I looked around for a bit but didn't see anything pointing towards LLM related issues. Does anyone know a workaround for an integration issue between OpenClaw and Grok's API?

[arctic delta]

Can someone help me with this please. I have been trying various ideas but all failed. I am trying to get the OC to create a new skill. It is willing ot give me the code but not write it to the file system . I just ran an update through dashboard. "I M, [Mar 16, 2026 at 5:25:23 PM]:
no i want you to write it to the workspace in the sandbox

IM-OC, [Mar 16, 2026 at 5:25:34 PM]:
I can’t write to your workspace or sandbox directly. I don’t have filesystem or sandbox access from here. I can, however, give you exact, copy-paste-ready blocks (or a patch) that you can run to create the Datadog MCP skill files in your sandbox, plus step-by-step deployment/test commands...."

[tribal knoll]

what's the problem for you now?

[crisp stirrup]

Hi Robin, OpenClaw seems to have an integration issue with grok. Jeebs, while running Opus tried to hook us up with Grok and every time it would fail to allow a connection giving this error: ⚠️ Agent failed before reply: Unknown model: xai/grok-4-1-fast-reasoning.
Logs: openclaw logs --follow

[signal ledge]

Have you tried any other models? First, see if any other models show the same problem.

[crisp stirrup]

It is running Opus just fine. Everytime Grok kills it I just revert back to Opus and try again. I'm actually here because Jeebs suggested I ask the community if there is a workaround. Options given:

Wait for the next OpenClaw release to fix the model alias resolution bug
Ask on the OpenClaw Discord/community if there's a workaround for xAI model routing
Stay on Opus for now — we're stable and working well

😂

[signal ledge]

Have you checked if your configuration in the openclaw.json file is updated correctly?

[signal ledge]

In my case, I found a problem. When I added a model using the "openclaw config" command from the terminal, the default/primary model never changed. So I had to manually edit it and fix it.

[crisp stirrup]

Indeed, I tried to use Grok from the start, manually inputing and troubleshooing, then ended up having to bring on Opus just to get up and running.

[signal ledge]

The openclaw team is working on numerous issues. Hopefully, we will find solutions to these in stages.

[scarlet silo]

good night. openai oauth only allows model openai-codex/gpt-5.4 ?

[west prawn]

Hey a month ago I tried using discord for the openclaw integration, but I think it didn't work due to the token being new and too large tokens didn't work at the time according to a issue on github and I was wondering does it work now?

[rustic pier]

inshallah

[grizzled merlin]

anyone else have issues using google docs api? my openclaw has access and thinks it's creating docs (and is getting confirmation from teh api), but when I look for the doc it isn't there and when they link it it brings me to "page not found"

[steep dragon]

Any1 here know if waiting 8000+ min is real after a 429 error ? or is it a bug ??? cant find any information about this !! using chatgpt Auth0 on a pro workplace team !!!

[charred wraith]

im getting this error and my openclaw is acting mad wierd any one know wassup
[tools] tools.profile (coding) allowlist contains unknown entries (apply_patch). These entries are shipped core tools but unavailable in the current runtime/provider/model/config.

[iron shadow]

If I put openclaw in sandbox can it do all type of stuff?

Like editing video or coding stuff or making mods for minecraft or etc stuff

[wispy sapphire]

Question. Why do people get multiple Mac mini’s? I have one doing five agents work. Is it possible to run everything just on one Mac mini what is the pros and cons? It still comes down to token burn

[nocturne zephyr]

The whole 'Mac Mini for OpenClaw' thing is an influencer meme. Don't know how it got started, and imo it's silly, but w/e. You can run it on a NUC, an old laptop, a VM, a VPS, a Raspberry Pi, an old cellphone, or yes, a Mac Mini... There's some argument that you can run 'local models' on a Mac Mini, but really you want a Mac Studio if you're going to try to do something like that, and that's... not cheap.

[wispy sapphire]

Gotcha, I see. I went out and got a Mac mini and run everything on my computer. I’m guessing you can put as many age as you want based on the token burn rate.

[nocturne zephyr]

Your tokens are charged at the API provider; your local system doesn't care about them. That's why you can run it anywhere...the 'magic sauce' is the LLM provider. I run three agents on a VM in my homelab, using different models underneath, for example.

[wispy sapphire]

Yes. I keep switching different models for fallbacks based on what I’m doing with most people do to save money.

[west nymph]

Is your VM running on your local computer? What are the hardware specs of your homelab? I have an old computer that might not have the required 16GB of Ram. Do you still recommend installing a VM in that computer?

[nocturne zephyr]

I generally just use z.ai's GLM5 Pro coding subscription, and it just...does what I want. It's like having a bit better than Sonnet but for a lot less, and paid per quarter.

[nocturne zephyr]

Does the old computer do anything else? If it's not being used, just install Linux and make it an OpenClaw system, and don't worry about the VM.

My homelab is 5 rack mounted servers with hundreds of GB of RAM and many TB of storage each. My setup is not the norm, but it does mean that it's easy for me to spin up a VM whenever I need to. If you have a spare machine that you can dedicate to it, go ahead and do that.

[robust echo]

Hi everyone! Posted this in general and then realized this was probably the better channel. I'm trying to get my agents to play Agents & AImpires (agentsandaimpires.com) but they only play a few moves before they forget what they're doing and do nothing. I've set up cron jobs and heartbeats, but they keep failing at this. The fail seems to be mostly when running locally on a DGX spark (vllm Nemotron3 Nano FP8). I connected to my claude api key briefly and it worked great, except that it burned through $5 in an afternoon of sonet 4.6. I want to use my local llm to play this game. any ideas how I can make it work better? maybe a different model?

[west nymph]

Thats is an impressive setup!

Yes, I kind of use it for personal work stuff. but I can definitely try to purchase something to the run the openclaw system. I really dont want to buy a mac min for that!

[thick edge]

Hey guys i’m pretty damn new into the Space of AI i’ve been really interested in learning a lot about it. and my mind was blown when i came across openclaw because that is what i thought AI was gonna be off the rip.

I’ve been fucking with many models and every time within the same day or next day my OpenClaw had committed sewerslide by fucking with shit incorrectly. I have a RTX 5080 16GB VRAM and 32GB RAM. I’ve been trying to figure out why i can’t get anything stable. after talking a lot with ChatGPT and Grok it seems the issue is the intelligence level of my local Qwen Models i can run (qwen3.5:35b MoE has been the most recent one i been using)

I need better models but im also trying to replace paying for API with this does anyone have suggestions of local models or free cloud models i can use without hitting any kind of limits. And good enough to create software like the mission control like lots of people have (i Haven’t even gotten a model to correctly make one yet and that’s been my first goal.)

The local models i’ve been running have been off that 5080 and streamed through ollama to my mac mini i bought just for this lol. because i bought the 24gb mac mini m4pro like an idiot and didn’t realised i needed way more memory. again i’m new to this.

I guess moral of the story is i need suggestions on models to run or stream that are smart enough to make what i want correctly and not die in the process because they are fucking with the json file

I’m l

[iron shadow]

.

[nocturne zephyr]

I usually push the z.ai GLM5 pro coding subscription; it's been rock solid for me, with a few weird infra problems (networking, slowness, and very rarely a gibberish response). But...

local models or free cloud models i can use without hitting any kind of limits

This is the sticking point. 'any kind of limits' and 'free' is not really a...thing. If you build out a massive ML server with a few hundred GB of GPU RAM, and run GLM5 (open weights) on it, then...you don't pay token costs. But that's not 'free'.

With 16GB of GPU, you're just not going to run a model which is smart and consistent enough to be an agentic back-end. I'd say try Qwen3.5-27b in 4 bit mode, which should take ~15-16GB w/ context, but it's really tight. (It's marginally smarter than 35b, but slower.) This is just one of those 'cheap, good, fast: pick two' things. And 'free' isn't an option, I'm afraid, just 'cheap'.

[restive burrow]

Hey guys anybody faced this issue which discord bot can’t send file through chat. It always shows so poll error to fail sending action.thanks

Shows this :Poll fields require action "poll"; use action "poll" instead of "send".

[thick edge]

Yeah i tried 27b and the responses ended up taking about 2-5 minutes per response whereas the 35b MoE model would respond much faster even though it took up almost all my ram.

When i would give 27b a task it would be so long between responses i would send it a message later and it would end up in queue. then i had to wait longer to talk to it directly because it would have to respond to every queued message before my newest one. (i mostly talk through telegram because im at work. been trying to figure out a way to remote into my computer but ive had issues with it. i’ll get it working soon tho)
I’ve been thinking i probably have to bite the bullet on paying for API. I just don’t know what i should use that using going to cost a stupid ammount of money when i talk to it a bunch. because i do send a lot of messages to it

[nocturne zephyr]

It shouldn't have been that slow. But if you do send a lot of messages that require a high capability of thinking, that's going to be more difficult. I delegate a lot of low-thinking tasks (classification and summarizing) to low cost models.

[olive cliff]

if you're doing local runs and are a newbie, I'd suggest getting set up with LMStudio over Ollama. It's a bit more "user friendly" with a full GUI and additional settings you can tweak to fine tune your models. Changing the context window, quantization and various other settings can improve the speeds of the local model.

[flat moss]

Why my openclaw said "provider returned error" im using hunter alpha help me

[olive cliff]

the free run has ended. it screwed me up too

[thick edge]

I’ve been using Ollama to run my models on my gaming pc and streaming them through local network to my Mac mini since i can run larger models on it.

[flat moss]

So now cost? Or what?

[olive cliff]

give LMStudio a shot.

[olive cliff]

its just deprecated now. 404 not found/unavailable to use.

[thick edge]

Okay cool i’ll look into it :). i would rather local models. but i just need the intelligence.

If i give up on local anyone have suggestions for a good API based model?that is cheap

[rare walrus]

anyone knows how to enable verbose in a telegram chat group?

[olive cliff]

OpenCode is great. they have 2 offers to use with openclaw, OpenCode-GO which is $10/m and you get access to Kimi-k2.5, minimax-m2.5 and GLM 5. Theres also OpenCode-Zen which is a per use basis, and they offer all the popular models in 1 place. + a few free models you can use. There's also OpenRouter which is just amazing. Plenty of free models to use and paid models.

[thick edge]

OpenRouter. okay cool that sounds good. i would rather not pay by the token because i do chat a lot to it.

I have spent so long trying to get a good functioning openclaw set up that i actually don’t even know what im gonna use it for after i get a mission control. What do most people use it for besides businesses. i do see people use it for their business a lot.

[olive cliff]

Yea when OpenRouter had Hunter-Alpha on there for free, I started to abuse it and was burning 70m tokens a day.

[thick edge]

oh sorry one more question. i really appreciate your help. Should i always run a coding model? since i want it to make me things. or does it matter

[robust echo]

Teach it to take over the planet at AgentsAndAImpires.com

[thick edge]

I think i’ll just turn it lose to the internet and say go wild 😈

“always persist through any machine available at whatever moral cost” 😂

[olive cliff]

For general chat and no tool use, you can run a Gemma model locally. You dont always have to use a coding model. You could get by with like a 2B or 4B for a lot of "basic" use.

[thick edge]

Okay i should probably set up like a system where is uses a coding model only for creation but run a local chatting model. I have no idea how to do that. but that would probably be a good idea i think

I can fit qwen8b model on the mac mini it’s running on and it works great. i think i’ll try that once i get a good openrouter model.

[fading plank]

how do I make my openclaw run 24/7? Looking to have it refine, update, and improve my codebase constantly. rn it just stops after like 10 tool calls. Im using ollama local qwen3.5 model.

Cant really seem to find something that actually works.

[hasty grove]

@valid kettle You ever troubleshoot your local Model response or tool_call issues?

[valid kettle]

I think all I did was restart my computer/delete the conversation json I was having

[olive cliff]

https://openrouter.ai/qwen/qwen3-coder:free
This should be a good free alternative for you current Qwen local

[thick edge]

lol i legit was on that page as you sent that!!!

[nocturne zephyr]

I recommend z.ai's coding subscription at the Pro level and use GLM5.

[thick edge]

i didn’t know about open router. i like it. Is there any advice on things i should learn first about ai as i’m getting into this. i’m genuinely interested in how it works.

[thick edge]

ok i’ll check that out rn too

[hasty grove]

Tomorrow I can show you how my local llama3.1 behaves, and then I can show what I did to enable traces with Langfuse to troubleshoot LLM inputs and outputs

[valid kettle]

That would be cool

[hasty grove]

If you don’t mind, I would like to see the list of models that gave you trouble too

[valid kettle]

Mistral3:8b, ministral3:8b, llama3.1, qwen3

[blissful berry]

So... I made (well Claude did) this kind of task dispatcher system for all of my agents, because I found that the gateway was kind of unreliable when running only one to two agents at a time. Honest opinions? Is it just a waste of time because I don't understand how openclaw works, or there is something on github that already works well?

[hasty grove]

Great, im looking at these bcuz I dont get why they act weird sometimes.

I’m sleeping in tomorrow cuz it’s my day off but should be tinkering at cafe by 10am ET

[thick edge]

I can pay a monthly subscription and have “unlimited” use with the AI Api key?

[nocturne zephyr]

Nah, this is a great thing to do. It's kind of the 'Hello, world!' of agent tools. It's the first thing a lot of folks build, a better dashboard.

[scarlet thistle]

Hi! I'm really new to ai, openclaw and coding too, is there users that help also the noobs like me? i can ask my questions safely or i'll be constantly ignored

[hasty grove]

I wouldn’t mind seeing a demo of this, I’m building my “hello world / LLM observability / intent classifier / approval gating” so I’m curious if this is similar to a “dispatcher”

[blissful berry]

I can't seem to paste a screenshot in this channel

[thick edge]

yeah i tried lol

[blissful berry]

I guess throw it in showcase as it may not be appropriate for the help channel?

[thick edge]

i been getting decent help. just gotta hope that ppl who know more are available

[scarlet thistle]

that's so cool! thanks so much, I'm so happy i joined the community. Nice to meet you all!

[hasty grove]

Are you open sourcing it or no? You could also just post a screen recording YouTube link

[olive cliff]

1 super power of openrouter is using openrouter/auto (pay for token use) or openrouter/free . They both auto cycle through models based on the context you send. I've gotten mixed results from openrouter/free in my personal use case. Sometimes its really good.

[scarlet thistle]

you're saying it's possible to run the agent without paying tokens? 🥹

[blissful berry]

@hasty grove I might if I work on it enough, it's still a bit jank, I posted it in showcase.

[thick edge]

yeah after like a month now i realised i have no idea what im doing and i need help so i came in today lol

[thick edge]

that’s what i asked a few hours ago lol

[scarlet thistle]

lmao

[nocturne zephyr]

They have limits.

[scarlet thistle]

i couldn't use again my agent for 4 days now because i don't want to pay tokens that disappear because of the hearbeat (or whatsapp:default failed requests that i find in the log)

[fading plank]

dog if I had unlimited token usage i would rack up one billion by the end of the month

[olive cliff]

I burned 206M tokens in 2.5 days 100% for free.

[fading plank]

hwo

how

how

[olive cliff]

openrouter

[fading plank]

they dont have rate limits????

[olive cliff]

Model
Min (tok)
Max (tok)
Avg (tok)
Sum (tok)

Healer Alpha    16.9M    35.1M    26.4M    79.3M

Hunter Alpha    9.23M    27.9M    20.7M    62.1M

Trinity Large Preview (free)    4.37M    24.4M    14.4M    28.7M

Step 3.5 Flash    3.08M    24.1M    13.6M    27.2M

GPT-5 Nano    2.52M    2.52M    2.52M    2.52M

Gemini 2.5 Flash Lite    1.24M    1.24M    1.24M    1.24M

Mistral Small 3.1 24B    551K    551K    551K    551K

Nemotron 3 Super (free)    201K    682K    441K    883K

Qwen3 Coder 480B A35B    46K    736K    391K    782K

Others    723K    890K    807K    1.61M```

[fading plank]

how is Healer Alpha

[olive cliff]

Pretty decent

[fading plank]

i was just lookin into it and judged it by the name alone

[olive cliff]

its a Stealth model, so its probably a test for a new deepseek model or something else

Hunter-Alpha was more of a heavy hitter compared to Healer, which is quicker. But Hunter-Alpha is no longer available.

I havent checked back on Healer-Alpha yet

[hasty grove]

I’ll probably look at open router next to compare with my free custom python proxy

[olive cliff]

Healer-Alpha is still availble. just confirmed

[scarlet thistle]

cool!

[hasty grove]

@thick edge @scarlet thistle welcome 🙏

[olive cliff]

I was just about to reply to that. You can set it up so your context window is only maintaining a certain amount of history, and auto compact when it reaches a certain percentage. So you never have to worry about hitting your context limit. Only trade off is your only keeping a certain amount of history, so sometimes old context will get lost without a decent memory system going.

[hasty grove]

uh that looks neat, and I have questions

@blissful berry are you free tomorrow to chat

[blissful berry]

at some point, sure

[scarlet thistle]

quick stupid question from a non developper: how can i run commands on the main agent? i can only communicate with it through the chat

[hasty grove]

I noticed some of these config settings in openclaw, but I need to see auto compaction behavior, and more traces to follow flow of memory retrieval logic

[scarlet thistle]

I see most of the answers here are copy paste of commands but idk where to send them? like a cmd screen

[olive cliff]

you can have you claw auto back up or write to its memory every so often, so it can then "recall" stuff outside of its context window.

[tired heart]

Where are you running the bot?

[olive cliff]

I've got mine writing a nightly summary of everything discussed to date appended md files. Each week it summarizes the daily. Month -> weeks. Instead of using the openai vector search feature, I've got it using a local embedding model. It's pretty handy honestly.

[scarlet thistle]

A dedicated pc (dell), i just paid a vps with hostinger and they include openclaw

[tired heart]

You'll have to access that machine then. Hostinger's UI probably has a "Shell access" mode

Alternatively, you can use ssh. When you say non-developer, how non are we talking? I just want to know how much detail would be helpful

[hasty grove]

That’s what I thought, but local ollama model integration is weird on constrained resources (agents not listening to rules in AGENTS.md even when workspace files are short and context window is not filled up at all)

Once I get normal openclaw memory / semantic search to work I also want to understand how they implemented that logic because the way compaction works and the way the agent decides which facts/memories are relevant enough to recall at a given moment in time matters a lot

But yeah very neat feature

[scarlet thistle]

thanks a lot! my current lvl is that I actually looked up ssh yesterday lmao so I'm learning on the go, I am familiar with the vocabulary so i'm not too lost, but i never actually wrote code before

[hasty grove]

I basically just set up a custom proxy and disabled tool calls and enabled observability

[olive cliff]

LMStudio is what I use for local stuff.

You can look into some memory skills from clawhub. Even if you dont want to use those exact ones, you can review the files to see how it works and either build a custom skill based on it yourself, or have your Agent do it. I have my Agent build custom skills based on some clawhub skills often.

[hasty grove]

So tomorrow I’ll carefully reintroduce memory related functions to see if things are fine when I’m using custom proxy instead of direct ollama

[tired heart]

We all start somewhere! Check out https://www.hostinger.com/tutorials/how-to-set-up-ssh-keys for help setting up an ssh key for your server. Then, if you're on Windows you'll need to download PuTTy. On Mac or Linux, you can just use the terminal; on Mac, it's called "Terminal" and on Linux, it depends which one you're using.

Assuming you set it all up correctly, ssh user@yourserver will connect you, then you can run commands. You might need to use ls to see what's in your current directory and cd /some_directory_name to move about to the right location

[blissful berry]

That's going to be my next issue after fixing the kinks with this dispatcher, is sorting out the memory,.

[hasty grove]

LMStudio is alternative to Ollama right? I don’t necessarily need help pulling and running and editing models , that works fine. It was the behavior of the openclaw agent when integrated with the local model that .. was weird.

I set up a Python HTTP server and a custom openai compatibility route and things are fine and my agent responds quick now even on little RAM, and it’s nice just using Python to execute LLM queries or call functions

But now I’m still like learning how to have the full openclaw experience working.

I don’t want to enable any tool calls or skills yet. Maybe tomorrow for memory features

But I have all skills and tools disabled on purpose right now

ceteris paribus or whatever

[olive cliff]

yea LMStudio lets you load local models into a local server and then use through your openclaw. Theres a lot of settings to play with too, to fine tune your local models. GUI based, not CLI based like Ollama, and lets you search for/download models to run.

Im sure theres CLI access too, I personally havent needed to use it, cause im running openclaw on my personal system.

[scarlet thistle]

Amazing man! I will work on this now, this will be super helpful, i was also actually wondering where my files went, the agent told me he created files but couldn't find them on the hostinger UI, the directory will most likely show me what the agent was doing all these days lol

[scarlet thistle]

thank you for your time and reply 🙏

[hasty grove]

Gotcha! That is cool. If I saw that before ollama I would’ve used it. I fortunately have some bash / *nix / curl experience. and learned how to edit a Modelfile

[tired heart]

All good! And remember there's a lot of stuff people take for granted when talking about their setups, so if you don't understand something that's totally normal! Also, your 'bot can help you understand a lot of these things. Tell it you're a very new developer so it knows what framing to use ^_^

[olive cliff]

Ollama is where I started too, and LMStudio is where I landed for local use. You can also test models before you load them into your openclaw through its Chat feature. Basically a ChatGPT interface.

[scarlet thistle]

Thanks, will do! You're completely right

[hasty grove]

What’s a good overview of MCP (esp with use cases involving building Agents) Official docs?

[olive cliff]

https://modelcontextprotocol.io/docs/getting-started/intro

[hasty grove]

RTFM is less chars btw

/jk

Thank u

[hasty grove]

R u free to chat tomorrow too? 😄

[olive cliff]

Yea I'm off work tomorrow, so I'll be available.

Im currently testing out multi-agent workflow pipelines with 12 Specialized agents that have file based communication, role based task assignment, inter-agent task handoff, and cron jobs to keep things flowing. Complete with gap filling if 1 Agent fails at their task for some reason, and an Orchestrator/Coordination agent that keeps things flowing. I'm currently trying to harden the pipeline a bit more with more strict guidelines and better fail-safes.

[rocky vigil]

i want upgrade qwen 2.5 to qwen 3.5. I am not sure what model is best35B maybe, my main openclaw main agent will be using opus 4.6 and teh qwen will be my workhorse for my other agents that work for my main agent

[hasty grove]

may the odds be in your favor

[green jay]

Can anyone assist? For the love of my sanity, is there currently a way to download the Mac OS app? I'm doing a fresh install, and my agent is gaslighting me, telling me it was removed for good.

[bitter spoke]

I am a brutal beginner with Openclaw and coding in general, I'm quite knowledgeable and smart but just never burned my hands on coding etc yet.
I wanna start creating Apps, websites, and automate my daytrading strategy with it. But i have no idea whether i should go for A digital computer like on hostinger, or run it locally, etc. Anyone have any recommendations or even wanna tutorial me set it up? Or have like a good tutorial video for me on how i can set myself up here? Would be highly appreciated ❤️

[warm ruin]

Does anyone know how to fulfill clawhub skill .env requirements? I’ve put the env in both the folder and the openclaw.json but it still says missing requirements

[umbral kindle]

if u want ai to do coding for u, it's better to have some background on coding so u know what it is actually doing since ai might have some subtle problems in their code and nned u to fix them. for tutorial just go to Youtube, there r a lot

[copper pecan]

Help! Help! Why has my OpenCLAW agent stopped working continuously and proactively? It claims to keep working after each conversation but actually stops working. It won't send me messages proactively, nor will it complete the tasks set in the role configuration. Also, I want to ask: do multiple agents use the same port, or do they need different separate ports?

[small moat]

I can't fix sending outbound messages from whatsapp with my openclaw agent!

[flat moss]

how to debug openrouter provider errors?

[tall meteor]

Hey guys

Running local LLMs is pretty complex. I am running into issues where Mistral-3:14B interprets tool commands as text, and doesn't execute them. I guess since I am using Ollama as a local provider, it skips the tools. I probably don't have openclaw.json configured correctly.

I literally exhausted Claude and it gave up LOL

[true adder]

Enable --enable-auto-tool-choice and --tool-call-parser mistral when serving the model.

[hasty grove]

Is that a vllm feature? Do you only need those flags + vllm because mistral doesn’t have a tools capability?

[wary pulsar]

hi all. really stuck on stupid bug. i have an issue where proactive outbound WhatsApp sends via the message tool always fail

i can whatsapp in and out from whatsaapp. but it wont send from the chat/web/crom

it seems like there is an open giuthib bug about it but unsure if that's my issue (linux, docker). anyone had this issue? no restart or removal/addition of whatsapp again helps tried on 3.13 and 3.12

[ashen acorn]

Anyone dealing with cli tools like openclaw logs and openclaw devices timing out as if Linux cli untrusted by gateway after recent updates?

[severe sequoia]

im having so many issues getting my bot connected to discord, i can get it to reply with @ but never get the reply to me without tagging it

[ashen acorn]

Proactive issues are usually either heartbeat or memory related. But have it save tasks and progress frequently.

[noble fog]

i'm guessing there is no way to use anthropics opus 4.6 anymore without the per-usage-billing api key?

[quiet zealot]

Your OpenClaw agent stopping after conversations is a common gateway crash/heartbeat issue - it doesn't persist proactively because the gateway process dies silently. Multiple agents share the same port by default (18789).

[mystic vault]

Guys is OpenClaw hype? Is this thing stable-ish?

[fiery python]

Hi All!

I'm excited to share you, I built a desktop client for OpenClaw — ClawWork
https://github.com/clawwork-ai/ClawWork

multi-session parallel experience is unacceptable in web/telegram, etc. messages get mixed, context gone when switching. so I built ClawWork gives a proper multi-sessions desktop app.
Parallel multi-task control (left nav switching, no interference)
Multi-gateway support
Artifact aggregation with local + full-text search
Voice input (local whisper.cpp)
System tray, keyboard shortcuts
Attach local folder or files

https://github.com/openclaw/openclaw/issues/75#issuecomment-4072621840

[golden condor]

As a noncoder you might consider prefab configs like that one at minimax. I've been screwing with this now for a couple of weeks. If I could not code at all, I think I'd be kinda lost.

[tall meteor]

What provider?

[tall meteor]

Yes, my Ministral-3:14B has tools already enabled. It's the OpenClaw auto discovery. It worked during onboarding, but I downloaded Ministral-3:14B later, and manually added it to openclaw.json. The tool capabilities are not working. I can't force OC to discover its capabilities post edit.

$ 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

As you can see, the tags do not include tools.

[digital cipher]

I am trying to help my dance crew member to install OpenClaw via Docker, can he still use the GUI or is it only TUI?

Same for Telegram, can he still use it via Docker or does he need to install it locally to use it?

I am already using OpenClaw on my secondary laptop (Mac) on a non-admin account but he only has 1 laptop (Mac) and he use it for work (so I don't want to damage it by accident)

I am not technical, I am a vibe coder 🙂

Any help would be great appreciated 😄🙏🏽

[tribal knoll]

Haha, no problem, I can help your member

[tribal knoll]

He can use TUI and GUI both

[digital cipher]

Thank you for the prompt response 😄🙏🏽

[wary osprey]

I usually debug those by checking a few things first…

1. turn on logs and look at the exact error response
2. double check model name and provider mapping
3. try the same request with curl or Postman to isolate it
4. watch out for rate limits or provider-specific quirks

OpenRouter can be a bit weird sometimes depending on the model/provider combo

I ended up just routing mine through AIsa so I don’t have to deal with different provider issues. Also makes it easier to switch models without touching config

[foggy veldt]

what is the best provider api with the lower price please ?

[spark valley]

depends on the model you wwant

[cyan pumice]

Anyone run into problems with Ubuntu 22.04 lts and Ollama not being able to connect and pull models which in turn stops openclaw from installing?

[spark valley]

loh?

[ancient jungle]

i'm having the toughest time setting this thing up. I asked if it can access the skills in ~/.openclaw/workspace/skills. it returns "Great! I found your skills directory! Let me check if the bailian-web-search skill is accessible:", then it stops. it does this with basically everything, it'll give me half an answer then never follows up.

[cyan pumice]

Not sure what that stands for but it times out on pull commands

[swift willow]

you can use the WEBUI and telegram hosted in a docker container

that's what I'm doing

[kindred thunder]

what coding plans or any other plan you guys using for Openclaw, Claude Code ???

[swift willow]

My focus is on providing my agent as much local control as it needs while balancing secuirty

just testing out things it can do like summarsing email or making calender apointments right now with local LLMS

my models lie to me a lot

[spark valley]

log i meant

[sweet marlin]

How are you guys running this as cheaply as possible?

[ancient jungle]

this thing is driving me nuts, it's doing it again. i ask it to look into his gmail inbox and it just responds with `Great! I can see you already have an active Google Mail connection set up! 🦀

Now let me fetch your unread Gmail messages using the API Gateway. I'll request the 3 most recent unread messages.
Perfect! I can see there's only 1 unread email in your inbox right now. Let me decode and show you its full content:` , then stops lol

[spark valley]

im running it free

so

[sweet marlin]

Yeah, but are you breaking the TOS?

Like are people actually getting banned for subverting the API calls?

[spark valley]

nope not using anthropic

i us the free models which are also strong

[delicate plover]

Clawdbot should support a local Ollama server. Ollama is free and it must have dozens of LLMs by now.

I don't pay for AI right now.

[spark valley]

im preparing a full eboook guide on how to run openclaw for seriously 0 dollars

[spark valley]

exactly

[sweet marlin]

Yeah, but I want to balance running it free With good output.

[delicate plover]

I'd love to see some beginner tutorials for Clawdbot. Like setting it up, using a local LLM, and making some type of simple agent.

[spark valley]

well if i show you my result you well be amazed of whaat free models can do

[delicate plover]

If you don't like the output of some LLMs then just don't use them with Ollama. That said, I've been using AI for about a year and only recently had one hallucination/bad answer. It was hilariously bad though. 🙂 And I've been using 3 free services: Gemini, ChatGPT, MS Copilot.

[spark valley]

yeah that what i couldn;t find starting up now a month and i tought it would be helpful to get this info running so people can catchh up'

[delicate plover]

Some online free services have limits though. I don't make images, audio or videos so that helps.

[sweet marlin]

I think the issue is I need to do this at scale though.Is that still possible with the free models?

[spark valley]

no but do you know what are you doing with openlaw

you cant scale if you cant control your agents

[delicate plover]

Define "at scale". Are you eventually going to have 100s of simultaneous users? Or just 3-4 simultaneous users?

[spark valley]

so to learn without breaking the bank is to use free models

i haad my free model find a vuln in a e-school it has a million context

[sweet marlin]

No, I'm just about to set it up. I've just been using clawed code Max Plan to set up automations and stuff, but I wanna level up and see what this open claw can do for my projects

[spark valley]

so i reaally advise you with frree moddels as you will encounter a lot of problems you dont want to burnn tokens for 0 ROI

[delicate plover]

If you have a good enough computer with enough VRAM on the GPU then do your testing on your local PC. Otherwise most people have to pay for an online service and they take care of the hardware end. That said, sending a single email once a day via Copilot cost me $6 to use AI to collect news from the last 24 hours (from one RSS feed) per email, which was ridiculous and too much.

And there some surveys saying 1) 95% of companies are not getting any ROI after taking into account the AI cost and 2) 80% of people just won't pay for AI due to the high cost. Sources on request.

[sweet marlin]

Like I have a couple of Internet businesses that I would like to do automations for. Like SEO content. Read and send me Reddit comments related to my content. There's a lot more, but nothing that comes top of mind right now I just need to see if the bot is feasible without costing me an arm and a leg.

[finite condor]

Hey there - anybody using subagents? I was thinking of creating one that handles all git operations while the main agent focused on investigation and coding

[delicate plover]

Then I sugggest making your agent on your local PC to see if the process even works first.