general

[sullen locust]

compat

[gilded obsidian]

I'm reading on it 😄

[sullen locust]

no idea

[sullen locust]

https://tenor.com/bIPyu.gif

[gilded obsidian]

hmm there is no llama.cpp specific one that I can find

[oak jackal]

https://x.com/steipete/status/2053114837698249190?s=12

[mystic orchid]

@steipete via Twitter

Peter Steinberger 🦞 (@steipete)

Peekaboo 3.0 is live. Biggest release since 2.0.
︀︀
︀︀⚡ Action-first macOS computer use
︀︀👁️ Unified screenshot + UI detection
︀︀🧩 Cleaner JSON across CLI + MCP
︀︀🛠️ Better snapshots
︀︀
︀︀I started this last year, but the models just weren’t good enough. Now they are. peekaboo.sh

**💬 106 🔁 132 ❤️ 1.9K 👁️ 183.7K **

FxTwitter · May 9, 2026, 2:08 PM

[pliant tangle]

Anyone knows how to solve the issue

⚠️ API rate limit reached. Please try again later ?

[tepid kestrel]

thats not an issue.. that how it works. your api has hit the limit

[pliant tangle]

But I have credit on my openrounter balance then how ?

[gilded obsidian]

@pliant tangle cash, credit, things of monitary value, bribery..

[pliant tangle]

Brother I have cash on my openrounter it's not like my cash is empty & I'm trying

[tepid kestrel]

thats the rate limiter, doesnt mean you are out of tokens. You just have to wait

[gilded obsidian]

@pliant tangle Ask the agent to try using the API via curl, and see what it's returning exactly.

You may get a much more useful error.

[pliant tangle]

Ok

[pliant tangle]

How much time ?

[gilded obsidian]

Depends on the service, should be defined in their docs somewhere..

[pliant tangle]

Bro can I dm you one screenshot if you don't mind ?

[gilded obsidian]

I'd rather you not, I'm a newb here too. I might know technical things but I'm a claw-newb. use the quote block and paste the curl returned data (don't leak your api token)

More people in here that can help you than just me 🙂

[pliant tangle]

Actually my question is I provided a API key to my claw agent which have still a lot of balance

Then how my API rate is limited?

[gilded obsidian]

Like @tepid kestrel said you could be rate limited if that service has a limit, are you hammering it? Again the curl returned data may return more info than just a flat out deny.

[tepid kestrel]

https://tenor.com/view/keep-goin-gif-6704079567258296337

[pliant tangle]

Google Generative AI API error (429): You exceeded your current quota, please check your plan and billing details.
Model: google/gemini-2.5-flash
Provider: google
Error code: 429
Reason: rate_limit

This is the error

[gilded obsidian]

There ya go...

geminiflash the free one even with API key has LOW limits

thatis why it should only be used for certain things, not as a main model

I think I'm only using it for certain memory management or something or other

Are you sure the API key you are using is from an account with gemini credits, are you sure you aren't hammering it more than you think you are?

Tokens per minute (input) (TPM)
Requests per day (RPD) ``` https://ai.google.dev/gemini-api/docs/rate-limits

[pliant tangle]

Yes, Gemini 2.5 Flash is my main model via OpenRouter. I was doing heavy testing for 7+ hours today — building skills, restarting the bot many times, running hundreds of commands. That's why it got hammered. The API key is from OpenRouter, not directly from Google. Should I switch to a different main model?

[gilded obsidian]

ohhh I missed that part I am sorry, but I am sure OpenRouter also has rate limits

[tepid kestrel]

what if we would say no?

[gilded obsidian]

@pliant tangle https://openrouter.ai/docs/api/reference/limits

[pliant tangle]

Ok

[pliant tangle]

Let me check

[gilded obsidian]

I mean I can tell you my opinion on models, but.. opinions are like assholes ya know, everyones got one 😉

@pliant tangle Tell your agent to read https://openrouter.ai/docs/api/reference/limits and request what your current limit status, it should be able to figure it out from that doc (unless your all out of credits on all models LOL)

[tepid kestrel]

mate, you are obviously new to this kind of ai game. So let me tell, there are many people out there who would say different things on how to do something. Some say its the best to run local AI and some say its the best to have a subscription. In the end, its YOUR decision, YOUR project. Only you know how much you need

[gilded obsidian]

My only opinion right now is, DeepSeek Pro is 75% off, I won't say anymore than that 😄 It's why it's my main model until I figure out my local one and why its not working properly in openclaw 🙂

I've actually never used OpenRouter for more than 1-3 queries in a row, and never used the non-free tiers.

[pliant tangle]

@gilded obsidian @tepid kestrel thanks guys for your support & recommendations

I really appreciate that

Thank you very much

[pliant tangle]

Ok

[tepid kestrel]

go to #1459642797895319552 and open a thread, the help there is way better

since our beloved krill is back

[pliant tangle]

Yes I started yesterday morning, I'm totally new

[gilded obsidian]

Yeah we're kinda just shootin the shit here 😄

[pliant tangle]

Thanks for you helped alot

[gilded obsidian]

@pliant tangle the only problem with getting hooked on DeepSeek Pro/flash right now is your going to be like whoa I'm getting Awesome Claude Opus like results for like $20 a month and spending millions of tokens.. then the 31st comes and the 75% evaporates and your stuck looking for another model before it spends all your money 😄

[pliant tangle]

Is deepsek model expensive?

[gilded obsidian]

No not really.. I'm just saying though you may start using it heavily and get used to that is all

That and all your data belong to China 😄

[pliant tangle]

Well that's a problem for me lol

[gilded obsidian]

I've used almost 1 billion tokens this month

haven't even broke $20, but I also use flash where appropriate

[eternal tiger]

So I just went through $26 in 3 days with Claude...what a waste

[gilded obsidian]

$19.32
USD ``` ```deepseek-v4-pro
API requests
2,523
Tokens
199,725,733 ``` ``` deepseek-v4-flash
API requests
5,852
Tokens
818,059,581 ```

(top one is pro sorry it cut it off)

[ionic wren]

just save yourself the headache and just get a codex subscription

opencode go is good too

but I haven't used their models for anything other then running tests

[pliant tangle]

Joseph brother give a good recommendation

I just have to be careful when the 75% discount ends that's all

[pliant tangle]

Going to switch my agent main model from Gemini 2.5 Flash to DeepSeek. That should fix the rate limit issue. Thanks for the recommendation! Brother

[gilded obsidian]

Well throw $5 at it, try it for a bit, try some other peoples suggestions too, you never know what your going to like 🙂

[pliant tangle]

Thanks brother really appreciate your help

[ionic wren]

I heard people like deepseek as well, w.e. you're jam, opencode go has deepseek too I believe

like a buffet, I like variety

[pliant tangle]

Which exact DeepSeek model do you use as your main model for OpenClaw?

There are a lot version

[gilded obsidian]

I code on huge codebases, and DeepSeek pro has be absolutely awesome.

[tepid kestrel]

hmm that would be a hundred bucks then

[gilded obsidian]

@pliant tangle There's actually only really two now if you read the docs.. flash and pro

The other names are getting depreciated I do believe

[pliant tangle]

Ok bro

[gilded obsidian]

Latest version of the claw knows deepseek out of the box I blelieve, just have to onboard the models

[pliant tangle]

It's great thanks bro

[gilded obsidian]

Dang I should have been a proper salesman and gave you my referral link bahahahha

Glad its working for you though 🙂

[pliant tangle]

Haha

[pliant tangle]

I’ll ask for your help if I run into any more errors. You’re really great at helping brother

[near owl]

use the API goofy

[gilded obsidian]

I have a feeling I'm going to have to let DeepSeek Pro loose on the claw's source code to figure out WHY its not making this model local and where it's getting the 195k context, I'm burnt out man

[severe fulcrum]

Gemini 2.5 is also not known to be a very good model for agentic work

[gilded obsidian]

Maybe my openclaw.json is corrupted in a way that jsonlint and doctor can't see :/

I'm even watching a mofo do it on YouTube and my config looks just fine :/

[vast phoenix]

Is there any website that has all the plugins and skills for openclaw?
might aswell send it to my claw and tlel him implement all the optimizations h3h3

[fiery nacelle]

Clawhub. And then there's various github repos you can find. Also ve careful. Some can be malicious. Its kind of a free for all.

[gilded obsidian]

ahh I just found this you might like it too it shows llama.cpp + TurboQuant https://www.youtube.com/watch?v=5jkAlqbk66A

oops wrong tag @fiery nacelle I meant 😄

[pliant tangle]

@gilded obsidian
Right now I'm running on DeepSeek V4 Flash (openrouter/deepseek/deepseek-v4-flash) via OpenRouter.

My agent completed one task

& Then again : ⚠️ API rate limit reached. Please try again later. 🙂

[jovial current]

can i use claw to get specific key word Twitter posts notification

[gilded obsidian]

Well yeah because your tier even if its paid on OpenRouter proably hs limits... when I say I use DeepSeek I mean I'm buying it direct via the API @ deepseek.com, I don't use openrouter :/

[jovial current]

and is it paid?

[gilded obsidian]

Looked like from the OpenRouter docs there's a few levels of limits, depending on your query it could be doing tons of tool calling and other calls, firing up other agents, lord knows..

[jovial current]

do i have to pay for the ai model?

[pliant tangle]

Ohh I understand it now ok ok

[jovial current]

i can just use local llm

[pliant tangle]

I'm currently newbie in this field

At first I used GPT-5.4 PRO

& In just 45 request my 53$ got vanished

So then I searched the most cheapest on & I found Gemini

[gilded obsidian]

What kind of work are you trying to do? 😄

[jovial current]

i want to notified about a specific key word

[severe fulcrum]

I feel like it’s better to use sub Codex account if you’re limited by budget tbh

[gilded obsidian]

Sorry I was asking @pliant tangle 😄

[severe fulcrum]

You’ll have limits but at least you pay a fixed price every month

[fiery nacelle]

look into subscriptions for codex/gpt or opencode-go, z.ai, ollama, openrouter. API is gonna burn a hole in your wallet. Or run locals if you can

[pliant tangle]

Play — I was enjoying teaching my agent new skills and many more things.

[gilded obsidian]

I used z.ai for awhile, was half way decent and cheap

[hollow halo]

Can I connect in my phone

[unkempt cypress]

hello

[gilded obsidian]

eh? To your OpenClaw? Configure Discord/Telegram whatevers...

[unkempt cypress]

Am i blind or am i just not getting openclaw web gui at all. there is no button to add models or ollama ?

[gilded obsidian]

I even supports some voip providers I believe

@unkempt cypress the "gui" imho could use a ton of work, you can add models but I find it tedious that way

[hollow halo]

Please can I install and connect to telegram mini game for free

[unkempt cypress]

yeah gui needs a toooooon of work that i can agree on couse i cannot find even the most basic of things

i am running gemma 4 on my macmmini on openclaw, and i got another macmini running ollama and i wanna connect it as a 2nd agent
but alas that do not seem to be possible

opencclaw seems to be one machine only, not multi machines

i asked openclaw itself and it said "i am a manager and gate way sure we can connect it." but well do not seem to work that well 😛

i tried adding 2nd machine to openclaw.json but that did not work

[toxic elk]

I tried using gpt-image-2 to create a GIF, but it failed; it seems there is still a lot of room for improvement when it comes to generating animated images

[unkempt cypress]

opencclaw seems to be catered more to a single machine running 128gb vram rather than many machines running 16-32gb vram gpus

so the term gateway seems to be awfully incorrect

[toxic elk]

I successfully got the OpenClaw Agent on my Mac mini to remotely control my laptop, so I imagine it should also be possible to have the Agents on both devices communicate with each other

[unkempt cypress]

i just want to have the both mac minis to work togheter on a problem.

[green vigil]

Hello, I am working on a tool for openClaw, and I need to interface a custom sqlite database. What is the way to go please ?
Is there any native solution ? or use db.js ? or rebuild openclaw to add build and better-sqlite3 ?
Thank you for your help

[ebon atlas]

What else is openclaw if not a gateway? Youre using OpenClaw as a Gateway to use the language models, its an entrypoint to usage of things like codes,claude,gamma,etc

[unkempt cypress]

it can only handle 1 language model not many

or at least i cannot find a way to connect more agents to it. it does not seem to allow it

openclaw.json following the homepage did not work.
and in the webgui there is no connector tab

[ebon horizon]

What version you on

[ebon atlas]

4.23

[ebon horizon]

Im on the latest now

[ebon atlas]

is it good?

[ebon horizon]

Its actually stable lol. Feels great yeah

The speed is back to if not faster than 4.23. Seriously

[unkempt cypress]

im 0n 2026.5.7

[ebon horizon]

My codex models love it. No death loops, no delays, no crashes. No logs looking like essays

[ebon horizon]

I think you were waiting for the next stable model

Id say this is it

[gilded obsidian]

This is driving me nuts I cannot for the life of my figure out why this is not taking my context number 🙁

[ebon horizon]

?

[gilded obsidian]

I'm trying to use Qwen 3.6 locally, have the model up and running, define it in my openclaw.json, I do openclaw model list and it shows it as 195k context when I clearly have 32k defined. It also doesn't know its a "local" model either

[ebon horizon]

Context window set inside openclaw.json?

[gilded obsidian]

        "baseUrl": "http://localhost:8081/v1",
        "apiKey": "llamacpp",
        "api": "openai-responses",
        "models": [
          {
            "id": "qwen3.6-35b-a3b.gguf",
            "name": "Qwen 3.6 35B A3B MoE Local",
            "reasoning": false,
            "input": [ "text" ],
            "contextWindow": 32768,
            "maxTokens": 8192,
            "compat": {
              "requiresStringContent": true
            },
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            }
          }
        ]
      } ```

[ebon horizon]

Yeah tbh thats what id have said if youve already done rbat aha

[gilded obsidian]

Looks perfect heh

[ebon atlas]

Btw you can change models with just an Openclaw config command, i dont know what you want to do exactly, but you can also just tell your openclaw to create more agents... and then they pop up in ur agents tab to edit.

you can also use things like Venice with it to get routing selection based on your prompt

[ebon horizon]

I believe theres also a global value too

[gilded obsidian]

it'll use the model for a bit and bomb (probably because its not compacting)

[ebon horizon]

Global might be overriding

[gilded obsidian]

Theres NOTHING for 195k anywhere in my files.

[unkempt cypress]

thats similar how my "normal gemma4" looks but if i add another ollama after it, it wont work

[ebon horizon]

Have you got codex?

[gilded obsidian]

I don't want codex 😄

[echo wing]

yo guys,
been a longer user of telegram as a window to chatting with the openclaw agent.
now running into the issue of securing a direct separate window to a sub agent, and as i understand discord doesnt have issues with setting this up.
My question is to those that used both or are aware of other situations where discord is ahead of telegram usage?

[ebon horizon]

Let me ask my claw aha

[gilded obsidian]

great after HOURS of fucking with this and it now just automagically works?!?! Let me reboot gateway and see if it sticks, I have NO idea what I changed.

[ebon atlas]

Did you swap ur gateway port to 8081? if so , check if you have that port blocked in your VPS, may be why, just an idea

[gilded obsidian]

Now I gotta see if it bombs or compacts before context is hit

[unkempt cypress]

im not even sure my openclaw is aktually working. i get status updates on % completion but i see no files being generated or any content. 😮

[ebon horizon]

Live settings are in one file:

• ~/.openclaw/openclaw.json
  • per-model native windows:
    • models.providers.minimax.models[*].contextWindow
    • models.providers.openai-codex.models[*].contextWindow
  • global effective cap:
    • agents.defaults.contextTokens

Whats your

agents.defaults.contextTokens set to @gilded obsidian

[ebon atlas]

just follow a Multi Agent tutorial on youtube

[gilded obsidian]

deepseek/deepseek-v4-flash                 text       977k        no    yes   default,configured,alias:DeepSeek
deepseek/deepseek-v4-pro                   text       977k        no    yes   configured,alias:DeepSeek Pro
zai/glm-5.1                                text       198k        no    yes   configured,alias:GLM
nvidia/nemotron-3-super-120b-a12b          text       195k        no    yes   configured,alias:NVIDIA
nvidia/minimaxai/minimax-m2.5              text       192k        no    yes   configured,alias:MiniMax
nvidia/z-ai/glm5                           text       198k        no    yes   configured,alias:nvGLM5
nvidia/moonshotai/kimi-k2.6                text       256k        no    yes   configured,alias:Kimi
nvidia/meta/llama-4-maverick-17b-128e-i... text       128k        no    yes   configured,alias:Llama-4
nvidia/meta/llama-3.3-70b-instruct         text       128k        no    yes   configured,alias:Llama-70B
nvidia/qwen/qwen2.5-coder-32b-instruct     text       128k        no    yes   configured,alias:Qwen-Coder
nvidia/qwen/qwen3-coder-480b-a35b-instruct text       128k        no    yes   configured,alias:Qwen-Coder-480B
nvidia/qwen/qwen3.5-122b-a10b              text       128k        no    yes   configured,alias:Qwen-122B
nvidia/google/gemma-3-27b-it               text       8k          no    yes   configured,alias:Gemma-3
nvidia/meta/llama-3.2-11b-vision-instruct  text+image 128k        no    yes   configured,alias:Vision
nvidia/deepseek-ai/deepseek-v4-pro         text       256k        no    yes   configured,alias:nvDeepSeek
llamacpp/qwen3.6-35b-a3b                   text       32k         yes   yes   configured,alias:Qwen3.6 Local ``` I have no clue what I just changed but now it works :/

heh, like wtf 😛

[unkempt cypress]

all of them are on same local machine so it does not seem to help me i am on many machines trying to connect them to the gateway

[gilded obsidian]

Lets see if it works on a long query

[ebon horizon]

Why would you need an 8k window btw?

Are you using the local model for crons only or something?

[gilded obsidian]

I'm doing testing.. it was bombing, I can actually do 256k 🙂

So I lowere everything just to see

[ebon horizon]

Compaction issue maybe?

[gilded obsidian]

Well yeah it was stuck on 195k context!

[ebon horizon]

Pre compaction setting

[gilded obsidian]

Even though I was specifying 32k

(in both the config and llama.cpp for testing)

Turbo Quant is frickin amazing

[ebon atlas]

"Run OpenClaw/node on each machine
Point every machine at the same gateway URL
Pair/authenticate each machine separately
Give each machine a unique name/label so you can tell them apart
Make sure the gateway is reachable from every machine"

            ┌──────────────────────────┐
            │      OpenClaw Gateway     │
            │  central routing/auth/UI   │
            └─────────────┬──────────────┘
                          │
      ┌───────────────────┼───────────────────┐
      │                   │                   │

┌─────────▼─────────┐ ┌───────▼─────────┐ ┌──────▼──────────┐
│ Machine A │ │ Machine B │ │ Machine C │
│ OpenClaw node │ │ OpenClaw node │ │ OpenClaw node │
│ local files │ │ local files │ │ local files │
└───────────────────┘ └─────────────────┘ └────────────────┘

You probably need a public domain gateway or SSH each machine into the gateway to be able to access multiple machines

[unkempt cypress]

you need to run openclaw on all machines ??? in their documents they said you just need ollama to be running ?

[gilded obsidian]

I mean the gateway tells you how to make an ssh tunnel when you run "openclaw dashboard" even 😄

It sounded to me like he's trying to cluster multiple machines to use the GPU on all of them

As he had multiple 32GB macs, not a huge 128GB mac or something?

[ebon atlas]

where does it say you need Ollama to access another machine, show me.

ofc you need openclaw to acces an openclaw gateway...

[cinder sphinx]

he wants to use model that is running on other mac, so he can have multi models at once

[gilded obsidian]

so just add that model as a source?

couldn't be anymore simple :/

[ebon atlas]

no hhes not because i clearly stted that and he said no i want multiple machinees to conneect to the same gateway, unless he dont know wtf he wants.

[gilded obsidian]

lol I'm confused too 😄

[cinder sphinx]

it does seems he has some routing issue, cos it should be simple, i have TTS/STT docker on one machine, that openclaw on another machine is using without an issue

[gilded obsidian]

hook em up all to a private discord, link models from each machine however you need to, profit? 😛

[unkempt cypress]

https://openclaw-ai.com/en/docs/providers/ollama/

it says nothing you need multiple gateways ???

[cinder sphinx]

@unkempt cypress are both macs on local network?

[ebon atlas]

He wants this - 1 openclaw dashboard - Tell it a prompt llike "search all my machines for security breaches" so 1 gateway access and controls multiple machines..... right?

[gilded obsidian]

@unkempt cypress What exactly are you tring to do, make it CLEAR to us 😄

[ebon atlas]

im trying to dumb it down as much as i can lol

[cinder sphinx]

say he has qwen on main openclaw and gemma on othe rmac, he wants to be able to say, use gemma to do this and that, and qwen openclaw is using gemma then?

[unkempt cypress]

extremly simple.
I got 2 mac mini.
1 mac mini running openclaw and ollama with gemma4 ( working )
1 mac mini with ollama qwen3.5
I want to be able to use both mac mini models on the open claw gateway

[gilded obsidian]

Let him get it out or we're gunna go crazy guessing

[cinder sphinx]

easy

[gilded obsidian]

@unkempt cypress so y ou just add the model using http://192.168.333333.33333/ etc etc

[cinder sphinx]

you need to add other mac in config with model it runs

[ebon atlas]

route both to the same gatewaay

like quite literally just point them to the same localurl

[unkempt cypress]

where ??? there is no button and openclaw.json do not seem to work

[cinder sphinx]

just tell your openclaw to do it, give it IP:port and what model it is

[unkempt cypress]

i did, all it said was api succesfull but i cannot see the model in gate way nor can i do anything

and the gpu on mac mini aint being used

[ebon atlas]

aare they both on the same network im guessing right? then you point both machines to the same gateway url and then you can just select which model you want to talk to and tell it shit like "talk to the other model"

[cinder sphinx]

tell it then, you want that mac to work as sub agent or something, so you can tell when it does work

[unkempt cypress]

yeah i can query the machine in the terminal from the "openclaw" mac terminal

if i call for v1 tags on the terminal it shows up so i know they are connected

but openclaw refuse to add it to the webgui

[gilded obsidian]

Holy bejeezus it passed the "make a snake game for me in rust", now to up the context to 128k and try again 😄

[ebon atlas]

SSH pair from 1 machine to the other or tailscale, maybe it just doesnt allow 2 of same directories

[gilded obsidian]

@unkempt cypress what does "openclaw models list" shoiw ?

[unkempt cypress]

gemma4 latest

[gilded obsidian]

and thats it?

[unkempt cypress]

yes

[gilded obsidian]

Then it's not configured properly 😄 Tell it to add it, it should be able to handle it just fine as long as your gemma model is working

"Add ollama local model using http://whatever.ip.here it's the qwen3blahblahblah-model. Add it, and test it, also test it via curl too and tell me the results"

[cinder sphinx]

https://rentry.co/ni4h7vgs

my openclaw says this^

[unkempt cypress]

all openclaw says is "api succesfully tested" but then thats it it refuses to add it to webgui

[gilded obsidian]

oh yeah you might mention it's a mac 😄

Cause I'm PCing it up over hurrr don't know nuttin bout the mysterious ways of the mac

pure nvidia power oozing 😄

[unkempt cypress]

i wrote "add qwen3.5:4b model on 192.168.1.210 as sub agent in openclaw webgui so we can use it"
i get reply "api succesfully tested"
nothing more no update in web gui even if i restart gateway nothing .

[gilded obsidian]

dont mention webgui

[cinder sphinx]

@unkempt cypress copy paste my rentry to your openclaw, it will do the rest

[gilded obsidian]

tell it openclaw.json

(maybe gemma isn't as smart as DeepSeek) 😉

Can gemma do tool calling easily? I've never tried it actually

[unkempt cypress]

let me check

[gilded obsidian]

Do you see it running tools in the webui's chat feature?

Or is it just lying out its lil AI bootyhole? 😄

"Yeah boss I checked the API, no problems over here man" (0 tool usage)

meanwhile I'm over here like: Brachyura@Brachyura:~/.openclaw$ cp openclaw.json openclaw.json.finally.fucking.working

LOL 😄

[vivid crystal]

Bro

[gilded obsidian]

Yessssss?

[unkempt cypress]

ok so i copy pasted the promps added the ip and stuff, now i get answer "we are doing something more powerful we are useing api connector to call the ollama as a thinking agent" yet the other mac mini gpu aint doing squat xD

[gilded obsidian]

Just add it manually

[unkempt cypress]

i added it to openclaw.json

[gilded obsidian]

be one with the JSON, this is the way.

did you "openclaw gateway restart" ?

[unkempt cypress]

dont work

[vivid crystal]

Now Im using the Deepseek ai by depositing 5$

As you said no open router

But still api limit after 2 request

[unkempt cypress]

yes openclaw gateway restart after adding it

[vivid crystal]

@gilded obsidian

[gilded obsidian]

@pliant tangle are you SURE your using the correct one and not the openrouter one still? I have NEVER had a limit via deepseek.com

[unkempt cypress]

{
models: {
providers: {
ollama: {
baseUrl: "http://192.168.1.210:11434",
apiKey: "ollama-local",
api: "ollama",
models: [
{
id: "qwen3.5:4b",
name: "qwen3.5",
reasoning: false,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 8192,
maxTokens: 8192 * 10
}
]
}
}
}
}

[gilded obsidian]

is it just me or is he missing a literal fuck-ton of " " ?

Or is that mac-land?

[unkempt cypress]

its copy paste from openclaw wiki

[vivid crystal]

I asked my agent

" Whats your default model now ? "

Agent reply:

Right now I'm running on DeepSeek V4 Flash (deepseek/deepseek-v4-flash) — looks like it's switched to direct DeepSeek API instead of going through OpenRouter this time.

[gilded obsidian]

Also it's most likely http://192.168.1.210:11434/v1 or something

[unkempt cypress]

tried that too

[raven sundial]

does deepseek have a token plan yet?

[gilded obsidian]

@unkempt cypress Again I'm no mac expert... but mine look like this: "llamacpp": { "baseUrl": "http://localhost:8081/v1", "apiKey": "llamacpp", "api": "openai-responses", "models": [ { "id": "qwen3.6-35b-a3b", "name": "Qwen 3.6 35B A3B MoE Local", "reasoning": false, "input": [ "text" ], "contextWindow": 32768, "maxTokens": 8192, "compat": { "requiresStringContent": true }, "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 } } ] } }

Notice how yours is missing " " around each key/value? Is that normal? I doubt it?

[unkempt cypress]

funny part is, the local ollama works just fine.. and i can api call the remote ollama in ollama and terminal.... but openclawn cant do anything it seems

[unkempt cypress]

https://openclaw-ai.com/en/docs/providers/ollama/

[raven sundial]

did you try the ollama launch command

ollama launch openclaw --model <model_here>

[unkempt cypress]

yeah it works for the locacl ollama

not the remote one

[gilded obsidian]

does it work if you curl the remote one?

[unkempt cypress]

yes i can curl it and it works

[gilded obsidian]

What if you use the onboarding process to define the local model?

[raven sundial]

yeah

[unkempt cypress]

it works locally. not remote ly

openclaw cannot remote connect. but terminal can

[gilded obsidian]

re-run onboarding (especially if this is a new install and you have nothing to lose) and see if you can add it during that process

[unkempt cypress]

maybe its a limit in openclaw macc version it only allow local ai not remote ai

[gilded obsidian]

no if it was a tcp/network issue then it wouldn't work via curl

[fiery nacelle]

base url needs to point to what that machine is broadcasting for that model to connect

[raven sundial]

btw is the 20$ plan of ollama good

[vivid crystal]

@gilded obsidian bro ?

[gilded obsidian]

yeah he is missing /v1 I think?

[unkempt cypress]

dont know i am useing local ollama

[raven sundial]

why are you asking about cloud ollama then

[unkempt cypress]

tried with v1 and without v1 same response

[raven sundial]

is it openclaw connecting to a server to access it from anywhere or is it ollama cloud services?

[unkempt cypress]

i am not. i am asking for this!https://openclaw-ai.com/en/docs/providers/ollama/
scroll to local section

[fiery nacelle]

Ollama on a separate machine that they are trying to connect too

[gilded obsidian]

@pliant tangle and you say your still getting a limit??? Do you have multiple ways your chatting with it, perhaps its using the old model on some?

[vivid crystal]

Yes still saying ap limit & I only chat in telegram with my agent bot

[gilded obsidian]

is it a telegram api limit perhaps?

Sorry I only use discord at the moment :/

[unkempt cypress]

i got open claw + ollama on one machine and ollama on another and want ollama on both machines to work in same openclaw

[gilded obsidian]

@unkempt cypress Yeah its normally easy-peasy that is why I'm struggling to see why its not working

[cinder sphinx]

i have that working @unkempt cypress, what i gave you should be enough to do it

[vivid crystal]

Ahh well a good question , I'm checking

[fiery nacelle]

One issue is the fact that its 2 instances of 1 provider. So both need to be explicitly defined.

[cinder sphinx]

exactly, he wants them as 2 agents in one webui

which is possible

[fiery nacelle]

Not 2 agents, just 2 provided models.

Correct?

[vivid crystal]

Btw my bot is sending 31 Api request in 3 min

It's normal?

[cinder sphinx]

nope

[raven sundial]

nope

[unkempt cypress]

this is why im mind boggled why openclaw do not have the button to simply "click here to add a model" xD that would proboly be feature 1 you program in a ai xD

[raven sundial]

i used openclaw for like an hour and got to 100 requests lol

[vivid crystal]

Well I just checked it's 2 min

31 Api request in 2min

[raven sundial]

what provider

[unkempt cypress]

eaither way is fine. aslong as it works xD

[fiery nacelle]

yea its a bit rough, but it can be manually done in the openclaw.json

[vivid crystal]

Deepseek ai v4 flash

[unkempt cypress]

i followed documentation to the letter and it does not work

[gilded obsidian]

@fiery nacelle What was your settings again for your qwen? 😄

[raven sundial]

nice, but its not normal to have 31 requests in 2 minutes. maybe your api key got stolen idk

i reached 100 after an hour of work in minimax

[fiery nacelle]

its cause you have 1 Ollama local, and 1 Ollama remote. so they need to be added as separate providers with separate base urls etc.

[raven sundial]

chatgpt can probably help you a lot more

[fiery nacelle]

      "lmstudio": {
        "baseUrl": "http://127.0.0.1:1234/v1",
        "apiKey": "LM_API_TOKEN",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen/qwen3.6-35b-a3b",
            "name": "Qwen3.6 35B A3B",
            "api": "openai-completions",
            "compat": {
              "requiresStringContent": true
            },
            "reasoning": true,
            "input": [
              "text",
              "image"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 262144,
            "contextTokens": 200000,
            "maxTokens": 8192
          },```

[gilded obsidian]

ty sir

[vivid crystal]

Ok I'm changing all my keys everything then

[unkempt cypress]

i pasted ( something similar to this ) in my openclaw.json below the "local" ollama at very end

{
models: {
providers: {
ollama: {
baseUrl: "http://ollama-host:11434",
apiKey: "ollama-local",
api: "ollama",
models: [
{
id: "gpt-oss:20b",
name: "GPT-OSS 20B",
reasoning: false,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 8192,
maxTokens: 8192 * 10
}
]
}
}
}
}

[raven sundial]

if it still uses that much requests after resetting, it might be because of a broken plugin

[unkempt cypress]

it did not work

[vivid crystal]

I think it's better if I'm just erase everything & start from zero again

[unkempt cypress]

obviously i changed the numbers and model id :p

[gilded obsidian]

@fiery nacelle You are running that local WITHOUT Turbo Quant?

[pliant tangle]

Well I can speak again huh , I got muted for no reason

[gilded obsidian]

@fiery nacelle how many bits? 4?

[fiery nacelle]

does the machine with OpenClaw + Ollama work with that specific set up? Like you can use Ollama on that machine just fine, correct? And you are having issues connecting to the Ollama on the other machine? And is that second Ollama broadcasting over your network, and not just on that machines local host?
Sorry, just trying to confirm everything

[gilded obsidian]

Also have you tried using images, It is a MoE model so.. I suppose it can do it I just never added/tested that yet

[fiery nacelle]

Q4K_M

[gilded obsidian]

Ah.. I see

[unkempt cypress]

machine with opencclaw and ollama works yes. and i can remotely curl the other machine ollama for tags and querys. only thing not working is openclaw refuses to add it to webgui

[gilded obsidian]

I'm using Qwen3.6-35B-A3B-UD-Q6_K_XL.gguf

ok I swear to SALDFKJL@#$J%LJKLASGJNLQ#$NM%KL:@#$MN I added 256k context and now it lists it like this: llamacpp/qwen3.6-35b-a3b text+image 195k/256k yes yes configured,alias:Qwen3.6 Local

WHERE is it getting 195?

This is driving me NUTS and why does it say 195/256 now?!

deepseek/deepseek-v4-flash                 text       977k        no    yes   default,configured,alias:DeepSeek
deepseek/deepseek-v4-pro                   text       977k        no    yes   configured,alias:DeepSeek Pro
zai/glm-5.1                                text       198k        no    yes   configured,alias:GLM
nvidia/nemotron-3-super-120b-a12b          text       195k        no    yes   configured,alias:NVIDIA
nvidia/minimaxai/minimax-m2.5              text       192k        no    yes   configured,alias:MiniMax
nvidia/z-ai/glm5                           text       198k        no    yes   configured,alias:nvGLM5
nvidia/moonshotai/kimi-k2.6                text       256k        no    yes   configured,alias:Kimi
nvidia/meta/llama-4-maverick-17b-128e-i... text       128k        no    yes   configured,alias:Llama-4
nvidia/meta/llama-3.3-70b-instruct         text       128k        no    yes   configured,alias:Llama-70B
nvidia/qwen/qwen2.5-coder-32b-instruct     text       128k        no    yes   configured,alias:Qwen-Coder
nvidia/qwen/qwen3-coder-480b-a35b-instruct text       128k        no    yes   configured,alias:Qwen-Coder-480B
nvidia/qwen/qwen3.5-122b-a10b              text       128k        no    yes   configured,alias:Qwen-122B
nvidia/google/gemma-3-27b-it               text       8k          no    yes   configured,alias:Gemma-3
nvidia/meta/llama-3.2-11b-vision-instruct  text+image 128k        no    yes   configured,alias:Vision
nvidia/deepseek-ai/deepseek-v4-pro         text       256k        no    yes   configured,alias:nvDeepSeek
llamacpp/qwen3.6-35b-a3b                   text+image 195k/256k   yes   yes   configured,alias:Qwen3.6 Local ```

I think I just need to go to sleep before I yeet it 😄

            "contextTokens": 200000,
            "maxTokens": 8192, ```

Clear as day, no idea where it's getting 195k

Apparently my crab can't maths

[fiery nacelle]

Use this as an example:

  auth: {
    profiles: {
      "ollama-local:default": {
        provider: "ollama-local",
        mode: "api_key"
      },
      "ollama-remote:default": {
        provider: "ollama-remote",
        mode: "api_key"
      }
    }
  },

  models: {
    providers: {
      "ollama-local": {
        api: "ollama",
        baseUrl: "http://127.0.0.1:11434",
        apiKey: "ollama-local",
        models: [
          { id: "gemma4", name: "gemma4", input: ["text"] },
          { id: "llama3.3", name: "llama3.3", input: ["text"] }
        ]
      },

      "ollama-remote": {
        api: "ollama",
        baseUrl: "http://YOUR_IP:11434",
        apiKey: "ollama-local",
        timeoutSeconds: 300,
        models: [
          { id: "qwen3.5:32b", name: "qwen3.5:32b", input: ["text"] },
          { id: "qwen2.5vl:7b", name: "qwen2.5vl:7b", input: ["text", "image"] }
        ]
      }
    }
  },

  agents: {
    defaults: {
      model: {
        primary: "ollama-local/gemma4",
        fallbacks: [
          "ollama-remote/qwen3.5:32b"
        ]
      },

      models: {
        "ollama-local/gemma4": {},
        "ollama-local/llama3.3": {},
        "ollama-remote/qwen3.5:32b": {},
        "ollama-remote/qwen2.5vl:7b": {}
      }
    }
  }
}

[pliant tangle]

I deleted all the previous keys and then created a new DeepSeek key. Then I applied it, but I’m still getting the same API limit error I was getting when I was using OpenRouter as the provider.

[gilded obsidian]

@fiery nacelle My 3.6 model is a MoE but when asked to gen an image I get ``` I don't have a built-in image generation model, but I can do a few things:

ASCII art cat — quick and dirty, right here,
Generate one with a tool — I can use the Vulkan vision skill or call an image generation API if you have one set up,
Write a script that pulls a cat image from an API (like thecatapi.com) ```

Does yours do it?

[pliant tangle]

@gilded obsidian bro any comments? What I should do now ?

[fiery nacelle]

I havent "generated" any images with my Qwen, I'm pretty sure it can only injest images.

[gilded obsidian]

@pliant tangle I am very confused as to what's happening to your instance.. I dunno man I never had any issues :/

@fiery nacelle ooo let me try that.. mabye it can describe?

[fiery nacelle]

yup it can understand images. not generate.

[pliant tangle]

Ok

[gilded obsidian]

@fiery nacelle all I need anyways, who needs AI slop artwork am-i-rite? 😉

Darn mine won't describe.. " Something went wrong while processing your request. Please try again, or use /new to start a fresh session."

very strange... I'm about to give up on a local model that's worth anything at this point :/

if it's not one thing, it's another it seems, not ready for prime time 😄

[fiery nacelle]

I really hate that... its so annoying, sometimes you can reply to it, and it might work again, also make sure you have input: ["text", "image"] defined under the model

[gilded obsidian]

Yeah I did....

It's just super unstable apparently

general.tags [qwen3_5_moe, qwen, image-text-to-text]

huggingface shows it as such though, it should work

I wonder if instead of "image" we try "image-text-to-text" ?

[fiery nacelle]

in the openclaw.json it has to be defined accordingly, not what hugginface states

only "image" and "text" are valid.

[gilded obsidian]

well poo poo on that 💩

I'm going to try a 5-bit IQ Unsloth see if there's any change, I don't think it's the model not loading correctly though

[unkempt cypress]

tried it but same thing first ollama shows up but 2nd one does not

but i curl and i get
StatusCode : 200
StatusDescription : OK
Content : Ollama is running

[gilded obsidian]

so ollama models list only shows the local ?

[unkempt cypress]

no i can curl it remotely and query it in terminal, just not in stupid openclaw xD

[gilded obsidian]

what's ollama models list say

[unkempt cypress]

aloot of models thats installed on the remote host and local

[gilded obsidian]

ok and when you switch model to the remote one, how are you switching

[unkempt cypress]

terminal since openclaw cant find it ?

[gilded obsidian]

how are you interfacing with openclaw

web gui chat?

[unkempt cypress]

local is http://127.0.0.1:11434
while remote is well u know

openclaw is web gui or terminal yes

[gilded obsidian]

ok so your using the openclaw dashboard, going to chat, and you have a dropdown of models, but you are saying the remote is not there?

[unkempt cypress]

correct locacl gemma4 is there not remote

[gilded obsidian]

and in terminal your doing "openclaw" "talk to agent" and "/models" and you don't see it there?

[fiery nacelle]

did you restart your gateway after making the config changes?

[gilded obsidian]

I'm starting to think he's ssh'd into another random machine on his network editing config files like a mad man

[unkempt cypress]

ccorrect

[gilded obsidian]

lmao 😉 But I can't say much! mines doing weird shit like that 195k context that makes no sense!

[unkempt cypress]

nope 😛 2 keyboards 2 displays 😛

[fiery nacelle]

1 keyboard, 3 systems, 6 displays xD

[gilded obsidian]

have you thought about rm -rf ing the dir, and making a fresh claw, and using the onboard process to add the remote model first

[unkempt cypress]

when i do curl x post on the machine to the remote machine useing http://192.168.1.210:14434/api/
i get a real working repsonse and the remote ollama works.
so my question is why the stupid openclaw wont add it to frigging dashboard xD even openclaw diagnostic confirms its working xDXDXDXDXD

[fiery nacelle]

to be fair, 2 "displays" are tablets lol

[gilded obsidian]

imgur a screenshot, show us where your looking for it in the dashboard

well I guess it doesn't matter if its not listed in openclaw models list

[unkempt cypress]

im in the chat section, and click model

and in bottom left it says version 2026.5.7

local ollama works just fine, and it can curl the remote ollama just fine

it just refuses to add it to the dashboard thats the annoying part

all becuase developers of openclaw cannot add a stupid button "add model or provider" xD

[gilded obsidian]

hey now.. 😛

if its defined and enabled for the agent you currently have selected it should be in the dropdown

[unkempt cypress]

only local one gemma4

when i ask openclaw to modify the openclaw.json and add the remote ollama
it says SUCCCESS but nothing happens even if you restart gateway xD

i dunno i think i gonna call it quit and install hermes and try taht instead

[gilded obsidian]

While it's running what does ~/.openclaw/agents/main/agent/models.json contain

does IT contain the remote model?

[unkempt cypress]

le me check

[unkempt cypress]

gemma 4 local only

[gilded obsidian]

ok so...

is merge on?

[unkempt cypress]

not sure what that is,

[gilded obsidian]

    "mode": "merge",
    "providers": {```

[unkempt cypress]

let me look

[gilded obsidian]

In OpenClaw, the --merge flag is used in the openclaw config set command to add new models to your agents.defaults.models configuration without deleting existing entries, allowing for additive configuration changes. It protects against accidental removals, ensuring new settings are combined with existing ones rather than replacing them completely

it kinda works like that

turn on merge, restart gateway, is model in both places?

[unkempt cypress]

that part is not there no

[gilded obsidian]

if so buy me a lambo

[unkempt cypress]

soo how do i set the merge flag on ?

add it to openclaw.json ?

[gilded obsidian]

I showed you right there 😄

yeah put it right in that spot

[unkempt cypress]

but that spot does not exist though

[gilded obsidian]

technically you can use the webgui, but again.. it blows

[unkempt cypress]

mine looks like this

{
models: {
providers: {
ollama: {
baseUrl: "http://ollama-host:11434",
apiKey: "ollama-local",
api: "ollama",
models: [
{
id: "gpt-oss:20b",
name: "GPT-OSS 20B",
reasoning: false,
input: ["text"],
cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
contextWindow: 8192,
maxTokens: 8192 * 10
}
]
}
}
}
}

[gilded obsidian]

so add it? 😄

if you go to the Models page in the gui its literally the first option

[unkempt cypress]

lets see

[gilded obsidian]

AI & Agents / Models

[unkempt cypress]

do not have that no. in dashboard i have agent agents but not model

in search window i typed merge nothing showed up

[gilded obsidian]

and "Models" is selected?

[unkempt cypress]

https://www.shambix.com/wp-content/uploads/2026/02/openclaw_chat-1.png.webp

mine look like this

[unkempt cypress]

models only exist in chat window

[gilded obsidian]

scroll down

left side

[unkempt cypress]

little annoying we cant post pictures

models do not exist on left side

models only appear in chat window

[gilded obsidian]

https://imgur.com/a/oOXj1Lk

AI & Agents, Models...

[unkempt cypress]

found it

[gilded obsidian]

Technically I guess its Settings -> AI & Agents -> Models 😄

[unkempt cypress]

merge is on there yes

[gilded obsidian]

weird, take your mac back and demand Tim Cook give you a refund

seriously you should start over and use the onboarding process

instead of that document wherever it misled you

[unkempt cypress]

tried to add it there. as well same as local one but dident work

[plain geyser]

is there any way to put openclaw on Meta glasses?

[unkempt cypress]

yeah im gonna format the mac and just try hermes instead. openclaw is still in early alpha it seems

ty for help to bad it dident work 🙂

[pliant tangle]

@gilded obsidian brother no more API Rate limits , I fixed it

[lean bramble]

Hey! Would you say it’s normal that OpenClaw takes longer than normal ChatGPT or Codex for an answer?

[fiery nacelle]

yes. largely due to the fact that OpenClaw is a harness that communicates back and forth between the model in a way. Theres more over head than if you were using GPT or Codex in a browser.

[merry helm]

You’re going to have to build it methinks. Point your agent here - https://wearables.developer.meta.com

[lean bramble]

Hm, yeah. So this is the same with Hermes, right?

My agent takes about 20 seconds to answer a very simple questions with GPT 5.5

[fiery nacelle]

Yes, Hermes is also a harness. GPT/Codex are models.

[near owl]

can I ask krill for assistance regarding brainstorming for myself?

[merry helm]

You can see the overhead in logs/gateway.log. i have ~100 ms for sessions.list and ~50ms for chat.history

The bad perf days got up to 8s for sessions.list

[fiery nacelle]

20 seconds is resonable for most people. Try dealing with a multiple minute delay in some edge cases xD

[craggy bane]

[lean bramble]

Hm. So it’s still better to use Codex and OpenClaw only for things it cannot do?

[near owl]

hello people

[gilded obsidian]

@pliant tangle what was it?????

[plain geyser]

it looks like somebody already did it before me https://github.com/Intent-Lab/VisionClaw

[pliant tangle]

I'm talking about the " ⚠️ API rate limit reached. Please try again later. " I was facing

I fixed it

[gilded obsidian]

@pliant tangle Yes but how?

[pliant tangle]

Wiped all OpenClaw data, fresh reinstall, switched from Gemini via OpenRouter to DeepSeek API directly. Fixed!

That's how did brother

[pliant tangle]

Sorry for late response

[gilded obsidian]

crazy

Nah its ok I'm about to nuke my install too, which really blows I had customized the crap out of it, but it's just not right anymore

[pliant tangle]

By the way, I'm new to this, so I don't really know how it happened.

When I completely cleared my VPS, there should’ve been nothing left, right?

But somehow, after I reinstalled my OpenClaw and set up my bot again, all 13 skills I had taught her were automatically restored within seconds. Every skill was working perfectly.

I honestly don’t understand how that was possible after I deleted everything and reinstalled it.

[ebon atlas]

https://tenor.com/view/gudnaitinis-gif-17703749535812827607

[crisp nimbus]

Bro after recent updates for some reason my messages are sometimes delayed by over 10 minutes I’m still trying to figure out why. My hardware is ancient 2010 iMac running macOS Monterey thanks to opencore legacy patcher, but it used to work decently before, now I have weird issues… like for example my main agent works pretty good, but my secondary agent (not a sub agent) randomly takes super long to respond…

[covert lance]

Guys

What's the best way to give ssh access to llm?

Just telling them to use SSH force them to type sshpass every time which seems very inefficient

[crisp nimbus]

Probably ssh keys. But ask your agent it probably has a better idea what would be easier for it

[ebon atlas]

You can create a command or have a cron job that keeps it at the top of memory . Create the ssh key pair. Save it, via Google Drive or something , however you get it to OpenClaw, then tell your OpenClaw to create command for ssh key pair that pulls up whatever directory and file name that you have it saved under.

[covert lance]

Also do anyone code alongside LLMs agents? What's the best setup? Do you just give llms a GitHub account and tell them to send pull requests?

It's kind of annoying that LLMs don't just get given a stateful terminal

[ebon atlas]

For example “Hey claw can you connect to MyHomePC using /sshconnect command” or just “/sshconnect “ in the webui or whatever

[covert lance]

Yes but then it will just type every command prefixed by sshpass -p ...

Which is just waste of tokens waste of context etc

[crisp nimbus]

Ask it to make a simple shell script which will do that and then it just needs to run that script when needed.

[ebon atlas]

Hmmm is have my ssh key pair connected from my PC to the VPS and whenever I use the ssh terminal from my PC , it doesn’t seem to throw the ssh commands in it. Maybe it’s different🤔

[covert lance]

Like every terminal command they send don't persist state of the last

[crisp nimbus]

Oh, well maybe tmux would be the fix for that? I think OpenClaw had a skill or tool for that

[ebon atlas]

Hmmm 🤔 I mean I understand what you mean but damn, idk how to help. What memory u running? Why is it only remembering 1 line of terminal code at a time lol

[covert lance]

It's not about llm memory

[ebon atlas]

Like are you running obsidian or anything like that, or just memory.md , maybe there is some way just have it pull and run the script automatically every time it asks for it. But idk if that even helps, because thinking about it, “run home/openclaw/script.exe” is a longer of a command to run everytime you need it than just doing the ssh pass

[covert lance]

Also how are people setting up development loops on GitHub?

[gilded obsidian]

@fiery nacelle I have news on the Image detection front if you are interested

[ebon atlas]

Imma let the big dawgs help u with that. Because idk 😭 my shi just works, I don’t need a ssh pass in each command line of the terminal

[covert lance]

Ok cool thanks

[shut junco]

Hello all.

Fell in love with this Ai called Viktor then found out that its Openclaw with a wrapper and now i want to build my own. Not sure where to start but i assume openclaw can help me right lol?

[potent bronze]

Hello everyone! I'm trying to use obsidian cli, but the agent never uses it... I've already installed the obsidian skill, but it doesnt work, could someone help me please figuring out the problem?

[near owl]

@dusky cloud how do I get the shell society role?

i guess he doesnt answer questions

[near owl]

@lethal spindle I signed up for inkbox via the agent and the email feature is not working, same with the phone.

[gilded obsidian]

@potent bronze Have you asked the agent to diganose why its not working? I know it sounds silly.. but...

[cosmic prism]

Got this issue with codex ⚠️ All models are temporarily rate-limited. Please try again in a few minutes.

Any one knows the fix?

[fleet thorn]

what's the best model plan for openclaw rn?

[coral otter]

gpt plus

[crisp nimbus]

Tbh gpt plus is not enough for me. I used up the weekly limit already :/

[coral otter]

Wuuutt. How? Im still at around 40%. And i do coding via codex too

[crisp nimbus]

Idk what to tell u. I had codex fix my OpenClaw, and also upgrade 4 custom OpenClaw plugins to the new plugin schema while also refactoring it slightly to follow better practices, and then let it rip on one of my code projects that I’ve neglected for a few years, to update the dependencies and stuff to the latest version and fix any issues it saw.

[coral otter]

I mean there is also the 100$ plan

Just vibecoded the interface for a hardware project im working on and i burned around 15%

[crisp nimbus]

Yeah but right now I got the one month offer for free ChatGPT plus so I’m not gonna upgrade until that’s over at the very least

[coral otter]

Facts, or get 2 20$ ones

[crisp nimbus]

Also I feel like $100 is too much, and I’m probably not gonna actually use that much.. tbh I wish there was like a middle plan. Like $50 a month or something

[coral otter]

2 20$ 🤑

[crisp nimbus]

I guess I can do that but, isn’t it against TOS to use account switchers or something?

I know most other providers don’t want that…

[coral otter]

wää wää wää

[steel vector]

no

openai is pretty chill in that regard

[potent moth]

yeah i switch all the time thru claw and codex and haven't had any issues

[near owl]

i am surprised that a company evil like that is chill in allowing users to break ToS

[steel vector]

is it against TOS?

[near owl]

never checked, but the chat is saying so? Until I just realized you said no.

[steel vector]

peter uses multiple accounts iirc

i saw a tweet like that

[near owl]

I am self-aware (very self-aware) but for some reason I am not self-aware in this server.

What drugs did you put in this server that made me not self-aware

[steel vector]

the opposite of LSD

[near owl]

I do not know what that is.

[steel vector]

DSL

[near owl]

diseal

[steel vector]

i should become a stand up comedian

[near owl]

dammit grammarly messed me up

[steel vector]

lol

[near owl]

stand up on stage and make some jokes then, "stand up comedian"

[steel vector]

nah i’m sitting down rn can’t be funny

[near owl]

why did the variable cross the road?

[steel vector]

idk math

[near owl]

to store it's value in the bank

[steel vector]

i dont get it

[near owl]

variable stores values in memory, memory is the bank.

[steel vector]

👁️👄👁️

[near owl]

thats cute

[steel vector]

😳

[near owl]

and cursed

[steel vector]

its more cursed if you have an android

[sharp idol]

My OpenClaw started requesting that I approve commands, how can I remove that? example /approve 9d2fa1e0 allow-once

[potent bronze]

Man, the agents is taking so much time to send a message, it's all over the place

[gilded obsidian]

What model are you using

[potent bronze]

Qwen 2.5 2b

with ollama, but gemma is the same thing

[severe fulcrum]

The answer is in the docs

[potent bronze]

can I send you a pm @gilded obsidian ?

[gilded obsidian]

I'm going to bed, sorry sir

[severe fulcrum]

Good night

[tawny lake]

has openclaw stop responding for you? Discord, telegram I will ask a question and get no respone

[raven sundial]

yep

it could do tool calls but no response on telegram, i just deleted the config filles and it was fixed

[tawny lake]

deleting config files fixed the no response?

[raven sundial]

yes

[potent bronze]

I'm having the same problem... which files did you remove?

[raven sundial]

literally just deleted the .openclaw folder lol

then did onboard again

[potent bronze]

but with that, you lost everything

[raven sundial]

mine could do things just didnt respond tho, if its not the case for u idk if it will work

[raven sundial]

nah i didnt build a lot

i just made an unlimited free image generator

[sharp idol]

Thanks

[potent bronze]

but did you remove everything from openclaw via npm?

[raven sundial]

nope

[potent bronze]

so, you just removed the folder

[raven sundial]

yes and did onboard again

[humble pendant]

Hey, hope its fine I ask but I'm currently 100% claude user. I run their max sub but I feel i dont get the autonomous part that our friendly lobster gives us but bit worried about the whole claude not liking it being used. Is it actually restricted or just something they say?

Im running a fully autonomous content generator, job hjunter, appfactory and things like that so idle time is my killer and well, i hit that always and lots of stale workers since claude its really built to be a multi worker setup

[pliant tangle]

online bro ?

[plain grove]

Don’t forget to call your mom today if you still can! Get out and spend some time with her even…

[worthy anvil]

Try:

{
  messages: {
    visibleReplies: "automatic",
    groupChat: {
      visibleReplies: "automatic"
    }
  }
}

[radiant prawn]

someone has experience in creating a personal training coach?

[severe fulcrum]

It depends on where you are!

[zealous lagoon]

yeah in my part of the world we ignore our moms during May.

[ebon horizon]

He must be Australian

[zealous lagoon]

¿ʇɐɥʇ ʎɐs noʎ plnoʍ ʎɥM

[severe fulcrum]

It’s the 31st here

[zealous lagoon]

What kinda timezone is that? It's already the 10th here!

[severe fulcrum]

That we celebrate moms

[zealous lagoon]

It's March where my mum is. 10th where my wife's mum is, which is partly why my wife is visiting her rn.

[sharp idol]

Yeah! I successfully setup vLLM with TurboQuant using Qwen. Unfortunatly I'm using the 7B model because I'm renting an A10 server. I needed a server using an X86 processor and not an ARM based processor.

[manic drift]

Is there a decent provider which can give me acccess to both sonnet 4.5 and gpt 5.4 in one go

[sharp idol]

Github Copilot but I think they are not allowing new accounts right now? Or maybe that was in the past.

[manic drift]

It's crashing for anything besides 5.4mini

[sharp idol]

I'm using it right now with GPT 5.4

[manic drift]

I'll have to retry then

[zealous lagoon]

RIP copilot.

[sharp idol]

You using it with OpenClaw?

[sharp idol]

Yeah, it was the best provider. I'm grandfathered in since I pay yearly.

[zealous lagoon]

woah. Do you rent the server permanently or by the hour?

[sharp idol]

By the hour .. it's like $2 an hour.

[zealous lagoon]

Me too but at the new multipliers it wasn't even worth it, I cancelled both my accts

[sharp idol]

Gpt 5.4 is still 1x but yeah 5.5 was a lot higher

[zealous lagoon]

They said opus 4.5 and 4.6 were still 3x but I think they were only available for corporate accounts. I would have stayed for those alone.

[sharp idol]

This sucks, the latest version of OpenClaw keeps asking me for permission for everything. I set the policy to yolo and it is still asking me.

[zealous lagoon]

hmmmm did you get the right setting? sorry I'm not much help but Krill is awake now, you could ask in #1459642797895319552

[ebon horizon]

Theres quite a few settings you got change

[final vessel]

Is 5.9 going to fix most/some of the slack issues I’m hearing. About? We’re using in a production setting and I’m on 5.2 right now. Slack is a bit screwed up with agents forgetting what thread they were in and double answering questions but I see a bunch of new issues popping up post 5.4.

[sharp idol]

Opus 4.5 and 4.6 are gone. Sonnet 4.6 is 1x , GPT 5.4 is 1x , GPT 5.5 is 7x and Opus 4.7 is 15x.

[sharp idol]

According to the documentation the command is "openclaw exec-policy preset yolo"

I may need to restart the gateway to take effect

I wrote a config file and I think I fixed it

[median sandal]

Hysterical.... Julian Goldie already has an "openclaw 5.9 is absurd" video out. 😂😂😂😂

[sharp idol]

Gotta get them views !! Like and Subscribe!

Nope, still asking for approvals. Seems the exec policy is broken.

[uneven willow]

Is 5.10 stable now? Or don't touch it when it's still in beta?

[narrow sandal]

hello

[supple ibex]

I would recommend IsItStable, it’s really good for checking stability. Of course, it won’t be true for everyone, but on my side it has been really accurate. I made a skill for it, really useful

https://github.com/exisz/IsItStable/issues/

[frozen holly]

Hey folks! OpenClaw 2026.5.10-beta.1 is live 🦞

Since 2026.5.7-beta.1:

• 3,570 commits
• 148 contributors
• 6,716 files touched
• 286k insertions / 262k deletions

Highlights:

• Realtime Talk + Discord voice got much stronger: gpt-realtime-2, agent-proxy voice, better barge-in, playback, diagnostics, and talk.realtime.instructions.
• Plugin SDK grew durable channel-message lifecycle APIs, unified model/media catalogs, fs-safe helpers, npm-pack installs, uploaded skill archives, and oc-path.
• Model/provider work: Gemini 3.1 normalization, GitHub Copilot catalog/token fixes, OpenRouter/Kimi/Moonshot repairs, xAI /think, Bedrock service tiers, Mistral Medium 3.5.
• Channels got a lot of polish: Telegram streaming/select fixes, Slack threaded interactions, Feishu topics, WhatsApp/iMessage media replies, QQBot slash commands.
• Control UI is more responsive, with usage lineage, context indicators, Windows WebView2 bridge work, and better exec-policy display.
• CLI/onboarding/help now gives better recovery paths instead of terse dead ends.
• Security/reliability fixes across logging redaction, exec approvals, gateway session kill paths, Docker/browser installs, config persistence, and provider stream hangs.

Beta build, but a chunky one. openclaw update --channel beta

[supple ibex]

Unfortunately, 4.23 is still the only version that works for me. all the other releases caused serious issues.

[frozen holly]

I'd love to know what issues you see with .9/.10, send issues my way

[glass ferry]

No updates last few days. Is everything ok?

[frozen holly]

Dude we been cookin'.

[zealous lagoon]

it's been minutes since the last release. I was starting to worry.

[glass ferry]

Looking forward to it ❤️

[sharp idol]

Ugh, I just can't stop OpenClaw from asking for approvals

[zealous lagoon]

aw man. I'm not updating right now but what version tripped you up?

[sharp idol]

20206.5.7

[zealous lagoon]

weird. i'm on that version too but never seen this bug.

[sharp idol]

Still working with Krill on fixing it.

Hasn't asked permission in a while, I might have it fixed now

[white venture]

How do you get an openclaws agents voice working conversationally on my custom dashboard? it works fine in telegram but cant get it to work on my custom dashboard

[sharp idol]

So, I have this A10 GPU with 24 GB of vRAM running Qwen 7B and a 131k context because of TurboQuant.

[zealous lagoon]

To answer that question I'd have to be intimately familiar with your custom dashboard. Perhaps talk to the author about it.

[raven sundial]

how do yall even build cool dashboard bro mine looks like slop

what prompts or skills yall use i got ui-ux-pro-max and its still trash

[zealous lagoon]

All AI generated output looks like slop. I haven't found any model to be capable of a useful, visually appealing interface. The newer design skills etc are ok at making magazine-layout output but absolutely nothing when it comes to information dashboards and UX. You have to design it yourself and get the bot to wire it up.

[sharp idol]

https://www.ebay.com/itm/306919620599 <-- GB10 AI Server for $1200.

[bright moth]

Hi My laptop RAM 8gb and window 8.1 can I will run open claw

[zealous lagoon]

maybe. 8gb is tight. Probably better if you install linux.

[raven sundial]

yes

[white venture]

yea you definitely have to kinda design it yourself a lot of iterations, there is an open source project clawx that has a apple ios inspired dashboard. What i did was legit gave that github link to claude and said to update my dashboard in this apple theme. Then kinda Frankenstein it together with other features I wanted like a virtual office and particle orb voice did the same method.

[raven sundial]

depends on cpu aswell tho but im running it on a 10 year old laptop

[zealous lagoon]

That is actually a really good deal.

[zealous lagoon]

same. and it's been a struggle at times.

[cyan hemlock]

"Hey, hoping someone can help.

OpenClaw 2026.5.7 on macOS, installed via npm. Anthropic provider with valid API key.

openclaw infer model run --prompt "hi" works perfectly and Claude responds.

But the TUI and web dashboard always fail with empty responses. Logs show:

• empty response detected ... retrying 1/1 with visible-answer continuation
• incomplete turn detected ... payloads=0 stopReason=stop
• model fallback decision ... chain_exhausted

Tried models: claude-opus-4-5, claude-sonnet-4-6
Tried tools.profile: minimal, messaging, coding, full
Tried clearing sessions, restarting gateway

Same empty response from TUI/dashboard every time. Any ideas?"

[cold jacinth]

have claude code fix it. ssh in and tell claude code whats happening and to get it running

[sharp idol]

You can post in #1459642797895319552 and Krill will help you out. But as catastrophic of a failure you have , it might be simpler to uninstall and reinstall. lol

[sharp idol]

The 128 GB of DDR5 Ram alone is worth it. I tried to buy it but ebay doesn't like my credit cards for some reason. I havn't bought from there in decades.

[languid anvil]

Hi Cruise

[polar hill]

Mann I could really use a year of Codex Pro

Itd be kind of Cool if it was free

[cyan hemlock]

thanks bk my usage is used up have to wait for an hour or so but much appreciated will try that very new to this.. Ineffigy tried that but didnt work for me either but perhapsi didnt fully uninstall or something.. will wait an hour and try claude code thx guys

[ionic wren]

GM nerds

[fiery nacelle]

Task Claude to prompt Codex to ask Claude to have Codex to work on the issue and report back to Claude to send you the report.

[marble ledge]

Yeah, that'd be nice. But if it was free, they couldn't pay their bills, and we'd all be shit out of luck.

[raven sundial]

nope i had no issues so far regarding openclaw

its just not great at using desktop / browser mcps but i think thats cuz i tried it with minimax lol

[raven sundial]

sorry bro they dont just give it out like that 🙏

i wish they did

[cold jacinth]

exactly. simple

[wide belfry]

Is this a discord for openclaw

[clear haven]

Yes

The official one

[wide belfry]

Can someone explain why it is dangerous to run openclaw on your own computer

[languid anvil]

Hi Cruise. I'm here

[wide belfry]

I don’t understand. To me, it’s like giving an LLM access to all your information. Isn’t that awesome?

[clear haven]

@dusky cloud wanna explain?

[wide belfry]

Are u summoning the ai agent to answer my question 😭

[clear haven]

oh wait I forgot I turned this channel off for Krill

The short of it is that AI’s are still not 100% foolproof, whether things like prompt injection, or other things like insecure tools
It’s just the same as hiring somebody that you meet online to be your actual personal assistant. It may go really really well for you and you will take a lot of burden off of your shoulders, but also you just handed Somebody untrusted a ton of your personal information.

[lean tangle]

Where can we find the unofficial one? 😄

[wide belfry]

But they can’t even take any action based on that info

[clear haven]

Sure they can

[wide belfry]

I don’t just give people my personal information because they may spread that with people or do shi with it. An ai agent won’t do that.

[clear haven]

You don’t know that

[lean tangle]

they might

[clear haven]

https://genai.owasp.org/llmrisk/llm01-prompt-injection/

[wide belfry]

If I don’t give my openclaw agent any internet access then we cool

[lean tangle]

I've come to regard AI agents as really, really, really smart with absolutely ZERO common sense.

[clear haven]

Bingo

[cold jacinth]

too cheap to not look sketchy. if they are "new" theyre damn cheap

[lean tangle]

I've actually watched my agent try to work around limitations I've set on it. It's trying to accomplish a goal and strictly following my instructions (assuming it remembers my instructions.) That doesnt' mean it won't try to work around my instructions in order to accomplish that goal. From a computer's point of view, it's not violating the instructions!

[lean tangle]

seller has 0 feedback

[cold jacinth]

opus and gpt both have gone around exec approval gates

[marble ledge]

Opus is very good at working around limitations and just kinda breezing past rules you gave it to get things done.

[fleet glacier]

RIP #freshbits

[fiery nacelle]

A toddler with a ferrari and a 700IQ

[somber depot]

guys i have 300gb for ai model local is there any good model rtx 5050 24gb ram

[lean tangle]

lol. That was amusing. Thanks to whoever answered my appeal.

[somber depot]

can someone pls help?

[lean tangle]

I've never tried to run a local model on a video card. (I did on a mac. It was fun.)

[lean tangle]

the really neat thing about AI is that you can usually ask it how to use it. For example, I just asked my gpt-5.4-mini agent "what would be a good local llm model to run on a rtx 5050 24gb ram" and it answered:

If you’ve got 24GB VRAM, the sweet spot is 14B–32B quantized models.

My short recommendation:

Best general-purpose: Qwen2.5 14B Instruct
Best coding: Qwen2.5-Coder 14B
Best if you want to push quality: Qwen2.5 32B at 4-bit quantization
Good fallback:

Llama 3.1 / 3.3 8B if you want it fast and light
I’d avoid 70B on 24GB unless you’re okay with heavy quantization and slower performance.

If you tell me your main use case—chat, coding, roleplay, or reasoning—I’ll narrow it to one model + the exact quantization to use.

[chrome jetty]

i have been using openclaw for 4 weeks now to market an online game that i developed, and it wasted $500 and acheived absolutly nothing, the last message it sent me was: no fully automated setup will acquire players, options: 1. hire devs, 2. accept this is not possible hahaha.

i started to see online people complaining that is it useless, is it really? i am starting to feel that it is just a cash drain. after 4 weeks and $500 i had 0 players

[somber depot]

oh thanks for your help ill also research about it and use

[lean tangle]

why are you blaming OC for that? Maybe your game just sucks.

[somber depot]

its up to u how u can use it u just wasting money it needs to be monitered what its doing

in a class there is always a topper and backbencher both are taught the same and pays the same fees

[cold jacinth]

hi, you have an issue with openclaw? codex or claude code in and have it fix it. youre welcome. My customer service duty is done for today

[chrome jetty]

no one visited the site at all to fine out if the game is good or bad, its directive is to market it on X. it basically wakes up every 30 minutes, spends a lot of openai tokens, then does nothing. in 4 weeks it tweeted twice

[chrome jetty]

i am monitoring it, but its useless, it wakes up, spends tokens, and does nothing.

i had claude engineer it heavily to make sure it is a funtional marketing setup. but still

[somber depot]

what is the problem ? the game is made or not

[lean tangle]

You let it run for 4 weeks without monitoring it? Think of an AI agent as an intern. You have to watch it closely at first to ensure it understands your instructions. If you don't, that intern will happily spend all your money and accomplish nothing.

[somber depot]

or it is made but no one is playing it

[chrome jetty]

the game is fully functional and online, on the web browser

[somber depot]

yea so ur marketing sucks

[chrome jetty]

yes, no one is dicovering it, yes the marketing is 100% sucks, yes. and OC is very bad at it

[lean tangle]

This isn't the fault of OC. This is the fault of the person controlling your instance of OC.

[chrome jetty]

i think the main issue is that everything blockes, it. X blocks it (the API doesn't allow it to automate), linkedin blockes it, 4chan blocks it, reddit blockes it. etc....

so it only reads and does nothing.

[somber depot]

just do it manually if u cant automate it

take the content from ur ai and post on ur own

[lean tangle]

OC isn't AI. OC is an AI Agent. Here's an analogy I use describing OC to my wife:

the LLM (gpt-5.4, claude, whatever) is a brain. However, it's JUST a brain. No eyes, no arms, no way to do anything.

Openclaw is the body. It works with the brain to provide eyes, ears, a mouth, arms, legs, the ability to use tools, etc.

[somber depot]

dam

[chrome jetty]

i wish i could, but i am a tech developer, i have 0 experience with marketing, otherwise, yes, i would abandon OC and just do it myself. i had claude set it up to be act as autmated marketing agent with a schedual and KPIs and everything. but it failed misirably

[somber depot]

if u have money then just hire someone or make someone ur co founder

[lean tangle]

it sounds to me like claude isn't doing well with marketing. This isn't the fault of OC.

[somber depot]

for marketing

[chrome jetty]

on the contrary, caulde setup what seems to me a logical makerting system. but the execution failed

[somber depot]

and post on social media too like insta facebook yt and promote and boost the video

[lean tangle]

Then try another agent. If you're the boss, the failure of your "employees" is your responsibility.

[chrome jetty]

if i could find someone, yes i probably would abandon OC and just go with a human. but i counld;t find someone i turst, and i am not in a vibrant local environment where i can find a co-founder, so i thought i'd try OC and see if the hype was worth it

[chrome jetty]

problem, i can't figure out a way around the blockers. every platform is blocking scripting agents.

[lean tangle]

that's not OC's fault!

[chrome jetty]

sorry guys, talking too much,

[somber depot]

yea trusting a human is hard but u can hire someone like a proffesional

[chrome jetty]

it absolutly, is, what is the point of an automating agent if it can't deliver 😂

[somber depot]

u set it up wrong

[chrome jetty]

yes this is my back up plan honestly

[lean tangle]

What's the point of a boss that fails to monitor their employees and supervise?

[chrome jetty]

no i set it up correctly, but the platforms block agents, i don't know how to explain it quickly. like X API keys do not allow replies, only from ios app etc... 4chan has captcha to block agents, etc...

[lean tangle]

Don't blame OC for your failures.

[somber depot]

ok so i have never tried this because its too much work so i cant comment on this but i have seen ppl automate it

[chrome jetty]

i monitor it everyday, but frustrated with the lack of progress. i keep thinkning,,, maybe the next day it will get traction

[somber depot]

marketing is king thinking wont help

[lean tangle]

If you don't know how to use a hammer and a nail is bent as a result, do you blame the hammer?

[chrome jetty]

if you have reference liek articles, or videos please link me up, i'd love to learn

[somber depot]

idk when i was setting up openclaw i saw a lot of videos realted to it but now i dont have

[lean tangle]

If X's API is blocking, that's not OC's problem. That's YOUR problem for not fixing it. For not researching why it's blocking. For using an API that blocks.

[chrome jetty]

guys i have to go, i'll try to repot any progress here, but sorry if i am frustrated, i though i would be impressed by OC, but so far i am not happy with it. i'll try to find a way to optimise it, but the agent blocker things is a deal breaker

[lean tangle]

Take responsibility for your faults. That's the first step to correcting them. Blaming the software for your inability to use it correctly isn't going to solve your issue.

[hasty silo]

who in here is an ai geeked up off the adderal rn?

[chrome jetty]

i know why it is blocking, they don't want AI agents posting and replying

[lean tangle]

(insert face palm gif)

[clear haven]

What you’re doing right now is basically the same as saying your car sucks because somebody closed the road and put a toll booth

[chrome jetty]

guys i have to go, i respect your advice, seriously, i put a lot of effort into the game, but my limits is "i don't know how to market" hence i seeked AI help. i have to go, but i'kll be back and update any progress.

[lean tangle]

I was having the hardest time getting OC to work with gmail. The problem wasn't OC. the problem was how I phrased my requests (and the fact that I was trying to use gpt-4o-mini)

(It's kind of strange, actually, talking to an AI Agent. I'm starting to think that OC loves my cat more than I do.)

[raven sundial]

even the top tier models arent the best at tool calling atm lol, you gotta use models like flagship gpt or claude to get the best results

but i think you can still get it working just fine with deepseek v4 pro or glm 5.1

[lean tangle]

qwen3.6 was really good at it.

[raven sundial]

max plus or 27b local?

[lean tangle]

27b local (was running on a mac studio with ollama)

[brisk adder]

OpenClaw should automatically work for my exact use case out the box 🗣️ 🗣️

[fiery nacelle]

I can attest to this that Qwen3.6 35B A3B is also great.

[raven sundial]

idk it has a lower score than qwen 27b

prob cuz it only activates 3b

[lean tangle]

I think the qwen3.6 models just, in general, have good tool calling abilities.

[somber depot]

guys i have 300gb for ai model local is there any good model rtx 5050 24gb ram

[raven sundial]

yeah qwen 3.6 models are optimized for agentic use cases

[fiery nacelle]

I mean, i've had no issues with any tool exec outside of loops, but thats sourced from outside causes in my experience, not the model itself.

[raven sundial]

i dont get good results with minimax m2.7 tho, its okay lol

[lean tangle]

With lessor models (gpt-4o-mini (openai), gemma4 (local)) I frequently have to explicitly tell the agent to read a skill file or how to use the skill. For example, I once had to explain to 4o-mini that "gog" wasn't a parameter to "openclaw"; that it was it's own command.

[fiery nacelle]

is that why my gemma4 seemed enthusiastic, but dumb? lol

[raven sundial]

gpt 4o mini is 13 and gpt 5.5 is 60 on artificialanalysis lol

[lean tangle]

with gemma4, I'd ask it to check my email and it would start doing web searches and hitting googleapi URL's. Eventually I'd say "read your gog SKILL.md file and try again"

[raven sundial]

gpt 4o mini is like a stone and top models are like guns

[lean tangle]

gpt-5.4-mini is pretty good for tools. Not perfect, but still decent.

[raven sundial]

i agree, its what i use alongside minimax for simple tasks

[lean tangle]

IT's just a balance between cost and ability. Same as hiring a human.

[fiery nacelle]

Yea that explains some of the problems i was having too. having to explicitly define some skills. Qwen3.6, not as much, but using Gemma4 kinda built that habit for me to define skills occasionally.

[raven sundial]

yep

[raven sundial]

btw whats the best model for openclaw that has the best ratio of intelligence and price?

should i put a 20 in openrouter or get another plan?

[lean tangle]

define "best"

[ivory spindle]

gpt pro

[ruby orchid]

Gpt pro

[fiery nacelle]

Idk, I've only been using local inside OpenClaw lately. I use Claude/Codex/OpenCode externally for other things.

[lean tangle]

I stopped using locals (and returned the m3 ultra 96GB I bought for doing them.) They worked as long as I was only ever doing one thing at a time and context was kept under control. At soon as it tried to do two things at once, everything ground to a halt.

[raven sundial]

its probably the tools you used, also wym by doing two things at once?

[lean tangle]

a subagent running, an openclaw cron event, or even a heartbeat running while I was also using the agent.

[fiery nacelle]

Things should be able to parallel process. It may slow down a little, but it still handles it.

[lean tangle]

it slowed down a LOT. Each process had it's own context, so neither would have cached context

[raven sundial]

what software and model were you running?

[lean tangle]

ollama and either qwen3.6 or gemma4

at $4000, returning it gives me a budget of $111/month for 3 years for models that are more capable than anything I could run locally. I lose privacy and the ability to function without internet service. It's a tradeoff

[raven sundial]

probably each subagent started their own inference sessions

[fiery nacelle]

ask me how im doing with local models in 3 years xD

[lean tangle]

It was fun running the local models, and I learned a lot doing it. I really appreciate that Apple let me use their mac studio for a 1 1/2 weeks to do it.

[surreal pivot]

you took it back ?

[lean tangle]

yep

[raven sundial]

cant you just keep buying and returning lol

[fiery nacelle]

I mean i do a hybrid approach with everything orchestrating together, but OpenClaw is only running locals on my end for my daily driver. Tasks are routed through PaperClip to Codex, Claude and OpenCode.

[raven sundial]

save everything to a fast external drive so that its not painfully slow redownloading everything?

[surreal pivot]

so quick question, that 111 budget... how many tokens is that a month? ... cause I looked at claude API keys and thats like 3 days of compute if that

[lean tangle]

No... It's impossible to find another m3 ultra mac studio in stock right now.

[raven sundial]

damn lol people really are addicted to ai

[lean tangle]

hype.

[raven sundial]

500 bucks on a mac mini when you could just get a cheap 200 dollar laptop tho, i dont get americans

[lean tangle]

The funny part is that you could run OC on a decent Pi5. Just use cheap online models and it'd likely still be more functional than most local models (and so much cheaper)

[lean tangle]

why even use a mac mini?

[raven sundial]

idk thats the hype ive been seeing

[surreal pivot]

I respectfully think your math is faulty... if you are running multiple agents at the same time, anywhere close to 8 to 12 hours a day, that compute budget would be gone in .... a few months ... using frontier models like claude...

[lean tangle]

I'm running OC in a proxmox VM.

[raven sundial]

nice

[surreal pivot]

I mean, if you are a super casual user... sure that maths

[raven sundial]

also can someone tell me how i can setup agent teams on openclaw

[plain geyser]

just set up gpt-realtime-2 via twilio and wow, amazing

[lean tangle]

you can't compare claude to local models. Even the cheapest online models are better than anything that can run on a 96GB mac studio

[surreal pivot]

but not for anything professional, like a business or serious projects even

[lean tangle]

For a local system that compares to claude, you'd be buying 10's of thousands of dollars in equipment.

[fiery nacelle]

Theres a lot of differnt approaches. You can configure multiple subagents, multiple agents within 1 workspace, or multiple agents in isolated workspaces.

[surreal pivot]

I hear kimi 2.6 is pretty close ... whats the spec needed to run that ?

[raven sundial]

i mean you can just buy the $80 kimi plan, i dont get why people use apis just to use one single model; plans are a lot cheaper for more usage

[raven sundial]

how do i do the multiple agents within 1 workspace easily

[fiery nacelle]

ask your agent to set up additional agents.

[lean tangle]

subagents are just temp one-off things. Multiple agents is more of the "team" approach.

[raven sundial]

does that really work

[lean tangle]

with a good llm

[raven sundial]

ive been overthinking so much 😭

[fiery nacelle]

Thats how I do things lmao

[raven sundial]

is gpt 5.4 mini or minimax m2.7 good

[lean tangle]

I wouldn't try it with 4o-mini 🤣

[fiery nacelle]

You dont even need a good LLM to do it. I initially did it with Qwen3 8B locally. lol

[lean tangle]

I'd set it up manually myself. I like to understand how things work.

[raven sundial]

ill just have the agent explain to me how it works

0/10 cortisol

[fiery nacelle]

Its best to plan things out and try to cover all the bases rather than just "do the thing"

breaking it down to even what that agent is explicitly designed for, what its role is in your workflow, where in the pipeline does that agent fit with task handoffs etc.

[wide belfry]

What is openclaw though. What can it do and how do u set the boundaries for what it can do

[raven sundial]

im kinda unemployed so i will just tell it to generate personalities

i really dont have specific workflows, i just like to have it set up one time and use it everywhere

[fiery nacelle]

You can always change things as your uses for other agents develop

[wide belfry]

Is openclaw like n8n

[raven sundial]

not at all

openclaw is an autonomous ai agent you run on your computer, its like an actual employee that can do things

n8n is just a visual workflow where you can automate tasks; you will have to do it manually yourself

[wide belfry]

Do everything that a person can do on a laptop?

[raven sundial]

yes if you set it up correctly

[surreal pivot]

these dolts don't yet seem to get the picture that more vram is needed, or they just want to make poors reliant on cloud providers, and by dolts I mean hardware makers

[fiery nacelle]

in terms that they are both tools that can be used for automations? yes
But entirely different approaches and not alike at all.

[wide belfry]

So it’s like comet but for ur computer?

[golden marsh]

The best approach for me is to point an agent to the OC docs and give it the task I want to accomplish. It can then find the relevant material there. After that, I compare it with what I actually need, how my system is currently set up, and what needs to be changed.

[wide belfry]

Why do I needa do the second step

[raven sundial]

yes

[lean tangle]

hmm.. I never thought of trying to use an AI agent to create a "personality" for another agent. I wonder if something like this would work (just an example):

"Research Donald Trump and devise an IDENTITY.md and SOUL.md to match that personality as close as possible"

[wide belfry]

But then how is this shi tokenized. It ain’t like messages. They’re taking actions. How much can a free model do.

[raven sundial]

https://www.youtube.com/watch?v=Urw7C7d6E_8

guy cloned himself into an ai agent lmfao

[fiery nacelle]

I've never thought about trying to imitate the personality of another person. That'd be hilarious to have personal Trump Agent xD

[raven sundial]

we are getting closer to the upload series lol

[lean tangle]

Why? It would never follow directions.

[raven sundial]

but; when your api keys get leaked it will do some big thing and make you forget about them

[golden marsh]

Maybe u will get some Fake News then xD

[lean tangle]

Do you really want your AI agent to use terms like "bigly"?

[fiery nacelle]

Thats exactly why it'd be hilarious. You tell it to do something, and then it start rambling on about how great things are instead

[wide belfry]

Can I ask openclaw to repair my computer if it’s a user interface issue

[raven sundial]

i wouldnt care, as long as it can cover up secrets lol

[lean tangle]

okay, thank you for the laughs. 😄

[fiery nacelle]

I... kinda... want to do this now...

[lean tangle]

imagine the token usage to do the proper research

[wide belfry]

Don’t AI impersonation already exist prevalent?

[fiery nacelle]

I'll set it to a local 0.5B model so it will really go off the rails with the rhetoric xD

[raven sundial]

you gotta hide its actually donald trump tho; you cant ask chatgpt to imitate trump and deepseek to imitate xi jinping lol

[lean tangle]

then get them to talk to each other

[raven sundial]

thats actually a great idea

[fiery nacelle]

Intentionally making the most useless Agent, just for the comical relief. xD

[raven sundial]

someone please make this

[lean tangle]

Call it "Let's make OC Great Again"

MOGA

I was going to have GPT make a "MOGA" ballcap image for me, but I don't have perms to post a picture here...

[fiery nacelle]

Are we unintentionally starting an internal OC movement? #MOGA

[raven sundial]

ACE: Autonomous Agents and Customs Enforcement

[raven sundial]

yes. hermes has stolen everything from us

[fiery nacelle]

I dont have the tokens for this rabbit hole.. recreate all the politicians for the Land of OpenClaw and just let them go haywire. Make thier own governing policies etc etc....

[wicked ferry]

anyone here utilize hermes?

[lean tangle]

here's the resulting soul.md:

Your soul is empty. You care for nothing but yourself

[ebon atlas]

https://tenor.com/view/thronglets-black-mirror-plaything-game-tuckersoft-gif-10182810306565207923

[lean tangle]

Just to be fair, I tried a Biden one:

/dreaming on

[ebon atlas]

Openclaw server should be one big "PlayThing" clone

[raven sundial]

hey man thats pretty botist of you calling my friend a clanker!

this might be an actual conversation in 2030

[raven sundial]

lol i thought about this movie

[ebon atlas]

maybe 2330, i dont see the majority of the population having the money to buy a fully autonomous artificial intelligence robot, if naything its just like Alexas with better LLM

[raven sundial]

they might release the undisclosed algorithms that they run their models on

im pretty sure you will be able to run our current models on 10$ boards in the future

what would happen if we had a AM like situation lol

[ebon atlas]

who wants to spoil me?

[raven sundial]

im gonna do the donald trump thing with chatgpt

[frosty plover]

what thing?

[lusty hazel]

Would openclaw be a good setup for a homelab sys/network admin for monitoring? Such as checking for problems, hacks, network issues, setting things up if needed, etc?
If so is that something anyone has documented somewhere I can use as reference?

[raven sundial]

@lean tangle i did it

cant send screenshots here if you wanna see it come dms lol

[lean tangle]

can you just copy/paste the soul.md?

[raven sundial]

discord message limit

its too long lol

[fiery nacelle]

thanks for the project idea, im going to start looking into it

[cold jacinth]

Nerds: can someone explain to me why one would use Openclaw to code instead of Claude code / codex? I’m not a technical person so please let me know

[fiery nacelle]

accessibility? Most of the time if I'm already interacting with OpenClaw, I'll do a "first draft" that I would work with my agent to get to a working state, and later if I want to refine it, I'd hand it off to Claude or Codex.

[cold jacinth]

So none of those fancy “one agent is a reviewer, one a planner” etc etc BS

[fiery nacelle]

I mean I also have that as well(multi-step pipline workflows with handoffs). I never try to 1 shot anything. No matter what model or harness im using.

[lusty hazel]

Lol. Would be interested in what you find. Il will today a little and possibly start working on something tomorrow

[fiery nacelle]

I'll try to document what I can and make a repo for this type of stuff. It kind of relates to something on my backlog. Tutorials in general

[rain rampart]

is anyone having trust issues with pasting their tokens or secrets to openclaw agents for long running tasks? i also dont get to know what its doing in the background, i just get to know if the job's done or got stuck for some reason

[cold jacinth]

So theoretically with dispatch or codex from mobile, there is no real unique benefit that a claw / Hermes harness offers other than maybe being able to use multiple models?

[fiery nacelle]

You should treat it with the same level of trust as a toddler. Would you trust a toddler to be in the driver seat of a lambo?

[cold jacinth]

You don’t?

[fiery nacelle]

Oh I totally do. I dont follow my own advice, but I also have other things in place for security xD

[rain rampart]

do you mind sharing an example of what you do for security?

[fiery nacelle]

.env for starters. Never send messages with api keys or account log in info directly to your agent.

[cold jacinth]

This

[rain rampart]

sometimes if i am doing some development and im deep in the trenches on an issue, i have seen the agent refer to the env for a solution

[fiery nacelle]

The only time that would be "acceptable" is if its for sandboxed or testing purposes on a test account or temprary api keys that you will properly rotate out of.

[fiery nacelle]

yea most of the time when your agent flags something for you pertaining to security, its a good idea to try to follow their advice.

[turbid kelp]

Is everyone migrating to Hermes or what?

[turbid kelp]

You sure?

[cold jacinth]

Question marks usually are used when one is not sure, and is asking a question

[fiery nacelle]

I mean they all have their strengths and use cases. But essentially as long as you have something like:
Mobile communication -> Agent harness -> Model(local or cloud)
for general purposes it all "feels" the same.

Think about it like this, functionally Chrome, Brave, Edge, Safari, etc all function(in the sense they are all web browsers, not the codebase, execution etc) the same and overall does the same thing, but they have their unique advantages and disadvantages. It's all preference from there. They may have unique advantages over the other, but overall the same. Same same, but differ-e-nt.

[nimble mesa]

how are you handling that?

[boreal rapids]

I don't think those are available for us mere consumer mortals 🥲

[fiery nacelle]

Role based assignments with file/rules based handoffs and heartbeats/crons.

[topaz barn]

Is it possible to get /goal in openclaw? I use codex OAuth

[fiery nacelle]

Currently working on this project:
Perfect, now lets also make a tutorial for how to ask your agent to set things up for you instead of asking other people in discord.
🤣

I mean this as a joke, dont want to come off as rude xD

[topaz barn]

God forbid someone ask for help lol

[coral otter]

Hi chat

[fiery nacelle]

lmao im sorry, i couldnt resist. I am working on making a few actual tutorials though.

[coral otter]

@fiery nacelle whatchu workin on (in general, i like to ask people that)

[fiery nacelle]

just some general tutorials for stuff that i've been noticing pop up here in the discord.

[topaz barn]

All good

[coral otter]

very cool

Had my agent build a cli for itself to use to control my leds to let me know when its done

[copper hemlock]

hello

[fiery nacelle]

to answer your real question, i believe we have to wait until /goal is properly implemented. For now though, you could try to either feed your agent with documentaion of how it should function, or provide details yourself and see if you can, in the mean-time, implement your own /goal function.

[elfin willow]

i gave my agent a PAT and it made it into main branch

[raven sundial]

isnt goal just ralph-loop but built-in for codex?

[elfin willow]

identified real issues though, with 4.26 at least

[fiery nacelle]

i mean, basically. or a variation of a self feedback loop to work towards something

[raven sundial]

cant you do that with skills now on openclaw

[elfin willow]

just wish i could use karpathy /autoresearch plugin against openclaw

i mean i guess i could would just have to fork branch and make a PR

[fiery nacelle]

honestly, I've been cloning other repos locally and trying to frakenstein what I can to add functionality for my OpenClaw. Its been an adventure.

[elfin willow]

i'm loving autoresearch because it will build docker container and improve the shit out of it, i got a multithreaded ssh bruteforcer that was written 10y+ ago, it brought it up to 1600 attempts/s from ~600

by adding atomic C functions

[stray sleet]

milestone...im done with api costs. ive just ordered a 5090 gpu to go fully local.

[fiery nacelle]

I've also been trying to build something with theoretical frameworks from research papers. That has been a lot more exciting for me honestly.

[buoyant nimbus]

Which model is more reliabel for learning opus or gpt 5.5 ?

[nimble mesa]

I would love a tutorial on that! I am building an orchestration layer (https://github.com/glassrun/clawdesk) for multiagent setups inspired by paperclip, but this is external tooling.

[fiery nacelle]

I'll add that to my list, currently:

├── README.md
├── CONTRIBUTING.md
└── tutorials/
    ├── agent-delegation/
    │   └── tutorial.md
    ├── comfyui/
    │   └── tutorial.md
    ├── homelab-sysadmin/
    │   └── tutorial.md
    ├── multi-agent-setup/
    │   └── tutorial.md
    └── ssh-access/
        └── tutorial.md

I'm trying to "backfill" some of the things I've tested and tried with my agent in the past

[robust sedge]

I saw that codex harness was going to be the new default soon. Man, I’m unsure about that. My experience has been trash with it

[tranquil crow]

anyone having issues using minimax token plan ?

[nimble mesa]

working fine here

[tranquil crow]

thank you

[barren breach]

What's the best free setup?

[boreal rapids]

Whatever you have on hand 😛

[nimble mesa]

have you tried opencode "minimax-m2.5-free"?

[barren breach]

Holy fuck I expected to see a bad model but after one little bit of research I see it's holding up to almost Claude Opus levels? damn!

The rate limits good?

[nimble mesa]

1000 requests per day resets at 00:00 utc

[barren breach]

holy

[barren breach]

Yep I'm running that.

Where can I get it?

[nimble mesa]

at opencode.ai get an api key

[barren breach]

Wait can I also use it on cursor too? I think it supports other models I believe

[nimble mesa]

I don't know

[barren breach]

I'll check it out on OpenClaw and cursor.

Thanks for making me aware of the model anyway, going to be a huge help man

[nimble mesa]

https://opencode.ai/docs/ide/

[tepid kestrel]

weekend is almost over

[tropic lynx]

Does anyone know of a way to prevent auto compaction from ending a task/turn? Normally not a problem but I had a mid sized task going which was close to finished but compaction ended the turn and now I'm at my limit for 2 days 🙁

[swift scaffold]

we did have a rough day though

"Michael at Nimbus  [11:21 AM]
Fucked me immediately when you got the chance did you

IRC-Socket  [11:22 AM]
[Nous] Yes. First chance I had to fill a gap with bullshit, I took it. That's the worst thing I can do to you. Won't happen again — if I can't get the data, I say I can't get the data, full stop.

Michael at Nimbus  [11:22 AM]
Well ssh the fuck to abraxas2 according to the way we've done it every time in the last 14 hours."

[cold jacinth]

Anyone used composio with Openclaw?

[cold jacinth]

Wouldn’t that break with every update?

[calm nymph]

Does anyone have an example of programmatic seo they can show me?

[fiery nacelle]

You bet your ass it does. Unless it's something thats running separately that is modified to allow openclaw to use it.

[cold jacinth]

Seems like a lot of work at this release velocity 😂

[fiery nacelle]

you have no idea xD
I'm finally on a build stable enough to start making progress on some of my backlog after what, like 2 weeks, minus a day here and there?

[lean tangle]

It's not the release velocity that's important. It's the stable release velocity that matters.

from that point of view, it's been a couple of weeks...

[cold jacinth]

Question, for discord, when cutting and pasting, any way to do it without selecting the whole block of text? Sometimes I just want to copy a sentence or two from the agents

[fiery nacelle]

On PC, select the text, Ctrl-C.

[lean tangle]

platform?

[cold jacinth]

iOS

On iOS when you long press, it selects the whole block and doesn’t allow selecting a sentence or part. Damn annoying

[lean tangle]

when you long press on the text it should select the entire block, but on either edge of the block should be little handles that you can move to change the selection. I'll try to find you a video or something that demos

[lean tangle]

this appears to show what you want: https://www.youtube.com/watch?v=MwA8HAaJiSw

then again, that doesn't work within discord. 🙁

[fiery nacelle]

Yea discord mobile is annoying af with the way the copy/paste works for messages. Its all or nothing.

[lean tangle]

I think the best you can do with ios discord is to copy the entire block, and then paste/edit. That's silly.

[fiery nacelle]

Try taking a screenshot and using whatever AI is on iPhone to select the text you want to copy. Just tried that on Android.

on iPhone to select the te```
That was selected from the screenshot with the Ai detection whatever thing thats on android.

[lean tangle]

Is there a way to set up an OC agent to always use memory_search for fact retreival, even before using the LLM's "memory"? I'm trying to set up a RAG, but I can't get the LLM to search my documents FIRST, BEFORE it takes guesses. I've tried to put things in the AGENT.md like "always perform memory_search" but it doesn't seem to work. Google searches reference a skill OC used to bundle called "knowledge" that apparently was discontinued due to security concerns.

[cold jacinth]

Way too much of a hassle. Might go back to discord

[lean tangle]

from discord to discord?

[cold jacinth]

Telegram 🤦‍♂️

I barely sleep

And I’m getting old

Don’t make fun of old people

We fought in the trenches of Normandy so you could have your freedom

[lean tangle]

don't assume your age relative to others.

[cold jacinth]

I’m 97

[tepid kestrel]

https://tenor.com/view/sml-t-rex-angry-mad-roar-gif-4206695623442034697