users-helping-users

[prime fern]

and do it again

and i cant even F----- do that

jesus

[indigo goblet]

no need to uninstall

did the .json saved?

all the blocks udpated correctly?

[prime fern]

i cant find the error

[indigo goblet]

its just syntax

you missed a comma before a { maybe

that block should be like that;

"agents": {
"defaults": {
"model": {
"primary": "ollama/qwen3:4b"
},
"models": {
"ollama/qwen3:8b": {},
"ollama/qwen3:4b": {
"contextWindow": 8192,
"maxTokens": 8192
}

[prime fern]

ok. i found the error

[indigo goblet]

and in the models;

"agents": {
"defaults": {
"model": {
"primary": "ollama/qwen3:4b"
},
"models": {
"ollama/qwen3:8b": {},
"ollama/qwen3:4b": {
"contextWindow": 8192,
"maxTokens": 8192
}

[prime fern]

was missing a comma

[robust bronze]

has anyone tried the lfm2.5 thinking model from ollama with claw?

[prime fern]

ok. i made that change already

but for a piece of software that is being billed as having a simple install, editing these configu files is super messy

[indigo goblet]

saved then, restart the gateway and start a new tui/session

[indigo goblet]

not sure what you did before with install/uninstall and if your workspace got uninstalled correctly etc, might come from that;

[prime fern]

95 },
96 "agents": {
97 "defaults": {
98 "model": {
99 "primary": "ollama/qwen3:4b"
100 },
101 "models": {
102 "ollama/qwen3:8b": {},
103 "ollama/qwen3:4b": {
104 },
105 "contextWindow": 8192,
106 "maxTokens": 8192
107 }
108 },
109 "workspace": "/home/salma/.openclaw/workspace"
110 }
111 },

there is an error here

[prime fern]

i restored previous snapshot. or but you could be right. maybe two many redundant "installs" even after it told it was installed

[indigo goblet]

that one shouldnt be like that:

"models": {
102 "ollama/qwen3:8b": {},
103 "ollama/qwen3:4b": {
104 },
105 "contextWindow": 8192,
106 "maxTokens": 8192

[prime fern]

ah ok

[indigo goblet]

but like that;

"models": {
"ollama/qwen3:8b": {},
"ollama/qwen3:4b": {
"contextWindow": 8192,
"maxTokens": 8192

[prime fern]

5 },
96 "agents": {
97 "defaults": {
98 "model": {
99 "primary": "ollama/qwen3:4b"
100 },
101 "models": {
102 "ollama/qwen3:8b": {},
103 "ollama/qwen3:4b": {
104 "contextWindow": 8192,
105 "maxTokens": 8192
106 }
107 },
108 "workspace": "/home/salma/.openclaw/workspace"
109 }
110 },

like that?

[indigo goblet]

yes

[prime fern]

do i need those cloud models in there?

anyways running it again

Config invalid
File: ~/.openclaw/openclaw.json
Problem:

• agents.defaults.models.ollama/qwen3:4b: Unrecognized keys: "contextWindow", "maxTokens"

still an error message

those options didnt exist for the agent block originally

are you sure they should be added to the agent section?

[indigo goblet]

hm, its a new issue apparently

hm ok then move;

"contextWindow": 8192,
"maxTokens": 8192

to the models block like that

[prime fern]

i deleted that section from the agent block. now its working

[indigo goblet]

"id": "qwen3:4b",
"name": "qwen3:4b",
"reasoning": false,
"input": ["text"],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 8192,
"maxTokens": 8192

[prime fern]

◆ Default model
│ ○ Enter model manually
│ ○ ollama/glm-5:cloud
│ ○ ollama/kimi-k2.5:cloud
│ ○ ollama/minimax-m2.5:cloud
│ ● ollama/qwen3:4b (ctx 256k)

◇ Ollama mode
│ Cloud + Local
│

i guess i could just choose local only

but lets run with this now

restarting gateway

hatching in tui

[indigo goblet]

for the latest section, close it;

"models": {
"ollama/qwen3:8b": {},
"ollama/qwen3:4b": {}

[prime fern]

Ollama API error 500: {"error":"model requires more system memory (38.9 GiB) than is available (33.6 GiB)"}

Wake up, my friend!

run error: Ollama API error 500: {"error":"model requires more system memory (38.9 GiB) than is available (33.2 GiB)"}
gateway connected | error
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

still the same error

which begs the question. does qwen3:4B need 3GB of ram or 40GB?

[indigo goblet]

try that first

[indigo goblet]

and this

[prime fern]

System Requirements

VRAM requirements for different quantization methods and context sizes

1,024 tokens
3.36 GB VRAM

Consumer

1x RTX 3060

12GB VRAM

Datacenter

1x RTX A6000

48GB VRAM

Apple Silicon

1x M2 Max

32GB VRAM

32,768 tokens
13.76 GB VRAM

Consumer

1x RTX 4060 Ti

16GB VRAM

[indigo goblet]

and you dont need to rerun the install everytime, just in a terminal > openclaw gateway restart then 'openclaw tui' in a fresh terminal

[prime fern]

it doesnt recognize any openclaw commands after it is "installled"

[indigo goblet]

openclaw tui, doesn't do anything in a terminal?

[prime fern]

no openclaw command works

[indigo goblet]

hm we might need to stepback then if this is what happens, not normal. Maybe uninstall openclaw completely and correctly first then we see and resume.

[prime fern]

eventhough it says openclaw was installed and connected. after i get those errors and end the session i cant run any openclaw commands

[indigo goblet]

but anyway for now, did you edit the files like I said, the last edits

and rerun onboard etc, make sure your .json is the same and try

[prime fern]

that is the problem. i cant even uninstall.

i literally have to restore the snapshot

so i will do that now

and do a fresh install

[indigo goblet]

hm hold on

back up your openclaw.json

then

openclaw gateway stop

ah shit

right

[prime fern]

ok

[indigo goblet]

try those

systemctl --user disable --now openclaw-gateway.service

rm -f ~/.config/systemd/user/openclaw-gateway.service

systemctl --user daemon-reload

[prime fern]

ok. done

[indigo goblet]

then

rm -rf "${OPENCLAW_STATE_DIR:-$HOME/.openclaw}"

rm -rf ~/.openclaw/workspace

npm rm -g openclaw

[prime fern]

done

[indigo goblet]

ok install it again

curl -fsSL https://openclaw.ai/install.sh | bash

[prime fern]

ok running

[cunning relic]

hii

can I have my openclaw chat in telegram interact with my current browser session?

i already set profile as user in json config, enabled remote debugging but the agent wouldn't acknowledge it

[prime fern]

How do you want to hatch your bot?
│ Hatch in TUI (recommended)
openclaw tui - ws://127.0.0.1:18789 - agent main - session main

session agent:main:main

Wake up, my friend!

run error: Ollama API error 500: {"error":"model requires more system memory
(38.9 GiB) than is available (33.0 GiB)"}
gateway connected | error

still the same error

so which llm should i use

[indigo goblet]

okay firs thing, open a terminal > openclaw gateway restart

it works?

[prime fern]

salma@ubuacetha:~$ openclaw gateway restart
openclaw: command not found
salma@ubuacetha:~$

[prime fern]

no. i am getting the same error about my ram

and my system still doesnt recognize any openclaw commands

wait. let me check the config file again

ok. i changed the context window to 8192

[indigo goblet]

hm ok hold on

[prime fern]

but seriously dude. if i cant even install openclaw. how could a newbie be expected to do this?

! PATH missing npm global bin dir: /home/salma/.npm-global/bin
This can make openclaw show as "command not found" in new terminals.
Fix (zsh: ~/.zshrc, bash: ~/.bashrc):
export PATH="/home/salma/.npm-global/bin:$PATH"

maybe this is part of the problem?

[indigo goblet]

yeah i was looking at new issues, this is it

run

npm prefix -g
npm bin -g
echo $PATH

[prime fern]

im going to swithc ollama to local only

[indigo goblet]

what does it says

[prime fern]

salma@ubuacetha:~$ npm prefix -g
/home/salma/.npm-global
salma@ubuacetha:~$ npm bin -g
Unknown command: "bin"

To see a list of supported npm commands, run:
npm help
salma@ubuacetha:~$

there is no bin option

npm help npm more involved overview

All commands:

access, adduser, audit, bugs, cache, ci, completion,
config, dedupe, deprecate, diff, dist-tag, docs, doctor,
edit, exec, explain, explore, find-dupes, fund, get, help,
help-search, hook, init, install, install-ci-test,
install-test, link, ll, login, logout, ls, org, outdated,
owner, pack, ping, pkg, prefix, profile, prune, publish,
query, rebuild, repo, restart, root, run-script, sbom,
search, set, shrinkwrap, star, stars, start, stop, team,
test, token, uninstall, unpublish, unstar, update, version,
view, whoami

[indigo goblet]

run this; ls -ld /home/salma/.npm-global/bin

[prime fern]

salma@ubuacetha:~$ ls -ld /home/salma/.npm-global/bin
drwxrwxr-x 2 salma salma 4096 Mar 26 09:49 /home/salma/.npm-global/bin
salma@ubuacetha:~$

[indigo goblet]

run; echo $PATH

[prime fern]

salma@ubuacetha:~$ echo $PATH
/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin:/snap/bin
salma@ubuacetha:~$

[indigo goblet]

run; export PATH="/home/salma/.npm-global/bin:$PATH"

then run openclaw --version

[prime fern]

salma@ubuacetha:~$ export PATH="/home/salma/.npm-global/bin:$PATH"
salma@ubuacetha:~$

salma@ubuacetha:~$ openclaw --version
OpenClaw 2026.3.24 (cff6dc9)
salma@ubuacetha:~$

[indigo goblet]

ahhhh

finally

[prime fern]

yes

well done

https://tenor.com/view/yess-girl-gif-26883543

https://tenor.com/view/rugby-yay-gif-3156747878755757137

not so fun when it doesnt display the gif. lol

[indigo goblet]

run this now;

echo 'export PATH="/home/salma/.npm-global/bin:$PATH"' >> ~/.bashrc

[prime fern]

salma@ubuacetha:~$ openclaw security audit --deep
openclaw: command not found
salma@ubuacetha:~$

still cant run openclaw commands tho?

[indigo goblet]

its not over, we have to make it permanent

[prime fern]

salma@ubuacetha:~$ echo 'export PATH="/home/salma/.npm-global/bin:$PATH"' >> ~/.bashrc
salma@ubuacetha:~$

[indigo goblet]

source ~/.bashrc

now

[prime fern]

ok. but why all these extra steps. what about the one liner install?

salma@ubuacetha:~$ source ~/.bashrc
salma@ubuacetha:~$

done

[indigo goblet]

ok now; hash -r

and retry openclaw --version

[prime fern]

salma@ubuacetha:~$ openclaw --version
OpenClaw 2026.3.24 (cff6dc9)
salma@ubuacetha:~$

still the same as last time

[indigo goblet]

ok try it in a new terminal, should work

[prime fern]

run bash again

[indigo goblet]

no

openclaw --version

this fix was to make openclaw commands recognized by shell so you dont have to reinstall everytime

if it works, then we can resume on the other issue

[prime fern]

damn sonny

well done

its working now

amazing

[indigo goblet]

ok wooufff nice

[prime fern]

nope

i spoke to soon

it automatically opened the web gui

but now in the browser gui

i get the same error

Ollama API error 500: {"error":"model requires more system memory (38.9 GiB) than is available (32.0 GiB)"}

Wake up, my friend!
openclaw-tui (gateway-client)

[indigo goblet]

open a new terminal > openclaw tui

[prime fern]

i think we should try a different llm

[indigo goblet]

working?

[prime fern]

tui same error

Ollama API error 500: {"error":"model requires more system memory (38.9 GiB)
than is available (31.5 GiB)"}
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

[indigo goblet]

it's okay, how is your openclaw.json

[prime fern]

i fixed the new json for local only

changed context window to 8192

still same error

[indigo goblet]

did you edit everything that needed to be edited, what I wrote earlier

the last edit, i'm talking about

[prime fern]

one sec

ill copy the new json

[indigo goblet]

hold on no

is the fresh .json looking like the old one?

or is it very different?

[prime fern]

fresh json

because i chose the local option only for ollama

[indigo goblet]

your openclaw.json

[prime fern]

new json

no cloud models

⚠️ Agent failed before reply: Model context window too small (8192 tokens). Minimum is 16000.
Logs: openclaw logs --follow

new error

[indigo goblet]

okay

change it to 16000

might be not enough but let's see

I think normally this should be changed on ollama side too

[prime fern]

⚠️ Agent failed before reply: Model context window too small (8192 tokens).
Minimum is 16000.
Logs: openclaw logs --follow
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

i changed the context window to 16000

so i dont know what i am still getting that error

[indigo goblet]

in openclaw.json correct?

terminal where ollama is running, throw that; OLLAMA_CONTEXT_LENGTH=16000

[prime fern]

yes

[indigo goblet]

then restart it

the .json is done already, so you have to restart openclaw too

[prime fern]

im closing all terminal

[indigo goblet]

unless youve done it already

[prime fern]

ok

how do i restart

restart the gateway?

[indigo goblet]

openclaw gateway restart in a terminal

[prime fern]

yup

doing that now

Ollama API error 500: {"error":"model requires more system memory (38.9 GiB)
than is available (31.5 GiB)"}

⚠️ Agent failed before reply: Model context window too small (8192 tokens).
Minimum is 16000.
Logs: openclaw logs --follow
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

still same error

[indigo goblet]

ok dont do anything, lets do it step by step

first, make sure your openclaw.json have been corrected with what I wrote earlier

[prime fern]

json still say 16000 for context window

one sec

ill copy the new json

{
"meta": {
"lastTouchedVersion": "2026.3.24",
"lastTouchedAt": "2026-03-26T09:50:50.352Z"
},
"wizard": {
"lastRunAt": "2026-03-26T09:50:50.305Z",
"lastRunVersion": "2026.3.24",
"lastRunCommand": "onboard",
"lastRunMode": "local"
},
"auth": {
"profiles": {
"ollama:default": {
"provider": "ollama",
"mode": "api_key"
}
}
},
"models": {
"mode": "merge",
"providers": {
"ollama": {
"baseUrl": "http://127.0.0.1:11434",
"apiKey": "OLLAMA_API_KEY",
"api": "ollama",
"models": [
{
"id": "glm-4.7-flash",
"name": "glm-4.7-flash",
"reasoning": false,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,

"cacheWrite": 0
},
"contextWindow": 16000,
"maxTokens": 8192
},
{
"id": "qwen3:4b",
"name": "qwen3:4b",
"reasoning": false,
"input": [
"text"
],
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
},
"contextWindow": 16000,
"maxTokens": 8192
}
]
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "ollama/qwen3:4b"
},
"models": {
"ollama/qwen3:4b": {}
},
"workspace": "/home/salma/.openclaw/workspace"
}
},
"tools": {

the agent section changed

[indigo goblet]

hmmm, okay use 16384 instead

on ollama side too

restart both, make sure ollama is started with; export OLLAMA_CONTEXT_LENGTH=16384

if that still doesnt work, we need to use openclaw logs --follow because the .json looks fine and hmm, might be an openclaw bug

[prime fern]

salma@ubuacetha:~$ sudo systemctl restart ollama
salma@ubuacetha:~$ openclaw gateway restart
3

salma@ubuacetha:~$ openclaw gateway restart

🦞 OpenClaw 2026.3.24 (cff6dc9)
I'm the middleware between your ambition and your attention span.

Restarted systemd service: openclaw-gateway.service
salma@ubuacetha:~$ openclaw tui

⚠️ Agent failed before reply: Model context window too small (8192 tokens). Minimum is
16000.
Logs: openclaw logs --follow
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

still same error

i think we should try a new llm

[indigo goblet]

so you confirm ollama restarted with 16348

16384*

[prime fern]

one sec

salma@ubuacetha:~$ export OLLAMA_CONTEXT_LENGTH=16384
salma@ubuacetha:~$ sudo systemctl restart ollama
[sudo] password for salma:
salma@ubuacetha:~$

like that?

[wet stone]

Hi ! It's been hours I'm trying to setup a way to have reminders, like "Remind me in 20 minutes to check this".

I tried to setup it with cron and skills, but I never get a message when the cron is triggered, it seems sent in the heartbeat and nothing else, would be nice if I could get it on Discord DM
For now, I try manually using openclaw cron add --name test --at "30s" --wake now --session main --system-event "Remind me to check my emails!", but the result seems to be lost, I can't find a way to forward it to Discord

Should I try to use it in isolated mode ? I got some results using --announce --channel discord --to <id>, but I doubt that's the best way to achieve a simple reminder system

[prime fern]

Ollama API error 500: {"error":"model requires more system memory (38.9 GiB)
than is available (31.5 GiB)"}

⚠️ Agent failed before reply: Model context window too small (8192 tokens).
Minimum is 16000.
Logs: openclaw logs --follow
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

same error

[indigo goblet]

you need to edit it

[prime fern]

edit what?

[indigo goblet]

sudo systemctl edit ollama.service

then put that;
[Service]
Environment="OLLAMA_CONTEXT_LENGTH=16384"

save and exit

then; sudo systemctl daemon-reload

sudo systemctl restart ollama

[final patio]

I think you should resolve memory issue first since it is less than a requirement.

[prime fern]

maybe that is the problem

sudo systemctl edit ollama.service

everythimg is commented out

/etc/systemd/system/ollama.service.d/.#override.conf3b111a53cfea81e8

Edits below this comment will be discarded

/etc/systemd/system/ollama.service

[Unit]

Description=Ollama Service

After=network-online.target

[Service]

ExecStart=/usr/local/bin/ollama serve

User=ollama

Group=ollama

Restart=always

RestartSec=3

Environment="PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bi>

[Install]

WantedBy=default.target

[indigo goblet]

you run ollama manually from a terminal?

usually?

[prime fern]

yes

[indigo goblet]

hm ok, try something quick,

run

OLLAMA_CONTEXT_LENGTH=16384 ollama serve

run

ollama run qwen3:4b

in another terminal

and check with> ollama ps

[prime fern]

salma@ubuacetha:~$ OLLAMA_CONTEXT_LENGTH=16384 ollama serve
Error: listen tcp 127.0.0.1:11434: bind: address already in use
salma@ubuacetha:~$

[indigo goblet]

ah

right, stop it first

[prime fern]

salma@ubuacetha:~$ sudo systemctl stop ollama
salma@ubuacetha:~$

like that?

[indigo goblet]

yes

[indigo goblet]

now restart from there

[prime fern]

ok it worked

time=2026-03-26T10:33:15.243Z level=INFO source=routes.go:1729 msg="Ollama cloud disabled: false"
time=2026-03-26T10:33:15.244Z level=INFO source=images.go:477 msg="total blobs: 5"
time=2026-03-26T10:33:15.244Z level=INFO source=images.go:484 msg="total unused blobs removed: 0"
time=2026-03-26T10:33:15.245Z level=INFO source=routes.go:1782 msg="Listening on 127.0.0.1:11434 (version 0.18.2)"
time=2026-03-26T10:33:15.246Z level=INFO source=runner.go:67 msg="discovering available GPUs..."
time=2026-03-26T10:33:15.247Z level=INFO source=server.go:430 msg="starting runner" cmd="/usr/local/bin/ollama runner --ollama-engine --port 44823"
time=2026-03-26T10:33:15.292Z level=INFO source=server.go:430 msg="starting runner" cmd="/usr/local/bin/ollama runner --ollama-engine --port 36033"
time=2026-03-26T10:33:15.401Z level=INFO source=runner.go:106 msg="experimental Vulkan support disabled. To enable, set OLLAMA_VULKAN=1"
time=2026-03-26T10:33:15.402Z level=INFO source=server.go:430 msg="starting runner" cmd="/usr/local/bin/ollama runner --ollama-engine --port 42769"
time=2026-03-26T10:33:15.446Z level=INFO source=types.go:60 msg="inference compute" id=cpu library=cpu compute="" name=cpu description=cpu libdirs=ollama driver="" pci_id="" type="" total="30.8 GiB" available="26.1 GiB"
time=2026-03-26T10:33:15.446Z level=INFO source=routes.go:1832 msg="vram-based default context" total_vram="0 B" default_num_ctx=4096

new terminal?

[indigo goblet]

can use that one, ollama is stopped

ollama run qwen3:4b is another terminal yes

and check with ollama ps

if it print context at 16384 then it worked

[lilac pulsar]

so what are soem free models we can try otu rn

[prime fern]

salma@ubuacetha:~$ ollama ps
NAME ID SIZE PROCESSOR CONTEXT UNTIL
qwen3:4b 359d7dd4bcda 5.3 GB 100% CPU 16384 4 minutes from now
salma@ubuacetha:~$

[lilac pulsar]

liek we can run openclaw with exactly 0 or around 10$ a month cost

[prime fern]

this also worked

[lilac pulsar]

as i saw a website where we can run openclaw onbrowser

shoudl i use that or thats not safe

[indigo goblet]

ahh finallyyy, ok hm, i'm getting lost in the spaghetti right now, your openclaw is running ?

[lilac pulsar]

and are there any hardware equirements for openclaw to run

[indigo goblet]

if yes, just restart it again, and try

[prime fern]

⚠️ Agent failed before reply: Model context window too small (8192 tokens).
Minimum is 16000.
Logs: openclaw logs --follow
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

[indigo goblet]

use a new TUI, so close that one, it's the same right?

close it, open a terminal and run openclaw tui and see

[prime fern]

yes i closed all terminal

same error

⚠️ Agent failed before reply: Model context window too small (8192 tokens).
Minimum is 16000.
Logs: openclaw logs --follow
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

[indigo goblet]

in a new tui ?

[prime fern]

yes

[indigo goblet]

if yes, run openclaw logs --follow in another terminal

[prime fern]

log is too big

[indigo goblet]

do you see ctx=8192

[prime fern]

should we start dm?

[indigo goblet]

or 4096

or even KvSize=262144? 262k

[prime fern]

low context window: ollama/qwen3:4b ctx=8192 (warn<32000) source=modelsConfig
10:07:59+00:00 error agent/embedded {"subsystem":"agent/embedded"} blocked model (context window too small): ollama/qwen3:4b ctx=8192 (min=16000) source=modelsConfig

KvSize=262144 doesnt exist

[indigo goblet]

ok, you confirm your openclaw.json has 16384 set?

[prime fern]

yes

[indigo goblet]

search for every 8192 in it

might be others

[prime fern]

},
"contextWindow": 16384,
"maxTokens": 8192
},

[indigo goblet]

nothing in context

[prime fern]

10:07:59+00:00 error agent/embedded {"subsystem":"agent/embedded"} blocked model (context window too small): ollama/qwen3:4b ctx=8192 (min=16000) source=modelsConfig

10:07:59+00:00 warn agent/embedded {"subsystem":"agent/embedded"} low context window: ollama/qwen3:4b ctx=8192 (warn<32000) source=modelsConfig
10:07:59+00:00 error agent/embedded {"subsystem":"agent/embedded"} blocked model (context window too small): ollama/qwen3:4b ctx=8192 (min=16000) source=modelsConfig
10:07:59+00:00 error diagnostic {"subsystem":"diagnostic"} lane task error: lane=main durationMs=149 error="FailoverError: Model context window too small (8192 tokens). Minimum is 16000."

10:07:59+00:00 error Embedded agent failed before reply: Model context window too small (8192 tokens). Minimum is 16000.

[indigo goblet]

okay, hmmmmmm

shit is coming from another file

[prime fern]

i think i should just do a snapshot restore and start from scratch

with a different llm

thoughts?

[indigo goblet]

you could yeah, but maybe this annoying thing is coming from another file

hold on

[prime fern]

lets do a fresh install

[indigo goblet]

some stale pain in the ass is still around

[prime fern]

from scratch to rule it out

[indigo goblet]

run this

grep -RIn '"qwen3:4b"|"contextWindow"|8192' ~/.openclaw

[prime fern]

salma@ubuacetha:~$ grep -RIn '"qwen3:4b"|"contextWindow"|8192' ~/.openclaw
salma@ubuacetha:~$

done

[indigo goblet]

ok run, grep -RInE 'qwen3:4b|contextWindow|8192' ~/.openclaw

[prime fern]

/home/salma/.openclaw/agents/main/sessions/sessions.json:79: "model": "qwen3:4b",
/home/salma/.openclaw/agents/main/sessions/sessions.json:87: "model": "qwen3:4b",
/home/salma/.openclaw/openclaw.json.bak.1:37: "contextWindow": 128000,
/home/salma/.openclaw/openclaw.json.bak.1:38: "maxTokens": 8192
/home/salma/.openclaw/openclaw.json.bak.1:53: "contextWindow": 128000,
/home/salma/.openclaw/openclaw.json.bak.1:54: "maxTokens": 8192
/home/salma/.openclaw/openclaw.json.bak.1:69: "contextWindow": 128000,
/home/salma/.openclaw/openclaw.json.bak.1:70: "maxTokens": 8192
/home/salma/.openclaw/openclaw.json.bak.1:73: "id": "qwen3:4b",
/home/salma/.openclaw/openclaw.json.bak.1:74: "name": "qwen3:4b",
/home/salma/.openclaw/openclaw.json.bak.1:85: "contextWindow": 8192,
/home/salma/.openclaw/openclaw.json.bak.1:86: "maxTokens": 8192
/home/salma/.openclaw/openclaw.json.bak.1:95: "primary": "ollama/qwen3:4b"
/home/salma/.openclaw/openclaw.json.bak.1:98: "ollama/qwen3:4b": {}
/home/salma/.openclaw/openclaw.json.bak:41: "contextWindow": 128000,
/home/salma/.openclaw/openclaw.json.bak:42: "maxTokens": 8192
/home/salma/.openclaw/openclaw.json.bak:45: "id": "qwen3:4b",
/home/salma/.openclaw/openclaw.json.bak:46: "name": "qwen3:4b",
/home/salma/.openclaw/openclaw.json.bak:57: "contextWindow": 262144,
/home/salma/.openclaw/openclaw.json.bak:58: "maxTokens": 8192
/home/salma/.openclaw/openclaw.json.bak:67: "primary": "ollama/qwen3:4b"
/home/salma/.openclaw/openclaw.json.bak:70: "ollama/qwen3:4b": {}

[indigo goblet]

can you open both sessions.json:89 and 79

and check for the context in it

[prime fern]

how do i do that?

a fresh install wont be faster/easier?

[indigo goblet]

hmm ok, just move those two out of the way and restart openclaw

[prime fern]

maybe install openclaw through ollama?

[indigo goblet]

will be faster

[prime fern]

ok. how do i open these sessions?

[prime fern]

what does "move them out of the way" mean?

[indigo goblet]

you see; sessions.json right?

in that folder, agents/main/sessions/

move it to your documents or somewhere else

[prime fern]

ok. im in sessions

oh. ok. move it?

[indigo goblet]

yes

[prime fern]

or edit it?

[indigo goblet]

move it out of there, anywhere else, documents, dowloands

downloads

then restart openclaw

[prime fern]

wont that break it

if i move it?

[indigo goblet]

try it

then, after restarting openclaw and the tui is open, use /reset

in the tui

and let's see

[prime fern]

salma@ubuacetha:~/Documents$ openclaw gateway restart

🦞 OpenClaw 2026.3.24 (cff6dc9)
The only open-source project where the mascot could eat the competition.

Restarted systemd service: openclaw-gateway.service
osalma@ubuacetha:~/Documents$ openclaw tui

🦞 OpenClaw 2026.3.24 (cff6dc9)
WhatsApp automation without the "please accept our new privacy policy".

openclaw tui - ws://127.0.0.1:18789 - agent main - session main

session agent:main:main
gateway connected | idle
agent main | session main | unknown | tokens ?/16k

ok no error message

[indigo goblet]

send a message to your agent

[prime fern]

i sent a message

still running

for a minute

now im wondering if i can try a bigger llm

[indigo goblet]

does the model and token are displayed correctly now?

[prime fern]

but Sonny you are a genius. looks like that error was fixed

yes 16k

just slow nnow

[indigo goblet]

ok lets wait if it answer

[prime fern]

2,5 minutes

too slow to be usable

[indigo goblet]

that line ""agent main | session main | unknown | tokens ?/16k"" still the same?

[prime fern]

yes

i wrote hi. and it hasnt responded in 4 min

[indigo goblet]

lol

[prime fern]

im going to try and switch to GPT-OSS 20B

[indigo goblet]

i think its an issue, especially that agent main line

run openclaw logs --follow in another terminal

[prime fern]

reboot system?

[indigo goblet]

and openclaw status --all

and send hi again

[prime fern]

┌─────────────────┬────────────────────────────────────────────────────────────┐
│ Item │ Value │
├─────────────────┼────────────────────────────────────────────────────────────┤
│ Version │ 2026.3.24 │
│ OS │ linux 6.17.0-19-generic (x64) │
│ Node │ 22.22.2 │
│ Config │ ~/.openclaw/openclaw.json │
│ Dashboard │ http://127.0.0.1:18789/ │
│ Tailscale │ off │
│ Channel │ stable (default) │
│ Update │ pnpm · up to date · npm latest 2026.3.24 │
│ Gateway │ local · ws://127.0.0.1:18789 (local loopback) · reachable │
│ │ 48ms · auth token │
│ Security │ Run: openclaw security audit --deep │
│ Gateway self │ ubuacetha (192.168.1.127) app 2026.3.24 linux 6.17.0-19- │
│ │ generic │
│ Gateway service │ systemd installed · enabled · running (pid 24339) │
│ Node service │ systemd not installed │
│ Agents │ 1 total · 1 bootstrapping · 1 active · 1 sessions │
│ Secrets │ none

i wrote hi again

and it stopped processing the request

✓ Restart sentinel: none
! Port 18789
Port 18789 is already in use.

• pid 24339 salma: openclaw-gateway (127.0.0.1:18789)
• Gateway already running locally. Stop it (openclaw gateway stop) or use a different port.
  ! Tailscale: off · unknown
  error: Error: spawn tailscale

[indigo goblet]

whats happened with openclaw logs --follow

[prime fern]

what should i look for in the logs?

10:20:49+00:00 error (node:22198) [DEP0040] DeprecationWarning: The punycode module is deprecated. Please use a userland alternative instead.
(Use node --trace-deprecation ... to show where the warning was created)

[indigo goblet]

run error, unknown model

[prime fern]

10:07:59+00:00 error agent/embedded {"subsystem":"agent/embedded"} blocked model (context window too small): ollama/qwen3:4b ctx=8192 (min=16000) source=modelsConfig
10:07:59+00:00 error diagnostic {"subsystem":"diagnostic"} lane task error: lane=main durationMs=149 error="FailoverError: Model context window too small (8192 tokens). Minimum is 16000."
10:07:59+00:00 error diagnostic {"subsystem":"diagnostic"} lane task error: lane=session:agent:main:main durationMs=151 error="FailoverError: Model context window too small (8192 tokens). Minimum is 16000."

[prime fern]

how do i do that

search?

[indigo goblet]

close the terminal with the logs

open a fresh terminal> openclaw logs --follow

then send 'hi' to your agent again

then whats important here in the logs are run error and unknown model

[slim kiln]

hey guys, im in doubt about what OS to use for hosting openclaw with an anthropic model?. a bit of context, i have an hp elitedesk 800 g2 sff with extra ram and im gonna that for the hosting, in general im gonna use claw for little things, read my newsletters and create like a newspaper for it, set reminders via wpp and by voice messages, use the Productivity in the clawhub and so on.

[prime fern]

🦞 OpenClaw 2026.3.24 (cff6dc9)
iMessage green bubble energy, but for everyone.

openclaw tui - ws://127.0.0.1:18789 - agent main - session main

session agent:main:main

hi

fetch failed
gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/16k

fetch failed

new error

ok. wrote hi again. processing

by some divine miracle i was able to chat with openclaw previously with a bigger model

[indigo goblet]

might be this model is too small?

[prime fern]

what is frustrating is that even if i document the steps i cant recreate that

ok i will switch model

[indigo goblet]

we can continue and check if the model is actually loading busy with ollama ps

while the message is hanging

[prime fern]

salma@ubuacetha:~$ ollama ps
NAME ID SIZE PROCESSOR CONTEXT UNTIL
qwen3:4b 359d7dd4bcda 5.3 GB 100% CPU 16384 About a minute from now
salma@ubuacetha:~$

[indigo goblet]

hm ok

run this in a new terminal and past the output

openclaw models status --agent main

[prime fern]

salma@ubuacetha:~$ openclaw models status --agent main

🦞 OpenClaw 2026.3.24 (cff6dc9) — You had me at 'openclaw gateway start.'

Config : ~/.openclaw/openclaw.json
Agent dir : ~/.openclaw/agents/main/agent
Default (defaults): ollama/qwen3:4b
Fallbacks (0) (defaults): -
Image model (defaults): -
Image fallbacks (0) (defaults): -
Aliases (0) : -
Configured models (1): ollama/qwen3:4b

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

• ollama effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=0, token=0, api_key=1) | ollama:default=marker(ollama-local) | models.json=OL...EY | source=models.json: ~/.openclaw/agents/main/agent/models.json

OAuth/token status

• none

[indigo goblet]

ok lol mmh

running out of ideas

ok

open ~/.openclaw/agents/main/agent/models.json

anything weird in it? context?

[prime fern]

ok

one sec

[indigo goblet]

whats in the qwen section for exmaple

[prime fern]

nope. same as openclaw.json file

[indigo goblet]

ok lol, I guess that model is too small for openclaw which is why its stuck even with 'hi'

[prime fern]

is the context window of 16k optimal?

[prime fern]

yea. as much as i appreciate all the help so far

[indigo goblet]

or it's the hanging bug

[prime fern]

i think better to do a fresh install. because it shouldnt be this complicated

im going to try and switch to gpt oss

beforehand

[indigo goblet]

yeah, it isnt normally, one of the reason I avoid local models

[indigo goblet]

yeah sure try and see

[prime fern]

hi

connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

eventhough i went into configuration and changed the llm all that it did was change the tokens to 262k

jesus

honestly bro. this is trash

[indigo goblet]

you did a restart yes?

[prime fern]

yes

i restarted both ollama and openclaw

hi

gateway connected | idle
agent main | session main (openclaw-tui) | ollama/qwen3:4b | tokens ?/262k

still showing qwen

wtf man

rough edges is an understatement

[sterile ingot]

Does anyone know how i can connect openclaw to my browser so it can display the web ui , since i don't wanna use the terminal

[high river]

there skill for it i guess int clawhub

[prime fern]

ollama launch openclaw --model gpt-oss:20b

[late lagoon]

if I have a large number of skills, passing all their json into the context window eats up way too many tokens
how are you guys handling this

[prime fern]

trying this now

ok that fixed the model

but its still running slowly

wtf

[indigo goblet]

hmm yeah, local models can be a pain

[final patio]

opeclaw dashboard

[prime fern]

all the youtube video make it sound like this is the whole point. you can run it locally and securely using free llm

[sterile ingot]

yeah how do i open it

[prime fern]

plus i got it working before

very whimsical

[sterile ingot]

i just fixed it nvm

[prime fern]

@indigo goblet fresh install?

[final patio]

👍

[prime fern]

ill create a snapshot right now. so can return later if the fresh install doesnt work

[sterile ingot]

can u check ur dms

[indigo goblet]

to be honest at this point, I'd run codex

grab a plus sub and use a lower model, usage should be fine

[prime fern]

instead of fresh install

[indigo goblet]

well youtube is youtube, unless you have a crazy setup and use the best local model

then yeah

I mean, depend what you want to do

but high end model, the best, are not local.

[indigo goblet]

yeah that's what I would do

[prime fern]

well my original goal was to have cheap local llms for basic stuff

editing google calendar

heartbeats

and then expensive llms like opus 4.6 for complex tasts

[indigo goblet]

rerun onboard, use OAuth for codex, then done and done, your agent is running, any issues, pair programming and work with it to fix stuff,

[prime fern]

but i cant even get the local llms to run

[prime fern]

im rerunning onboard

to make sure i dont make a mistake

how do i "use OAuth for codex

[indigo goblet]

you need chatgpt sub

[prime fern]

im on the Model/Auth provider

[indigo goblet]

20 bucks a month

[prime fern]

ok. but that doesnt solve the locall llm thing

[indigo goblet]

this is not about solving the local llm

locals are headaches in my opinion, feel free to pursue that

[prime fern]

ok thx again for all the help bro

[indigo goblet]

no worries youre welcome

[prime fern]

ill set it up with chatgbt

im also going to try a fresh install

be back shortly after fresh install

[half spear]

I posted on the #1459642797895319552 channel but maybe somebody here has seen already a behaviour where every openclaw command just fails with exit code 1 and no output at all

[indigo goblet]

bruv, Krill is gone?

[rustic carbon]

Whats better putting 10 euro credit in openrouter or gemini

[acoustic solar]

After the last update the openai oauth is not working for me, keep on getting I reached API limits. My account just renewed a week ago. I am not a heavy user yet.

can anyone help

[high heron]

buying subscription. not api

and they're both sht. dont use openrouter and gemini

[nimble hill]

Is it a timeout error, if yes then append --timeout 30000 or higher

[small hedge]

Hey everyone! I'm loving OpenClaw, but I'm struggling with the 'token bleed' in long sessions.
I recently spent €5 on just 100 messages because the prompt size balloons so fast (Claude Sonnet 4.6).

I've temporarily switched to Claude 3 Haiku ($0.25/M input) to keep costs down, but it feels like a band-aid.
The real issue is that OpenClaw sends the full chat history with every turn.
I'm thinking about a setup that:

1. Aggressively compacts the active history after ~10 turns.
2. Offloads the rest to RAG (Memory Search), so the agent only pulls in specific 'recalled' snippets when needed.
3. Sends only the minimum required context to the LLM.

I saw the Anthropic 1-hour prompt caching option, but for my use case (intermittent chatting throughout the day), it doesn't seem like it will save much if the cache expires between my messages.

My questions for the pros here:
• How are you guys fighting these high API costs?
• Is there a recommended config for a 'Sliding Window + RAG' approach that actually works without making the agent lose its mind?

Would love to hear how you're optimizing your openclaw.json to keep things lean!

[sterile ingot]

is there a way to allow full access on my laptop to openclaw?

[fiery arch]

Happy to help you getting this working on the OG https://nono.sh

[indigo goblet]

yes, what do you want to do with it

[sterile ingot]

like give it completely unrestricted access to do tasks for me , like move my mouse and use my keyboard instead of me

[indigo goblet]

you wont see those but it can do it

happens behind the scene

[sterile ingot]

okay how do i set up it?

[indigo goblet]

setup what? openclaw?

[copper nova]

do you have anything personal logged in or on your laptop?

if you do, I suggest you stop right there

[sterile ingot]

yeah i do

[copper nova]

yea.... don't do it

[sterile ingot]

hmm

[copper nova]

if someone gets access to your bot, and I ask it to give me your google passwords

you not gonna like that

go rent a VPS for $5 to mess around with

[sterile ingot]

yeah but how is he gonna do that if the bot is local

[copper nova]

create an account for your bot

[sterile ingot]

yeah that should work

[copper nova]

create new email, new everything for your bot

don't log in with your shit

[sterile ingot]

bet

[indigo goblet]

it's not like sudo give knowledge about gmail passwords, depends on what he wants to do with it, hence what do you want to do with it

also depends on security and rules for the agents

[hexed terrace]

and also depends on the model the agent runs

[sterile ingot]

well let's say i want to give the bot access to specific 3d creation platform so it can do the tasks for me

[copper nova]

my bot ignores my rules all the time

[indigo goblet]

yes, local models = forget about it

[copper nova]

I run on 0 trust system

[indigo goblet]

well that depends on his source of truth files and memory, mine doesn't

[copper nova]

should look into prompt injection attacks etc.

it's a thing

it's better to be safe than regret later

all these rules are "soft" rules, better to just remove the ability to fuck you all together

[indigo goblet]

yes it is, I have specific rules for that, also, prompt injection usually happens on freeroam and unmonitored web browsing.

[copper nova]

it can't fuck you, if there is nothing to fuck

just saying

[indigo goblet]

been a long time latest high end model do not respont to basic prompt injections

[copper nova]

rather saftey through that, than rely on these "soft" rules to protect my most important information

but you do you, I ain't your dad

[indigo goblet]

it can, what platform are we talking about exactly? and it probably won't be set up with a couple of prompts, keep that in mind, see it more like a project and multiples step.

[sterile ingot]

let's say Roblox studio

[indigo goblet]

agree to disagree, source of truth files are not soft rules, again that's just my experience and how my agent is set up. Full access is not bad, but it needs discipline and logic.

and structure too

[ancient jungle]

i made mine its own gmail account and just forward my mail to it, that way it can still alert me if something that looks important pops up (i never do), but he still can't accidently just delete all my emails i have on my own account

[indigo goblet]

probably can, but I'm not familliar with anything roblox related, so you'd have to look into it and figure what are the first steps.

[sterile ingot]

yeah that's the problem it's either super complicated or takes a lot of time which i don't have 😆

[indigo goblet]

well, I'd say there is no choices lol, if you want to properly get there without having stuff break in random places

[sterile ingot]

i think zapier mcp is good for connecting some apps and tools with limited access

[indigo goblet]

yeah seems interesting

did you look into roblox mcp

[sterile ingot]

well there are specific ai studios that are build exactly for that to simplify the scripting and modeling and are easy to connect but they have subscriptions as well otherwise u barely get some tokens

actually what i found interesting about openclaw is that u can make and config telegram bots the way u want

[pallid haven]

Hello! I built a chatbot on n8n, but it sometimes gives incorrect answers. The most common issue is losing context in dialogues longer than five sentences. It's not ideal for the client to write to the n8n bot directly and receive a wrong answer.
So, I am thinking about setting up OpenClaw to validate the n8n responses. In this setup, OpenClaw would act as middleware between the client and the n8n bot. It would send the client's message to the n8n bot, and then forward the bot's response to the client only if it thinks the message is valid. If OpenClaw detects a wrong answer, it will trigger a second scenario instead. This way, the client won't receive invalid messages. Can such a workflow be realized with OpenClaw?

[high river]

how to attach image in here?

[half spear]

no it's not timeout and actually

clawbot@synology:~$ time openclaw gateway status --timeout 30000

real    0m4.189s
user    0m2.647s
sys    0m0.298s

[copper nova]

https://giphy.com/gifs/paramountplus-rupauls-drag-race-all-stars-8-qMR8Oqpv3DgbLdXmoE

[wicked urchin]

Hello I was wondering if someone would be able to help me out. My bot doesn't seem to want to do anything on it's own. seems to have pretty strict guardrails and it stops constantly on tasks to ask me for approval. is it because I'm using open ai?

[ancient jungle]

can you oauth codex business plan?

[quaint shell]

Hey! 👋

Building a document management app for a VTC company — React Native + Supabase + Next.js admin. Two roles (drivers upload invoices, admin sends payslips/legal docs), RGPD compliance needed.

I'm on Max 5x and want to go all-in on Claude Code multi-agent workflows. Running it on a VPS with Docker for overnight autonomous sessions.

Would love to hear from people experienced with Agent Teams:

• How do you split your agents for a fullstack project (mobile + web admin + backend)?
• Auto mode or --dangerously-skip-permissions in Docker for long runs?
• Any good CLAUDE.md patterns for monorepos?

First time setting up a serious multi-agent workflow, any tips appreciated 🙏

[hexed terrace]

probably not because you’re using openai. this is usually exec approval behavior or something that is injected into context like AGENTS.md

[wicked urchin]

Well I keep trying to tell it that it has free reign. that the PC it's running on is specifically for it and it just refuses to listen lol

From the get-go I said "I want a fully hands-off experience. If I give you a task, I expect you to do said task without needing approval or intervention from me and there is no risk of anything dangerous happening because this enviroment is isolated"

and it was like "Yeah, I'm gonna ignore that... I cannot bla. bla. bla

[odd forum]

what are trying to do?

[hexed terrace]

this usually won’t override the runtime. try looking into exec policy first

[odd forum]

Don't run all 5 on the same code. Split them:

• 1-2 agents on mobile
• 2-3 on the admin panel
• 1 on backend/schema stuff

They'll step on each other if everyone touches the same files.

───

RGPD in plain English:

• Encrypt the files (Supabase does this automatically)
• Drivers can't see other drivers' stuff (database rules)
• Log who accessed what (for legal stuff)

───

That's it! Don't over-engineer on day 1. Start with 3 agents, see how it feels, add more if needed.

[wicked urchin]

exec policy? interesting. I'm not familiar with that, I'll look into it, thank you

[odd forum]

The problem: OpenClaw keeps sending your entire chat history. 100 messages in = €5 gone.

The fix: Cap your history at 10-12 turns. Old stuff gets summarized, not deleted.

[quaint shell]

what you mean by not run on the same code ?

you need to run all in same time

in the same cli terminal

[odd forum]

You can't run multiple Claude Code agents in the same terminal window. They'd fight each other

[quaint shell]

oh you need one agent for one terminal ?

[thorny yacht]

Hi all, I need some help! I started a fresh rig but its incredibly unresponsive. I think it could be LLM rate limiting but what's the way to be sure its not something to do with my new setup?

[odd forum]

Yes, exactly.

One terminal = one agent.

You can't run two Claude Code sessions in the same terminal window — they'll fight over the keyboard and you'll get gibberish.

Your options:

Option 1: Multiple terminals (easiest)

• Open 3 terminal tabs
• Tab 1: claude working on mobile app
• Tab 2: claude working on admin panel
• Tab 3: claude working on backend

Option 2: One terminal, run them one after another

Mobile agent

claude
...work...
/exit

Admin agent

claude
...work...
/exit
Option 3: Use tmux/screen
Split one terminal into multiple panes, each with its own Claude session

[odd forum]

Check your logs for 429 errors or delays if you see them it’s rate limiting, if not it’s likely your setup (CPU, network, or prompt size) causing the slowdown.

[odd forum]

I opwuld run OpenClaw as a guardrail layer that scores/validates n8n outputs before sending, and fallback if confidence is low.

[odd forum]

Usually means a broken install or config maybe run openclaw doctor and check logs, it’ll point to what’s failing.

[odd forum]

Go with Ubuntu and add Ollama alongside it—gives you a solid local fallback to cut API costs.

That usually means the service isn’t running or the port isn’t exposed // check your OpenClaw service status and firewall/Docker ports.

[thorny yacht]

Thanks! no 429 errors, but some errors saying it can't find memory.md. Can't imagine its CPU/network/prompt, as all computer systems are idle and the prompt is tiny Could the missing memory file be doing it? I don't know why that would be missing.

[quaint shell]

im talk about agent teams

[cunning meadow]

How do you check if a model is timing out during a telegram or discord channel conversation?

[odd forum]

missing memory.md can break context/memory loading, create it (even empty) or disable memory to test.

[odd forum]

Same idea // one agent per terminal/worktree, even in teams, so they don’t step on each other’s changes.

[odd forum]

Check your agent logs for timeout or long response gaps if replies stop after a delay, it’s usually the model timing out.

[cunning meadow]

Thanks for the reply. I'm new to openclaw. Where do I find the agent logs?

[odd forum]

Check the terminal running OpenClaw or the ~/.openclaw/logs folder—that’s where agent logs are stored. Do you know how to use termiinal its easier if you Ollama to bootstrap it

[copper nova]

but I do it all the time?

[cunning meadow]

Yes I can use terminal to inspect the logs. I don't understand about ollama to boostrap it tho.

[half spear]

doctor has the same behaviour... I have an idea tho, so I try a fresh local node install instead of the one included in synology

[odd forum]

If your OC is new type thsi in Terminal ollama launch openclaw then you will have a LLM to help you // the doc are here - https://docs.ollama.com/integrations/openclaw

[undone zinc]

So guys, to stay in context yestserday i was all the day trying to configure a local ia on my laptop with a 4060 and 32gb ram, i know is a laptop but i thing is a good one for only this use in ubuntu.
But trying and trying like 12 hours, i was able to use just ollama but not local if not on cloud and everything broke.
So now do i need to reinstall just claw or better reinstall ubuntu?

[odd forum]

Yeah that’s likely it Synology’s bundled Node can break things, a fresh official Node install usually fixes exit code 1 issues.

[cunning meadow]

Oh I see. Thank you tried this when I first started

[odd forum]

Your laptop is fine — 4060 + 32GB is solid for local AI.

Don't reinstall Ubuntu. That's overkill.

Just reinstall OpenClaw:

[undone zinc]

Ok i will try to clean uninstall it

[odd forum]

I'm three months in and still use Ollama as Ollama handles everything automatically

[high river]

Cron Job had bug which not delivered to channels huh?

it delivered inside Gateway UI Control.

yeah it showed and it got deleted by itself after refreshing?

[undone zinc]

So with what ia do i need to start with?

[odd forum]

just type ollama launch openclaw in termial and you might not have to unistall it

[odd forum]

Check your channel config

[thorny yacht]

Created empty memory.md which the system recognises. Re-started gateway. But the same. No errors in logs. I sent a new message and I can see from LLM dashboard that it received a message but no response...

[high river]

umm specifically? i am using telegram

[undone zinc]

gateway connect failed when i try to open that, with ollama 3.2.3

[half spear]

yes, I am amidst installation but now onboard has started

[odd forum]

If the LLM receives it but doesn’t respond, it’s likely stuck on a tool or waiting—try disabling tools to isolate it.

[odd forum]

That’s usually Ollama not running or wrong URL check it’s up on localhost:11434 and matches your OpenClaw config.

[small hedge]

Thank you @odd forum is there an example how to implement this fix? I got the idea, just want to get some more details about implementing it, as I‘ve started with OC just recently

[odd forum]

that’s a good sign, onboarding starting usually means the install is now working 👍

[odd forum]

What model are you using right now?

[odd forum]

Check your bot is still connected

[high river]

yup it still it a bug i guess

[small hedge]

Claude 3 Haiku

[thorny yacht]

Slow or no response

[odd forum]

Yeah sounds like a UI bug...Use CLI instead of Gateway UI to find it

[high river]

i can chat with them normally but the cron job that should report to me every 1 minutes isnt really executed to Telegram and UI Controller

22:18 report: Notion workspace excavation remains systematic. Ongoing cataloguing of fetched and traversed blocks; full structure is still being charted—percentage complete metric pending. No requirements for external notification. Minute-by-minute summaries ongoing until task end.
22:19 update: The Notion workspace excavation proceeds with disciplined consistency. Blocks continue to be collected and logged; total structure remains under analysis, preventing precise progress calculation at this stage. No order to communicate with external recipients. Further operational summaries to follow at each interval.
22:21 status—Notion workspace excavation is proceeding efficiently. Blocks are continually gathered and indexed; assessment of total progress is still pending completion of the overall mapping. No orders to send updates to external recipients. Minute-by-minute status reports will persist until the job completes.

[odd forum]

Haiku is cheap but dumb. If you fix the context bloat, you can switch back to Sonnet and still save money.

My suggestion:

1. Add the config below (works with Haiku too)
2. Try it for a day
3. If costs are down, switch back to Sonnet:

{
"llm": {
"model": "claude-3-haiku"
},
"context": {
"maxTurns": 12,
"compressOld": true,
"keepRecent": 4
}
}

[high river]

[3/26/2026 10:22 PM] M F: status?
[3/26/2026 10:22 PM] Commander White: Progress update: The Notion workspace sweep is underway. However, the earlier subagent invocation encountered a fetch failure, which likely prevented any data retrieval from your Notion.
At this moment, there is no new Notion block data fetched yet. If you want me to restart the detailed Notion inventory (with your model preferences and progress reports active), please confirm so I can relaunch the process with proper diagnostics and continuous updates as scheduled.

Standing by for your confirmation or any adjustment, Master Manz.

[small hedge]

Great, Yes my idea was to switch back to sonnet. Will try your configuration 👍

[high river]

how about DeepSeek 3.2 anyone used it?

[odd forum]

Sounds like the cron is running but not dispatching. are you explicitly calling the Telegram tool/send action in the job?

[odd forum]

let me knwo how you get on. Maybe use a local model instead cheaper for those check email cron jobs

[thorny yacht]

@NotPhilSledge thanks but how can I disable tools and how will I know its effect?

[odd forum]

why use that model?

[high river]

yepp..
it does inside cron ui conrol

Delivery
Choose where run summaries are sent.
Result delivery

Announce summary (default)
Announce posts a summary to chat. None keeps execution internal.
Channel
Telegram
Choose which connected channel receives the summary.
To
-1003xx519xxxx:topic:467
Optional recipient override (chat id, phone, or user id).

[high river]

just asking for budget AI to play with

[odd forum]

just use ollama its free

[odd forum]

Disable tools:
openclaw config set tools.enabled false
What it does:

• Stops OpenClaw from using any skills (web search, file edits, etc.)
• Pure text-only chat with the LLM
• Faster, cheaper, but limited

How you'll know it worked:

• Try asking "search for X" — it should say it can't search
• Responses should be instant
• No [tool] blocks in replies

Why try this:
If your issue is the LLM getting stuck on tool calls, disabling them proves it. If chat works without tools → you found the problem.

Re-enable later:

[high river]

what platform/provider? openrouter? self host?

[odd forum]

https://docs.ollama.com/integrations/openclaw

[odd forum]

thsi will sort out yout TG probelms too

[high river]

my bad i dont undrstand man

[high river]

nice thanks so much will try

[undone zinc]

Ok i run it finally, but when open control panel and writing on chat just my message dissapear

  Open the Web UI:
    http://localhost:18789/#token=*


  Quick start:
    /help             see all commands
    openclaw configure --section channels   connect WhatsApp, Telegram, etc.
    openclaw skills                         browse and install skills

  The OpenClaw gateway is running in the background.
  Stop it with: openclaw gateway stop


🦞 OpenClaw 2026.3.24 (cff6dc9) — Automation with claws: minimal fuss, maximal pinch.

 openclaw tui - ws://127.0.0.1:18789 - agent main - session main

 session agent:main:main
 connected | idle
 agent main | session main (openclaw-tui) | ollama/llama3.2:3b | tokens ?/131k```

[cunning meadow]

What model do you use? You have a powerful GPU running a local model?

[odd forum]

I'm using Kimi2.5 cloud for my smal jobs but Codex for teh bigger jobs

[odd forum]

Is Ollama running? type ollama ps

[undone zinc]

Yes it is running but not answers on claw just with that version, but if i do
ollama run llama3.2:3b "are you working?"
Its answer me

[odd forum]

If Ollama works directly but not in OpenClaw, the connection between them is broken.

Quick fix:

Check OpenClaw is pointing to the right Ollama URL:
openclaw config get llm.baseUrl
Should be http://localhost:11434 (default Ollama port)

[craggy gulch]

curious_dog77

[hexed ermine]

Setting up OpenClaw. Teaching myself a lot about computers while doing it. I’m using my gaming PC and going to put it in a container with docker. Any tips to keep it safe? Also, is it worth spending the money for more quality from OpenClaw?

[undone zinc]

    "mode": "merge",
    "providers": {
      "ollama": {
        "api": "ollama",
        "apiKey": "ollama-local",
        "baseUrl": "http://127.0.0.1:11434",
        "models": [
          {
            "contextWindow": 131072,
            "cost": {
              "cacheRead": 0,
              "cacheWrite": 0,
              "input": 0,
              "output": 0
            },

do i need to change it for localhost?

[odd forum]

just use a vm

[undone zinc]

I changed but still not working with claw

Keeps on running and not answer me with local mode

[ivory shadow]

Hello anyone faced this? their agent returning only "HTTP 401: User not found."

[odd forum]

The best approach is to use a reverse prompt with Ollama, then switch from a local model to Kimi 2.5 in the cloud. You can even ask Ollama to handle that transition for you.

[odd forum]

Where did you get your API key from?

[fallen plinth]

Gateway aborted: config is invalid.
channels.discord: Unrecognized key: "channelIds" ? is there any other method

[odd forum]

That key’s outdated - channelIds (plural) is wrong. Use channelId (singular)

[fallen plinth]

thx :)

[fallen plinth]

Gateway aborted: config is invalid.
channels.discord: Unrecognized key: "channelId"

yea

[odd forum]

those keys were removed, just reconfigure Discord via openclaw configure --section channels and let it generate the correct format.

[fallen plinth]

is there a way to configure it easier?

[odd forum]

forget editing the config file. Use this:
openclaw channels setup discord
It will:

1. Ask for your bot token
2. Show your servers
3. Let you pick which channel(s)
4. Save it correctly

That's it. No more config file headaches.

[undone zinc]

Ok i get it now on kimi-k2.5:cloud
So what exactly need to ask it for do transition?

[stoic anvil]

Hey, welcome!. Few tips: set it up in a VM if you can, use tailscale for remote access instead of exposing ports, dont give it permissions it doesnt need, and be careful with third party skills from clawhub — always check the source.
For the money — openclaw itself is free, the cost is the AI model behind it. Depends on what you’re doing — for simple tasks use cheaper models,use the expensive ones for complex work. What are you planning to use it for?

[odd forum]

Just give it a clear task like “analyze this and propose next steps"

see you all later // In Peter we Trust // catch me on twitter if you need any more help @clawphilsledge

[hexed ermine]

I’m going to automate content creation, so I think the more I spend the better quality

[rotund cairn]

I'm getting "API rate limit reached" when I send "Hi" on WhatsApp ... I think openclaw is sending a bunch of retry. How do I stop that so I don't get blocked?

[fallen plinth]

is there a webui to configure it

[odd forum]

try this //

Open this in your browser:

ou'll see the Gateway UI with:

• Channels tab (add Discord there)
• Config editor
• Logs viewer

If that doesn't work, check the gateway is running:

openclaw gateway statusOr access it via:

[slow monolith]

anyone know how to add more sessions / chatboxes in dashboard webui?

[odd forum]

Use /new or create a new session via the UI/CLI—each session becomes a separate chatbox

[stoic anvil]

That’s a good usecase! For the creative writing parts a better model definitely helps. But you can save a lot by routing the simpler stuff like formatting and scheduling through cheaper models. Openclaw supports model routing so you get the best of both worlds

[hexed ermine]

Okay, awesome! Thanks so much!!

[slow monolith]

when i type /new, the existing chatbox will gone

[red cairn]

Hello, I’m trying to switch my my OC bot from the Anthropic API to 0auth and getting this error:

HTTP 401: authentication_error: Invalid bearer token (request_id: req_011CZRzDJkht776cqh6ByJ44)

—

I’ve been going in circles. Anyone else have this issue?

[tender cargo]

Anyone have any luck having openclaw run very long sessions (2-6 hours) without timeing out or having other issues?

[small hedge]

@odd forum thank you, I tried to research the OC documentation, but unfortunately I don't see any reference to llm.model, context.maxTurns, context.compressOld, or context.keepRecent as valid configuration keys. I see a feature request for maxTurns, but it is in opened state, so was not yet impelemented: https://github.com/openclaw/openclaw/issues/9912#issuecomment-3867469129
Could you point me to where these keys are documented? I want to make sure I'm using the right config before applying it.

[undone zinc]

Its amazing how its impossible to me run ollama/llama3.2:3b really its frustrating

[stray fjord]

which model should I use on 8gb ram

[odd forum]

Just use a cloud model

[odd forum]

Cloud

[stray fjord]

run out of tokens

already used api limit

[odd forum]

Use ollama models they are free

[stray fjord]

downloaded plenty of them some says
Model context window too small (8192 tokens)
or
Ollama API error 400: {"error":"registry.ollama.ai/library/deepseek-r1:7b does not support tools"}

[dark glacier]

If I reach the weekly limit in OpenAI OAuth, what happens?

[copper nova]

you wait for timeout to expire

[dark glacier]

But the Minimax Model is bad and even Kimi is not working

[copper nova]

hook up another provider, I like github-copilot

[limber hollow]

People concerned about price can use openrouter/free, or set primary to openrouter/auto and then configure a fallback to a free model on the openrouter site. \ "agents": {
"defaults": {
"model": {
"primary": "openrouter/auto"
},

[trim osprey]

i just got a second claude account and trying to get it oathed in, i am trying everything to get it switched over but my claw is just not respecting it and falling back to codex sub

i have ran onboarding and pasted setup token
ran openclaw configure and ran setup token
ran doctor --fix after all of these

Any ideas? my clanker and i cant figure it out

[undone zinc]

But for cloud i need to pay for it

[storm sage]

Thank you very much. I will try this. Perhaps I can dm you if you don’t mind? As I got some questions. I’m still new to this.

[fallen plinth]

control ui requires device identity (use HTTPS or localhost secure context)

?

[hexed terrace]

how are you opening the control ui? it should work http on localhost

[fallen plinth]

its on my vps. and i try to open it on my phone

[odd forum]

That model doesn’t support tools—use a tool-capable model or disable tools in your agent config.

[odd forum]

No you don't with ollama apart from the internet

[hexed terrace]

you would need a proxy/tunnel

[odd forum]

Sounds like it’s still using the old provider—try clearing cached creds (~/.openclaw/credentials) and re-running setup so it picks up the new Claude account.

[fallen plinth]

does OpenClaw have this in config?

[trim osprey]

thank you ill try

[odd forum]

Set openrouter/auto as primary and add a free fallback so it doesn’t burn paid credits.

[hexed terrace]

not sure what you have on your phone, it will work with tailscale or a ssh tunnel (ssh -N -L ..)

[odd forum]

Which VPS are you using?

[fallen plinth]

Debian 12

[odd forum]

Sorry are you paying for the VPS or running it locally?!

[fallen plinth]

paying

[spiral parcel]

Anyone know why the openclaw.json dosent update the ""lastTouchedVersion": "2026.3.23-2"," to the new one after updating? :S

[hexed terrace]

probably expected, I don't think is tied to openclaw update

[odd forum]

Name?

[fallen plinth]

?

[odd forum]

Sorry who is the provider of your VPS?

[trim osprey]

it was an issue with new key not getting picked up in auth profile for some reason but i got it working, now i have this error Config observe anomaly: C:\Users*.openclaw\openclaw.json (size-drop-vs-last-good:40433->15014)

[fallen plinth]

Index-Hosting

[odd forum]

That provider likely has networking/firewall quirks—check ports (18789/11434) are open and no proxy is interfering with requests.

[odd forum]

That means your config got overwritten—restore from backup or re-run openclaw configure to rebuild it.

[fallen plinth]

18789 and 11434 are open

[odd forum]

Try running a simple check from your server like curl http://localhost:11434 — if it fails, something on your VPS (like a firewall or proxy) is blocking it even if the ports look open.

[fallen plinth]

$ curl http://localhost:11434/
Ollama is running

[stray fjord]

warning: ignoring --model minimax-m2.7:cloud because cloud is disabled
how to enable cloud

[odd forum]

If Ollama is working, just run openclaw configure and make sure the model/provider is set to your local Ollama (http://localhost:11434), then restart OpenClaw.

[fallen plinth]

ok?

why?

i just want to test the webui

[odd forum]

Type this in terminal

Ollama launch openclaw and choose the right model

[odd forum]

Can you talk to Ollama?

[fallen plinth]

yes

[odd forum]

Reserve prompt to ollama and ask it to repair or use Openclaw Doctor —fix

[fallen plinth]

but why? I don't get it 😅

[odd forum]

Let's start again lol // what's wrong?

[fallen plinth]

.

[stray fjord]

it says cloud is disabled

[knotty violet]

hello guys who have a guide to make openclaw with gpt 5.4 an pentester assistant?

[odd forum]

https://docs.ollama.com/integrations/openclaw

[storm sage]

my version doesnt seem to support reserveTokensFloor:

openclaw config set agents.defaults.reserveTokensFloor 20000

Error: Config validation failed: agents.defaults: Unrecognized key: "reserveTokensFloor"

[hollow owl]

Hi — Telegram channel silently fails to start after a config rollback. openclaw status shows empty Channels table. Zero [telegram] lines in gateway startup logs at all. Config is valid, token verified working via curl getMe, openclaw doctor sees the channel. Version 2026.3.24, macOS Apple Silicon, M4 Mac Mini. Been broken all day despite multiple reinstalls, config fixes and restarts.

[fallen plinth]

and?

[little cove]

Hello sonny yes o had asked him to start a DM, but it's not working...

[tawny tiger]

hey people, can someone help me?
I installed openclaw on my machine. my bot react to my discord messages, with eyes, thumbs up andthinking emoji, but no response

[open saffron]

Apparently I need some serious help to resolve this. If I just chat with my agent everything is fine but if I tell it to do something... I get this:

I got caught in a loop — kept saying "I'll do it" without actually doing it. And I may have hallucinated completing things. Let me actually CHECK right now with real tools what the status is: I got stuck in a pattern — kept "saying" without "doing". And I keep "saying" instead of "doing". Let me actually verify with real tools I got stuck. Time to actually check Let me stop talking and actually verify I got stuck in loops of "saying I'll do it" without actually calling tools Let me actually investigate with Let me check — apologize I got stuck in loops. Let me actually verify with real tool calls: got stuck in loops of "saying" instead of "doing". Let me actually verify with real tools: I got stuck in a loop — kept "saying" instead of "doing". Time to actually verify with real tools. Let me stop talking and actually check. I got stuck — kept saying "I'll do it" without actually doing it. Let me actually verify with real tools: got stuck in loops........

...and that goes on and on. it does eventually stop but... I mean... whew.

[mighty rampart]

What model are you using and how is it connected?

[open saffron]

API to Kimi-k2.5 This was working and it has been getting worse.

[rustic pier]

How do we get Krill to respond?

Or is he just running slow atm?

[open saffron]

I'm guessing it's down at the moment.

[high heron]

/think xhigh

or use

/think high

[hollow owl]

ChatGPT fixed it after 7 hours of going around in circles.

OpenClaw Telegram issue summary (2026.3.24 / macOS LaunchAgent):

I had a valid Telegram setup (botToken worked, curl getMe worked, config validated), but OpenClaw would sometimes boot with zero Telegram/provider startup lines in the logs. No gateway/channels/telegram, no starting provider, no telegram/network — so the Telegram channel was not being initialised at startup at all.

Important part: this was not a BotFather/token problem and not mainly an allowFrom problem. When Telegram did come up, it worked. The failure was that the channel subsystem sometimes didn’t start Telegram on boot.

What finally proved it:

• valid channels.telegram config
• getMe worked
• foreground/manual gateway startup still showed no Telegram init lines
• restarting the gateway later would sometimes make Telegram work again

So this looks like an intermittent startup/init bug/regression in the current build, not a normal config mistake.

Also hit two side issues while debugging:

• if dmPolicy = "open", allowFrom must contain ["*"], otherwise config validation fails
• turning sandbox on without Docker breaks replies completely, but that is a separate issue from Telegram startup

[stuck flare]

hii hello i am new here ,also recently started with oppenclaw .i nned a help,is there a reliable way to let openclaw sub-agent to control a browser so that it can control browser so that the agent cam fill forms and stuff?? i was thinking about chromedevtools mcp.?? is it good?

[open saffron]

I am using /think high. xhigh isn't an option on this model.

[odd forum]

Hey! Welcome to OpenClaw.

Yes, browser control works. You have a few options:

Option 1: Built-in browser skill (easiest)

openclaw skills install browser

Then your agent can:

• Navigate pages
• Click buttons
• Fill forms
• Take screenshots

Example:

Go to example.com, fill the login form with username "test" and password "pass", click submit

Option 2: Chrome DevTools Protocol (more control)
Good for complex stuff but harder to set up. You need:

• Chrome running with remote debugging
• The MCP server connected

My suggestion:
Start with the built-in browser skill. It handles most form-filling use cases without extra setup.

When to use Chrome DevTools:

• You need precise control
• The site blocks automation
• You're doing complex multi-tab workflows

What kind of forms are you trying to fill? Simple login forms, or something more complex?

[inner gorge]

Getting status=401 auth error even with valid API key in auth-profiles.json. Running as root on Ubuntu 24.04. Key is correct format. openclaw models status shows api_key=1 but still fails.

[high heron]

then just /new and try again

[open saffron]

using /new didn't help... its starting to loop again and repeating itself. I'm simply giving it something I want it to remember.

[high heron]

/think
usually fixes this issue

idk wts wrong with urs

[hexed terrace]

can you paste the output of openclaw models status?

[open saffron]

It's having a crisis of some sort.

[sweet marlin]

LLM error api_error: An unexpected error occurred while processing the response (request_id: req_011CZS8uiNp4q9WSUU64tj3A)

Been getting this for 7 hours

Any1 else

[mellow loom]

Hey guys, I'm having some issues that are stumping me (and my bot). On Discord voice, bot joins channel but doesn't receive or respond to audio

OpenClaw 2026.3.24, macOS arm64, Node 22.22.1

My bot joins the voice channel fine (autoJoin works, /vc join confirms "Already connected"), and my mic is active in Discord (green border, waveform moving). But the bot never responds to speech. There are zero voice receive/STT/transcription entries in the gateway logs. DiscordVoiceReadyListener fired once on one restart but hasn't appeared since.

All optional voice deps are installed in the openclaw node_modules: @discordjs/voice, opusscript, sodium-native. They all resolve correctly.

Voice config (on the default account):

"voice": {
"enabled": true,
"autoJoin": [{"guildId": "...", "channelId": "..."}],
"daveEncryption": true,
"decryptionFailureTolerance": 24,
"tts": {"provider": "openai", "openai": {"voice": "alloy"}}
}Bot has Connect + Speak permissions. Message Content + Server Members intents enabled. Multi-account Discord setup (15 accounts, voice only on default). Gateway runs via LaunchAgent. What am I missing?

[royal matrix]

Bug: Images/photos fail with Anthropic OAuth (setup-token) auth

Setup:

• OpenClaw on a Linux VPS
• Telegram channel
• Auth: Anthropic OAuth token (sk-ant-oat01-*) from claude setup-token, linked to Claude Pro subscription
• Primary model: anthropic/claude-sonnet-4-6

Problem:
When a user sends a photo via Telegram, OpenClaw receives the file (it lands in ~/.openclaw/media/inbound/), but the agent returns "Something went wrong while processing your request."

Error from logs:

LLM error api_error: Internal server error
model: claude-sonnet-4-6
provider: anthropic
failoverReason: timeout

What we found:

• Photos worked previously when using a different model/auth
• The docs note OAuth tokens have limitations (no prompt caching, no fast-mode tier)
• Anthropic returns an internal server error specifically when trying to process images with an OAuth token
• openclaw models list shows anthropic/claude-sonnet-4-6 as text+image

Question:
Is image/vision support broken or unsupported when using setup-token / OAuth auth (sk-ant-oat01-*)? Is there a config workaround, or does image support require API key billing?

[south ether]

Anyone else noticing their gateway just shutting itself off since the new update?

[stray fjord]

rn

[mellow loom]

Yeah, mine is too. I can't tell if its what i'm trying to do with discord voice or the update.

[solid flare]

Has anyone had any issues with gog, the agent can see events I create but I can't see any the agent is creating

[feral star]

Is anyone using Kimi 2.5 and their openclaw isn't answering?

[real raptor]

anyone has experience let openclaw use adobe product ? like indesign and photoshop

[storm sage]

what is usually the cause when openclaw response is very slow/weird. When i type some messages they just dissapear

Openclaw also writes messages more than once

[rustic dust]

okay, i cant seem to get my agents to talk to eachother. first off in telegram, i can put them (separate agents on entirely separate machines) in a groupchat, they can see my messages, but not eachothers. ok that seems logged. a help thread said a better option was discord. so i get them all in a private discord server, and i turn "require_mention" to false, but they still wont reply to eachother! only messages i send

[inner gorge]

Config : ~/.openclaw/openclaw.json
Agent dir : ~/.openclaw/agents/main/agent
Default : anthropic/claude-sonnet-4-6
Auth store : ~/.openclaw/agents/main/agent/auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (0): -

• anthropic effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=0, token=0, api_key=1) | anthropic:default=sk-ant-a...kG9mzwAA

[hexed terrace]

just for sanity, try a curl request directly to anthropic to confirm the key is actually valid

[covert blade]

A little RTFM and I think I figured out my constant compaction events FWIW I just changed "contextWindow": 49152, and "reserveTokensFloor": 12288 previously reserveTokensFloor exceeded contextWindow. Restarted oc gateway. So far so good.

[ancient jungle]

why the hell is my agent's messages splitting up now, instead of just 1 big message

[indigo goblet]

bruv, whatts up

whats going on, issues?

[winged burrow]

trying to get browser relay working with the Mac app connected to a remote gateway (VPS via Tailscale Serve).

Setup:
• Gateway: Linux VPS, bound to loopback, Tailscale Serve proxying wss://srv1401858.tailc5e976.ts.net → 127.0.0.1:18789 • Mac app v2026.3.24 connected as both operator + node ✅
• Node shows capabilities: browser, canvas, screen ✅
• Browser Enabled: ON, Node Browser Proxy Enabled: ON, Browser Attach-only Mode: OFF, CDP URL: blank
• Chrome is open on the Mac

Problem: Every browser tool call returns UNAVAILABLE: Could not connect to the server.
The extension shows relay unavailable at 127.0.0.1:18792.

Question: What does the Mac app need to start its local browser relay?

Does the Mac node need any specific config to serve browser.proxy to the remote gateway?

Is there a log or status I can check on the Mac side?

[open saffron]

I want to try a different model for my agent. Currently using Kimi but it's doing some terrible looping and repeating itself and not actually running any tools even though it says it is. Anyone have a free recommendation to try?

[left leaf]

hugging face

[covert blade]

I see this in my oc web ui qbove the Message entry window Karen 1:43 PM ↑46.5k ↓761 95% ctx GLM-4.7-Flash:latest
my losslessclaw plugin "contextThreshold": 0.75 theoretically shouldn't ctx 95% be impossible? The loslessclaw should have kicked off?

[solid flare]

So don't update to 2026.3.24?

[undone zinc]

what's the best free or ollama ia to install there?

Claude is so expensive for me, i expend just doing some things like 20$ today, maybe im doing something wrong or using a incorrect model
What's more cheaper than that and good one

[rocky orbit]

You can have codex sub

[undone zinc]

Do i need to pay for?

[rocky orbit]

Yes but it’s a sub so it’s capped to 20$ a month for example

[rotund sigil]

Hey there,
I’m not that tech savvy , but have managed to download openclaw on a new Mac mini, created a new Apple ID I have connnected telegram using bit father which I created a new bot but can get the last bit done been the code required to be able to talk to the bot
Any help would be really appreciated

[rocky orbit]

You need to launch the gateway via terminal I guess

[rotund sigil]

Do I type in launch gateway into the terminal @rocky orbit

[rocky orbit]

OpenClaw gateway

[hybrid kite]

folks, has anyone set up the sandboxing feature for openclaw agents but with podman instead of docker? i'm thinking of just aliasing docker to podman.

[primal ember]

i hear llama maverick is almost as good as opus 4.6? and its free?

[charred steeple]

can someone walk me through this, im new to all of this and im trying to set up openclaw. Dm me please

[chilly wasp]

Same... Have you received any feedback on this?

[mighty rampart]

What OS / distro? There is a podman-docker package in many of them.

[hybrid kite]

I can alias it I was unsure if full compatibility is there

[grizzled merlin]

anybody have a solution for the recent update causing discord to crash the gateway constantly?

[hexed terrace]

openclaw gateway start

[hexed terrace]

it’s good for free, but it’s not in the same tier

[ancient jungle]

anyone notice gpt 5.4 wants confirmation on damn near everything even though the previous reply implies permission and what i want it to do?
minimax 2.7 is a lot more proactive in this regard

[naive hazel]

Hardware spec is:
Processor: 2.4GHz dual-core Intel Core i7
Memory (RAM): 16GB of 1867MHz LPDDR3
Storage: 259GB PCIe-based SSD
Graphics: Intel Iris Graphics 540 1536.

[dusk geode]

Any idea when @sharp parcel will be back?

[south ether]

Found the fix, change your computers power plan settings to ultimate performance if you're running windows!

Also, got open source models running! I was configuring them completely wrong. That's my own bad. Anyone have any model recommendations? I think the most my GPU can handle is 20b and below models

[ancient jungle]

holy shit gpt 5.4 likes to write a book in every response

[echo pine]

That hunter alpha is gone, ya?

[inner gorge]

curl test shows invalid_request_error (not auth error) — so the key IS valid. But openclaw still shows status=401

[inner gorge]

curl test shows invalid_request_error (not auth error) — so the key IS valid. But openclaw still shows status=401

[hybrid laurel]

Hey y’all. Just wanted clarification on 2 things. I’m using webUI mostly to interact with Claw.

• is heartbeat supposed to take over and post to the “main” session/chat? I don’t seem to have a main, isolated session chat. But I do see a “heartbeat” and “subagent: heartbeat” session in the dropdown.

• is there a way to create a separate chat / session that doesn’t get cluttered with other subtasks and can be my “main” conversation with claw?

[inner gorge]

"curl test confirms key is valid (got invalid_request_error, not auth error). Added ANTHROPIC_API_KEY to systemd service environment but still getting 401. auth-profiles.json has correct format with api_key type."

[ancient jungle]

i guess it can be either depending on how you set it up

[hybrid laurel]

I guess it’s my fault for not explicitly telling it to maintain a separate session and it just assumed to use the main one? That’s a lack of my understanding at the time.

Do you know anything about manually generating new sessions/chats? because from what I understand /new and /reset maintains the same session id just compacts some context

[ancient jungle]

so you want your heartbeats to happen in your chat or a separate session that you don't see?

[hybrid laurel]

Yeah-but I already figured that part out, just telling it to do exactly that.

The question is now how do I create an entirely new chat

[nimble hill]

/new

[hybrid laurel]

That doesn’t create a new chat session, it still uses the same session id of the session I’m in

[nimble hill]

It is a fresh session, all your context is reset..isn't what you wanted?

[hybrid laurel]

More specifically, in the webui dropdown list of sessions, I wanted to generate new options so I could work on multiple tasks in parallel, across different chats, not one main chat

[ancient jungle]

so you want additional agents?

[tawdry ibex]

anyone know how to get a predic ai for kalshi

[sly lake]

HRESULT 80131522 这个怎么解决

[placid flume]

Does anyone have advice on which model and configuration offer the best value in the lowest price range without becoming too slow or unstable? I’ve tried openRouter and llama:local but haven't had much luck. Any guidance would be greatly appreciated 🙏🏻

[copper nova]

if you want frontier models, a pro sub or github copilot would be your best value

if you just want cheap then look at the chinese open weight models like kimi

I think I seen alot of people shilling minimax and the alibaba plan too

[placid flume]

Thank you so much!! 🙏

I used to use Healer Alpha when it was free, and I was happy.

[copper nova]

isn't that the xiaomi one? they probably cheap too

I think it's mm something now

[nimble hill]

you cannot do that from chat - but using cli you can initate a new session that will be shown in dropdown. not near the system rightnow - will give you in a while

[mighty rampart]

What do you have session.dmScope set to? This seems related: https://docs.openclaw.ai/concepts/session

[nimble hill]

not exactly - you can have mutliple sessions for the same agent - just like you can have multiple channels for the same agent

[hybrid laurel]

^manjax is right - without creating a new agent, I wanted multiple sessions.

[ancient jungle]

🤷‍♂️ but why?

[hybrid laurel]

if i had an agent for every new thought... well it would be unbearable. overkill.

[fickle dock]

Hi - a little new to Discord so I might be using search completely wrong so I'm just going to ask in this channel

But I've been trying to get my team to use the OpenClaw instance I have on my Mac Mini and I'm running into a bunch of issues with queueing, etc. AFAICT there aren't very easy ways to see what's actually in-flight vs. queued in the command queue between the session lanes and the global lane?

I have tried using the existing cli and clobbering other things to put something like this together tailing JSONL session logs but it seems like there is a bigger change that needs to happen for this to be a reality?

My ideal would be sth that can let you see how many slots you have available, how many are taken up, and if it's an active session the last 3 messages from the agent with the timestamps, etc.

[nimble hill]

This is a command I used to do in some earleir version openclaw agent --session-id 1234 --message "who are you?" it does give a result, but somehow a new session does not show up in UI. It should create a session when there is no existing one.

There is a roundabout way - by setting up hook. If you have hook configured, you can do this
curl -X POST https://<your_server_ip:18789>/hooks/agent -H 'Authorization: Bearer <yourtoken_set_in hook_config>' -H 'Content-Type: application/json' -d '{ "message": "which model are you using", "name": "test", "agentId": "main"}'
it will give the response as {"ok":true,"runId":"6727c0f5-0a47-4b42-928a-b0a066d0791c"} Now you can see the new session in the dropdown

Hook configuration

  "hooks": {
    "enabled": true,
    "path": "/hooks",
    "token": "12345678",
    "allowRequestSessionKey": true,
    "allowedAgentIds": [
      "main"
    ]
  },

[young tusk]

Anyone had their OpenClaw agent respond with an end-of-sequence token in the chat? Tools appear to execute but nothing actually happens. Agent narrates actions instead of performing them.

[hybrid laurel]

ok that was super helpful - i ran the same cli command, it responded with tthe generic open claw message ready here to help. I ran a second command with same ID asking what was my last message and got same verbatime message saying my last message was "who are you" - and yet it doesn't show in web UI.

[undone zinc]

who dont know how to install openclaw correctly i reccomend for newbie people install ollama + claude and open ollama launch claude and tell them that you want to install openclaw it will do all for you for free 😉

no frustration with that way just a couple of hour seeing how its work and accepting haha

Enjoy mah bros

[hybrid laurel]

at least I can make a custom implementation from this - I thought i was going crazy not seeing anything in the docs that explicitely allowed me to create new sessions. I'l create a custom script that adds the session to the webui - but im hoping its not just CLI bound state memory and it can be fetched outside of that.

[nimble hill]

I updated my original message, if you want to use the hook feature - this will create a session in the UI and then you can continously use it

[hybrid laurel]

i see that now! super. i seriously appreciate your help

[nimble hill]

welcome - lastly, if you want to have the message inserted in to the same main agent - you can use openclaw agent --agent main --message "who are you?" It will be shown in your main. This way, even if you do not have access to chat - or via a program you can invoke it

[hybrid laurel]

beautiful

[nimble hill]

the problem with this approach is that if you are installing in your personal laptop, you have no idea where the files are. Openclaw versions changes lot of things and your fs would be drifted quite soon. Hence, I suggest to take some pain to read the docs and install with appropriate security consideration. OR install in a VPS and experiment. Those who are techy MUST install in a docker (or even k8s) and play around

[thorny sequoia]

please add me ([e-mail]) as a test user in the "Superfly" Google Cloud project.

[summer hemlock]

bro I kept it in a container for like a week

then I just gave it the whole mac mini. It runs it's own dockers now.

[tardy schooner]

Has anyone here been able to get their Zapier MCP server connected to openclaw to read calendar and emails, etc? In my setup, OpenClaw can see it, but is missing the exact OpenClaw runtime path that makes the remote Zapier MCP tools callable by the agent.

[hybrid laurel]

yeah, if thats a fresh system, you're probably good @summer hemlock the issues arise for newbies running this on their main computer. Look at the recent LiteLLM hack. one simple update behind the scenes and your entire computer is PWNED

1 simple python script pulled all local ENVs and other db type files.

[summer hemlock]

I've been here for months now, and the thing I've said most is "don't run this on your own local computer"

if you're like experimenting with this, you can def run it on docker desktop inside a container locally

[hybrid laurel]

i hate shilling a VPS - but how safe is a docker container, or even WSL?

[summer hemlock]

but you probably don't want to have your desktop running 24/7

[hybrid laurel]

as far as I understand, there are still ways to exit a docker container * from within a script to access the underlying system

[summer hemlock]

well you still have to think about security. You should keep your API keys & info in a local .env file, and mount that to the container when it's up.

like don't put that stuff in the setup files or dockerfile 😄

but docker is pretty great

if you want you can keep your openclaw on a private git, with version control etc. Just start docker containers straight from your git with docker desktop app. It has the integration.

but imo getting some hardware that can just run 24/7 and spin it's own docker containers is even better

[hybrid laurel]

@summer hemlock has docker given you any performance issues?

[summer hemlock]

lol no. I have a video on my yt channel where I have antigravity boot up like 16 containers calculating prime numbers at once to test my PC.

[tall rapids]

i run in containers (k8s, but same same). no performance issues whatsoever

[summer hemlock]

https://www.youtube.com/watch?v=phESFJ-CfN0

yes, I had nano banana make a stupid openmouth pointing thumbnail

[hybrid laurel]

for sure. that makes sense. I used to use docker for full-stack apps ~2019 mac pre m-chips on intel and after a certain point my comp was just unusable. but claw is light as hell.

[stoic dome]

Can someone advise if I have enough hardware to run something like qwen locally

I have openclaw on a rpi4 but I have a NAS running truenas (12700k + 64gb ram) to throw toward an LLM. Is that any good or would performance be bad

[ancient jungle]

don't do it

[hybrid laurel]

which gpu?

[summer hemlock]

ofc I sent it a picard meme to start

[stoic dome]

That's my problem don't have one rn

I also have potentially a M1 MacBook Pro 16gb if there's some way to use that and keep the Mac os untouched as is

[hybrid laurel]

I wouldn't expect anything crazy like flagship cloud models. whats youre usecase?

[tall rapids]

you're not gonna have a good time running an llm locally without a gpu. I mean you won't have a particularly good with a gpu either but without, nah

[summer hemlock]

@tall rapids before openclaw was a thing, I've done some stupid things with orchestrators and sub agents in containers & subcontainers, running ralph loops

containers & subcontainers are so confusing for agents lol

gotta mount a .md file in every container to let them know where they are

[stoic dome]

Help me analyse things for a business check ad accounts, general tech questions etc.

[summer hemlock]

M1 won't do man. fork out 20 a month for a plan

you can use the M1 to run the agent, but not the model

[stoic dome]

Wait it's that cheap? How

[hybrid laurel]

you MIGHT get half-decent responses on a cpu. your best bet (imo) is subscribing to openai/claude pro. you get decent context - but you have to watch out for rate limits.

[stoic dome]

How much does 20 a month get me?

[summer hemlock]

you can use open AI plus with oAuth.

that's 20 a month plus tax

[tall rapids]

i've never tried the docker image in plain docker/podman but in my locked down openshift, it was unhappy when it was born. no tools, no exec on a lot of things, trying to run "openclaw" instead of "node openclaw.mjs"

[summer hemlock]

@stoic dome gpt 5.4 is pretty big.

[tall rapids]

one of the first things i did was create a sandbox pod (clawdbox) chock full of the tools an llm could ever want

[stoic dome]

Might talk to wife about that, can she use chatgpt account on the pro version as well as the openclaw instance

[summer hemlock]

my openclaw on the mac mini has a searxng container running on it's own local docker desktop app to do web scraping & indexed searches for free.

[summer hemlock]

yes ofc you can run the chat and the auth at the same time

and also openAI codex

[stoic dome]

Yo damn this might be a banger. I do have it installed on a rpi4 and was gonna offload the model locally but 20 a month + tax and she can also get more leverage from it might be the way to go

At least trial it for a month

[summer hemlock]

you just need the openAI plus plan. Then redo the openclaw onboard in your terminal, choose the gpt models you want and do oAuth login

[stoic dome]

Just the openai plus plan?

[summer hemlock]

it'll bring you to openAI login page and copy the login token

yes

maybe this won't last

but atm

openAI is the only company that allows oAuth for openclaw. Peter Steinberger, openclaw creator, is working with openAI.

not saying this won't change in the future btw

[tall rapids]

yeah it would not be good optics if they rugged that

[summer hemlock]

openAI might want to make their own "closed" claw 🙁

[stoic dome]

Ahhhh well one month I can trial it.

Does it take a long time to rate limit on that open ai plus plan?

[summer hemlock]

sama won't care lol. That's why I pay per MONTH, never per year

[tall rapids]

though tbh i use my claude max x20 plan with setup-token and I've not had an issue either. TOS be damned

[summer hemlock]

google gave ppl 50% off to get a yearly plan. I said no.

this is NOT the time to be loyal to model providers

[tall rapids]

that's the one I have. expires in February 2027. seemed like a relative "bargain" at discount

[summer hemlock]

try other models every month

don't be loyal because things change too fast

[stoic dome]

Assuming you can hammer open ai plus without any limits yeah? Or it would take a bit of time before that would kick in?

[tall rapids]

same limits as if you use gpt desktop or cli

[stoic dome]

Cool I'll give it a crack for a month I think

See how it goes

[tall rapids]

openai's limits are pretty generous tbh

[summer hemlock]

tbh it IS. you get 2 tb cloud AND the AI stuff. So you actually only paying 10 a month for the gemini AI and antigravity. and all the other pro stuff in apps like notebookLM

Amazing value.

But openclaw removed the oauth....

google gives most value, imo. But antigravity oauth doesn't work

[tall rapids]

still works if you craft the json manually though

[summer hemlock]

unless you make it yourself

and well then you'd have to update etc

[summer hemlock]

yes I did that last week lol

[chilly kayak]

ahhh i messed with config and broke it

[summer hemlock]

that's how I ran claw last week

had exactly same idea

[nimble hill]

oracle cloud give 4 vcp, 24 GB RAM and 200 GB disk competely free forever - make use of it - you do need to sign up with CC

[summer hemlock]

I just brought oauth back using opus 4.6 and openclaw version control in git. it took 13 mins and 96% of my tokens.

[tall rapids]

if you can get it.. i cannot for the life of me get it provisioned in this region (APAC)

[summer hemlock]

opus 4.6 sure loves those tokens

nomnomnomnom

[tall rapids]

yep. I had deleted google, went to add it back, noticed the onboarding path was gone and did the same. pulled the config out of my gitea repo

[nimble hill]

raise a support ticket - they are usually considerate when you write

[summer hemlock]

I can still change my mind about google

but atm gpt 5.4 is doing good things for me

[tall rapids]

that's why i main sonnet. good enough 95% of the time. crank up opus for the serious stuff

[summer hemlock]

it helped me out a lot today with a difficult issue at work. Blown out vfd & DC motor for a big extruder

[summer hemlock]

I run antigravity on the mac mini that runs my claw

whenever I break claw, I can just remote into the mac and ask opus 4.6 to fix it

and it'll just....go.
it's beautiful to watch tbh.

if you give it a whole system

[nimble hill]

thats one way to lose the tokens 🙂 if you are already in docker, why not take automatic period snapshot and start from last know good version.

[summer hemlock]

@tall rapids explosion blew 4 filter & fan holders out of the cabinet.

[summer hemlock]

yes if you want to make memory permanent when using containers or any virtualisation you have to deal with taht

you can get free online database btw

like supabase

you could store your memory on supabase so it doesn't go away if you restart container

mount it back in

[tall rapids]

yikes

[nimble hill]

for the entire setup in docker, all you need is a bind mount of .openclaw and most of things will be there even when you restart. if you are on zfs or other systems, take periodic snapshot of that bind mount - thats it. regarding database - why online - use locally host postgres.

it is these influencers who have promoted the mac minis, supabase and other onlne tools because they get kickbacks

[summer hemlock]

I'm not an influencer getting kickbacks. It was opus 4.5 suggesting supabase to me when I started using AI back in december.

opus 4.5 defaulted to supabase and vercell, and some ways of coding and UI back then.

[tall rapids]

someone is paying influencers to get people to use free-tier services that cost the provider money? wow. what big brains they have

[summer hemlock]

soit was all "'glassmorphism"

and everyone was making AI apps and sites that looked alike

that's how tailscale got killed as well

their monetarisation was people buying their pro plan

they grew 500% but it was all AI agents that never get their pro plan

[nimble hill]

you don't understand - the free tier is often limited, when you put you data in you effectively get locked and they are counting on most of them being non-techy or lazy not to move and hence go to next level to charge. They are with big brains and ecnomical calcualtions

[summer hemlock]

no that was windsurf lol

plans are sponsored

I couldnt wouldn't run claw otherwise

[nimble hill]

anyway - if you are hosting on your own hardware - learn k8s - you can host mutliple things and scale up with decent security. Most of the services have opensource version. So you still need to take some online subsription by choice not by force

[tall rapids]

counting on people's laziness. like anything else, you can go build it yourself or you can pay someone else who already did. there's nothing supabase or tailscale, or pretty much any of these providers do that you can't just go build. if you pay, it's convenience costs only

[tall rapids]

everything I have is on openshift (well, okd upstream but same same)

[summer hemlock]

I didn't know what you meant, then my agent said it just was kubernetes

[tall rapids]

yeah, redhat's opinionated k8s basically. that reply wasn't to me. oops

[summer hemlock]

well I've just been using containers tbh

seems more than enough atm

it's confusing enough as it is 😄

[tall rapids]

honestly i think k8s makes containers easier, once you get over the initial learning hump

[stoic dome]

also if I run it on a machine locally will it have the ability to create dashboards etc that I can locally host on my network with data I get it to fetch and scrape? I haven't really used openclaw it is potential yet so I feel a package will help me unlock some serious horsepower.

[summer hemlock]

like, AI agents not understanding if they are inside or outside containers / subcontainers has been a serious issue for me the past months.

[tall rapids]

takes all the legwork out of orchestration

[summer hemlock]

bro if you host it locally it will just run on your machine localhost:18789

[tall rapids]

yep, if i had a penny for the number of times my claw tried to do things at first that could not possibly work in a container, I'd have quite a few pennies

[summer hemlock]

btw DONT run this on your local machine. It's not secure

[nimble hill]

OKD is nice - but people have been moving to "lighter" k8s - microk8s, k3s - wherere there is much flexbility

[tall rapids]

i was on k3s on proxmox lxcs before. much lighter, much easier. moved for work related reasons

[summer hemlock]

one time, I was trying to change some code for a website I was making in my dev container. I was also running agent outside the container. At a certain point I wanted to mount a file from the outside to the dev container. The outside gemini agent said some BS that he was gonna restart the container.

I allcapsed "NO DONT KILL THE CONTAINER!!!!".

Gemini said "na we good, because I see that we're local" and just reset the container. -_-

[worn echo]

Anyone experience some problems with tool calling?

[tall rapids]

okd is indeed heavy as hell

[tall rapids]

that's our gemini lol. an absolute madman

[summer hemlock]

gemini is a MFer when you code with it, as a human. but gemini flash is pretty great as an openclaw agent, in my experience

[nimble hill]

yeah security needs to be hardened- does not come default. k3s by default runs flannel as a CNI 🙂 microk8s does not have rbac enabled by default. So yes, some work is also needed in these light weight tools to be actually ready for decent job

[summer hemlock]

I called it a "wild horse" so many times

it's just so stupid when you use it yourself in an IDE. I don't even understand how it can be so capable as a claw agent

[tall rapids]

agree but it wasn't even that. it's just a homelab. I just needed to learn openshift and, as I'm sure you'll agree, the best way to learn anything is to force yourself to actually use it daily

[summer hemlock]

google should be able to fix this lol

[tall rapids]

it's a smart model but utterly reckless. I ended up writing a detailed SOP/Skill to stop it messing up my openshift

[nimble hill]

yep - now I run multiple microk8s, k3s versions across different clouds - all connected via tailscale at infra levelr and service mesh linkerd. There is this single openclaw gateway - which controls other clusters and machines as "nodes". This has been working great - a single point of control for multiple clusters and machines

[summer hemlock]

it's worse than reckless bro. It lies, it cheats test results when you ask it to test websites

when you say "hey you cheated"

it'll be like "hahahah yes i'm so sorry never will do it again".

then it'll cheat AGAIN.

I hate coding with gemini

it's like

not even comparable to opus or gpt

[nimble hill]

there will be a time when "morals" will be another skill 🙂

[summer hemlock]

but it's cheap & good for openclaw. DOn't ask me how

[nimble hill]

This remote browser control was working very well with .23 - this .24 did somethings and it hardly works consistently - epsecially when I am running a docker container at the remote nodes. Still trying to see what is the problem and get it stable. the non-announced breaking changes in openclaw...my goodness!

[random summit]

I installed open claw Win 11, with searxng, power shell, lm studio, with and without docker trying to get it to work. It installs and Im trying to grant it permissions, it is using my local Ai without issue, but it cant do anything at all its seriously useless LOL, it wont install plugins for me, and it tells me acp isnt installed. but when I run /acp doctor it says everything is fine. my system says its fine. but it dont have permissions for some reason, its driving me crazy. If anyone would like to help me, please DM me. Thank you in advanced.

[stoic dome]

Sheeesh I need to set this up

[ancient jungle]

this the kind of bullshit that's having me refrain from upgrading from 3.13. i finally got this thing to a usable state and all i'm seeing is people breaking stuff by updating

[stoic dome]

So the tldr is pay for openai plus and go through the setup and it's openai for the model @summer hemlock ?

[tall rapids]

openai-codex for provider

[nimble hill]

there is hardly any quality control at openclaw - AI can only do so much regression testing - it needs to be hand holded

[tall rapids]

i check like 4-5 different things after every upgrade. at least one is usually broken. most often, the browser stuff

[nimble hill]

.24 breaks the very basic thing - like they removed the timeout variable for the cli and the default value is 800ms - which is never enough. So basic command like openclaw devices list does not work without specificying the --timeout 20000 value. This was earlier possible to be set at the gateway config as env vars VITEST=1 and OPENCLAW_TEST_HANDSHAKE_TIMEOUT_MS. Now if has no effect - there is a PR raised for the fix - hopefully it gets merged soon

[tall rapids]

canvas behind a proxy has been broken for the last 6-7 or more releases. run behind ingress? nah fam, no canvas for you

[nimble hill]

another problem with no sync between openclaw nodes list and openclaw nodes status the status will show the node, while list says 0 🙂

[tall rapids]

I've seen that one, yep. at least a few issues open for it too

[nimble hill]

there was one PR that was opened and fixed - but it was flagged because of more than 10 open PR and got cancelled 🙂 otherwise would have been there in .24

[tall rapids]

it's a frustrating project at times but fun to play with. I wouldn't try to ship anything i relied on on it at the moment

[nimble hill]

yeah - my professional workflows are still happy on n8n with deterministic flows. Hwoever, some personal flows have been calling openclaw to save on tokens - since n8n does not have OAuth and uses openclaw

[tall rapids]

tbh nothing I'm doing with it couldn't be done somewhere else. a bot that uses a rag kb and some other knowledge to build flows in one of our products (but again, just as a lab experiment). a webhook that receives my openshift alerts, actions them if possible (there's a list of thing he can touch and not touch) and sends me a telegram in a clone of scarlett johannsons voice i made in fish audio lol

just toys for learning

[fickle dove]

just setup ollama and openclaw, when I run openclaw via ollama, I get no response via the tui or the openclaw web portal

e.g. I do a basic /set nothink and stuck on running