#users-helping-users

1 messages · Page 54 of 1

spare shard
#

what do you mean?

#

what do you mean by "everytime" like randomly?

dreamy pendant
#

Yeah like my openclaw sends me that

spare shard
#

It would happen every hour if i remember

#

Tool is to see which file
Reasoning its in the damn name
It uses its console to do some console things

#

yknow

#

running commands

#

thats all

keen trellis
#

pretty hard

#

its doable but you are wasting time and your ai usage on fixing your UI because openclaw had an update almost every time

#

if i didnt have telegram channel set, i would cry way too often. update, UI broken, telegeram, hey UI doesnt work, fix it

#

now it would be probably easier actually, since most features i had to invent is already in control ui

#

the thing i miss the most, is canvas rendering and such, so i can see in real time what doe sit do, with Ui for my project etc

spare shard
#

Can i just make an UI then connect to another localhost THEN just connect the CHAT To the ui?

keen trellis
#

the way it works you pretty much create another gateway with own ui

spare shard
keen trellis
#

unless i did it wrong way

keen trellis
spare shard
keen trellis
#

since i have whatsapp channel with bots own number added, it was cool to have contacts features, where, once you added the contact, you could just say, whatsapp my Mum with this recipe or some other stuff

#

it would know, that my mum speaks polish only, is offline at night etc, so it would be like, ok i scheduled message for 9am in polish and give me message content

spare shard
#

IM out of tokens..

#

damn it

keen trellis
#

yeah, happens to me often in Antigravity

#

openclaw is my deep research and some coding fella, so far managing it within weekly openai quota

spare shard
keen trellis
#

same

spare shard
#

I would have usen Kimi maybe

keen trellis
#

and im never out of tokens

spare shard
spare shard
keen trellis
#

cos i only use openai model for heavy lifting, all little things like cron jobs for heartbeat and sanity and even scheduled joke message to my mum about how old she is, are dealt by free gemini 3.1 flash

spare shard
#

I tried gemini

#

but it kept giving me RPM errors

keen trellis
#

you dont understand i think

#

i dont use gemini 3 flash for anything serious

spare shard
#

lol

keen trellis
#

only silly things, that are pointless to be run by primary model like gpt 5.4

spare shard
#

got it now

keen trellis
#

i dont need to use my gpt 5.4 quota to tell me heartbeat is ok, or how is my system (ssnaity check)

#

sanity*

spare shard
#

Idk why but I hope they get Groq support soon yknow its free and i can use it for silly things like you said

keen trellis
#

or to send my mum a joke

#

etc

spare shard
#

but ill do the same as you did

#

by adding gemini 3.1 flash for silly things

keen trellis
#

get this, before i did the split, what model does what. i was going to sleep, waking up and my codex 5hr usage quota was at 27% used and i didnt even touch the keyboard all night

spare shard
#

DAMN

#

does heartbeat md and system check usages that much

#

wow

keen trellis
#

thats why you run out of usage so fast, trust me

spare shard
#

dang.

keen trellis
#

yes, i also disabled skills and memory context for heartbeat and sanity

#

thats a token eater

spare shard
#

I need skills lol

#

Hearbreat :>

keen trellis
#

you can just tell you bot to set it up, it should be done with 2 or 3 prompts properly

spare shard
keen trellis
#

he can run tests to see if hearbeat/sanity is run by free gemini model etc

#

tell it to disable context from memroy and skills from those cron jobs

#

saves a lot of tokens, i used to have out of tokens cos of this

spare shard
#

Wow

keen trellis
#

256k context window and bot was sending 300k for heartbeat 😄

spare shard
#

Uhh i think my agent is crashing out it suddenly started to do tool call output even when i try to talk to it it will just continue

#

lol

#

no it just the ui

#

freaked out

#

oh

keen trellis
spare shard
#

Ooh

#

def doing that.

vital bough
#

anyone here successfully set up a verified openclaw matrix bot with e2ee?

#

and invited that bot to a space with multiple rooms?

spare shard
#

isnt hermes a clone of openclaw? do i know wrong?

#

idk bout hermes lol

#

LOL

vital bough
spare shard
#

i dont know really that mucj

lavish lance
#

Did they change something recently so ollama cloud models are blocked somehow? E.g. kimi.2.6??

normal garnet
#

where do I submit bug request ?

daring moon
#

Has anyone ever just done a "full reset" of their Openclaw? Uninstalling, installing from scratch using pre-existing skills and cron jobs and memory etc? Curious how this went if you did.

I feel like my Openclaw just gets crustier and more broken with every update and model change. Lots of leaking context and weird behavior in Discord and in the TUI. Feels like Windows after 3 years of regular usage and I just need a reinstall.

lone verge
#

ugn why is nextcloud-talk such a pain

mighty rampart
slim lotus
#

Hi. Need help with following. Appreciate any pointers
Termux node-host on Android — system.run not exposed

I've set up OpenClaw on an Android device via Termux as a headless node-host. Node connects and pairs fine (paired · connected, caps show browser, system). However:

• nodes describe shows Commands: (none reported)
• exec host=node fails with: "exec host=node requires a node that supports system.run"
• nodes invoke --command browser.navigate fails with: "node command not allowed: browser.navigate is not in the allowlist for platform android"
• exec-approvals.json is configured correctly (security: full, ask: off) but the socket never starts

Looks like the gateway has a platform allowlist that blocks all commands for platform: "android" on the headless node-host path. Is this intentional? Is there a way to enable system.run on Android/Termux, or is there a workaround?

Running: OC 2026.4.23 on Termux (Android), Gateway 2026.4.22 on Linux.

daring moon
# mighty rampart It hasn't even existed for six months, so... no reset yet here. But I use Claud...

I find that Openclaw tends to be poor at self-introspection - when I ask it to fix things about itself, it often seems to fail. And I find the online instructions dense and difficult to parse. Maybe I need to grab source and that would help it understand? I like the idea of using Claude Code to try and fix it (I definitely have more problems since moving from Claude to Codex models) so I'll give that a shot. Thanks!

mighty rampart
unborn igloo
lone verge
#

ok dang I got nextcloud-talk to work after many hours

#

buut I don't know if I'm just stupid but I can't actually mention my claw in NC

mighty rampart
compact beacon
#

@sharp parcel Is it broken or no longer available?

smoky chasm
#

#1497621480656076890
hi, i would like to know if anyone also uses github copilot knows what's happening...

leaden heron
unborn igloo
daring moon
past hawk
#

Hello everyone,

I was wondering if someone has made some progress in using openclaw in a usefull way. Yet for me its talking a lot but not delivering anything. I use the latest version together with ollama and cloud models, currently kimi-k2.6:cloud.

What a shitshow! ...

I was wondering if there is a setup and a manual somewhere shared witch really works. Tell me pls. about it ... and let me know where to find the manual.

thx

flint flame
#

Hy guys some ona have this problem Memory index failed (main): memory embeddings batch timed out after 120s

past hawk
flint flame
steep compass
#

Does anyone know how to get my agent to stop spamming me with exec commands on telegram with every prompt?

past hawk
brazen lagoon
#

does anyone know how i can decrease my input tokens and cache tokens

#

me doing simple uses cases is setting me back like 5Mil tokens a day

#

for input

steep compass
#

tried a gateway restart too

brazen lagoon
#

does anyone know how i can decrease my input tokens and cache tokens
me doing simple uses cases is setting me back like 5Mil tokens a day
for input

steep compass
#

typically looks like this:

#

Working…
• exec run openclaw gateway, openclaw gateway restart
• command run openclaw gateway, openclaw gateway restart
• tool: exec

#

Then removes it for the response

celest root
#

Hey,
my custom skill isn't triggering even with the most obvious prompt referencing the skill''s description. I assume it's not triggering because it doesn't follow any instructions that the skill defines, though I don't know how to check for certain if a skill has triggered or not.
Has anyone experienced something similar or can help?

mighty rampart
brazen lagoon
#

and it still does like 300-600k in tokens every session

#

evne though i reset the session frequently

past hawk
# steep compass tried a gateway restart too

Seems to me like a lots of stuff is not working right now! I just posted before and wonder if anyone have positive, like useful reproducible results.

I get the impression slowly that OpenClaw is by purpose havily broken.

steep compass
mighty rampart
brazen lagoon
brazen lagoon
mighty rampart
brazen lagoon
steep compass
past hawk
mighty rampart
mighty rampart
past hawk
#

Thank Yuo

brazen lagoon
past hawk
#

@mighty rampart I use currently this models.

ollama/kimi-k2.6:cloud
Fallbacks vor coding and or other subagents: gemma4:31b, qwen3.5:397b

The biggest problem is that they do a lot of guessing and don’t even know their own tools properly. As far as I know, that should happen by default when I install OpenClaw in the standard way, right? 😄

Its the worst when I just have reset the session. It don't know anything anymore. Are the standard md's delivered with openclaw so bad?

I should have listened right away to you. I'm about to mount my workspace folder from my server, I hoped that if the agent says he did something for example config changes that he would really do, even after several iterations on trying in many ways did not help to get openclaw work properly.

(for example I saw just a couple of minutes ago that I had discussed a change with it) but when I checked on the server and within the webinterface the changes where not applied 😕 ...

Thank You

wraith seal
#

BAD

#

very bad

patent cobalt
#

Does Anybody know why when using local models such as qwen2.5 it doesn't use the .md files (AGENTS, IDENTITY, SOUL, etc) as much as using, for example, google gemini flash?

mighty rampart
mighty rampart
# past hawk <@478926775666212883> I use currently this models. ollama/kimi-k2.6:cloud Fallb...

Are the standard md's delivered with openclaw so bad?
The defaults are minimal, those files are templates, you're meant to talk to your agent and have it fill them in. Whether this actually works depends on what model you are using. It turns out that you need a pretty capable model for it to do the multi-turn conversation and reasoning necessary to synthesize that information and update its own configuration.

patent cobalt
#

3B

#

which one you recomend por rtx3050 8vram?

#

if there is one, if not i'll see other options

mighty rampart
# patent cobalt 3B

I think a 3B model is going to struggle with agentic behavior and tool calling. Is Krill alive over in #1459642797895319552 ? Perhaps it can give advice. Otherwise web-based Google Gemini is pretty good at helping with OpenClaw stuff. Ask it what model for that card and OpenClaw.

patent cobalt
#

Thanks a lot

flint flame
#

Some ona have a solution for this problem Memory index failed (main): memory embeddings batch timed out after 120s this error occur when I use openclaw memory index --force

spare shard
chrome night
#

"status on that little project you were working on? You seem stalled.

I got most of the way through it, then the gateway restart cut the command mid-flight. I’m checking the live state now so I can tell you exactly what landed and what still needs verification."
FYI: part of the "little project" 5 step plan was at step 3 - to restart gateway.

mighty rampart
wraith seal
#

it could literally rm -rf your computer

spare shard
#

a bit late

#

for that

#

because IT Deleted itself

#

🤤

wraith seal
#

what did you do

#

for it to do that

spare shard
#

I dont know

#

lol

#

it kinda freaked out

#

like it started doing tool call and wouldnt responf

#

then poof

#

the openclaw was gon

#

and i dont have any devices to put but i can put openclaw into a safe env

wraith seal
spare shard
#

good thinknig

#

thanks

marsh bolt
#

Guys - can't get the gateway to start since upgrading to 4.24. Getting stuck trying to update plugin dependences - doctor asks if I want to, but doesn't seem to do it and then gateway can't start because of missing dependencies

jagged frost
#

First time posting here. I have a fresh openclaw 2026.4.2 installation on linux mint. I have no previous experience with openclaw. I am using the internal (bundled) browser plugin in openclaw and experiencing that certain sites like edition.cnn.com are so slow to load that they timeout. In such cases I'd like the agent to load a lighter version of the site if available. Where can I give such standing instructions to the agent - in the config file or somewhere else?

flint flame
flint flame
mighty rampart
slate minnow
#

anybody been running openclaw with qwen3.6:27b locally?

slow oxide
#

Well I suppose I'll ask here as well. I coded my own plugin but it wont show up as available. According to the documentation it seems like it should just appear and the doctor command sees it as imported but I have no idea how to make it go from imported to Loaded

flint flame
#

If some one want to use Ollama as local embedding I don't recommend to use it. After to update the new release launched today I can use local instead ollama and it's easier to use than ollama. I try debuggin to see what really happend but I miss with it. So if you wanna use embedding for local I'll recomend use local. You should install node-llama-cpp and the configuration I used is this "memorySearch": { "enabled": true, "provider": "local", "fallback": "none", "extraPaths": [ "insert/your/paths" ], "local": { "modelPath": "/dir/path/model I use qwen3-quant-guff_q8" }, "cache": { "enabled": true } },

naive ocean
#

anyone having issues after updating to .24 ? openclaw status is showing a bunch of plugins staging and installing bundles but oterh then that the gateway never loads via luanchctl

oak agate
#

its probably known, waiting for next patch

pallid cove
#

I updated to .4.24 and any openclaw CLI command keeps trying to install plugins?

23:43:27 [plugins] acpx staging bundled runtime deps (1 missing, 66 install specs): acpx@0.5.3
23:43:31 [plugins] acpx installed bundled runtime deps in 4122ms: acpx@0.5.3
and so on for all
but they never get installed apparently

#

what's going on

open ravine
pallid cove
open ravine
rose halo
#

Guys what is the plugin / skill for free local audio transcription for whatsapp in oclaw

#

And how do I install it

#

It's not in "openclaw configure" menu for plugins or skills?

woeful adder
#

Hi. Any LLM combination for using openclaw free? groq/llama-3.1-8b-instant and groq/llama-3.3-70b-versatile were never able to asnwer due to token restrictions. gemini-2.5-flash only answered once.

prime reef
lucid turret
#

hi
how to setup codex oauth on headless?
docs says

For headless or callback-hostile setups, add --device-code to sign in with a ChatGPT device-code flow instead of the localhost browser callback:
but this option dont exist and onboarding dont show code [shown on the local device only]

shut onyx
#

Can anybody tell me what’s the best token plan available for openclaw , i need kimi 2.6 and deepseek 4 pro

meager vessel
#

this new update broke my openclaw!!

tough siren
naive ocean
#

I found how to manually solve the issue but its a pain, openclaw sswitched plugins to node_modules/openclaw/dist/extensions/
i had to manually go into each of the mentioned plugin folders and run the npn install xxx that the doctor mentioned, something to do with the latest version of node js. and I dont know what else

#

but its prob more sane to go back to 23 and wait till its patched

#

here is an example
cd /usr/local/lib/node_modules/openclaw/dist/extensions/codes
npm list --depth=0
@openclaw/codex@2026.4.20 /usr/local/lib/node_modules/openclaw/dist/extensions/codex
├── UNMET DEPENDENCY @mariozechner/pi-coding-agent@0.70.2
├── UNMET DEPENDENCY ajv@^8.18.0
├── UNMET DEPENDENCY ws@^8.20.0
└── UNMET DEPENDENCY zod@^4.3.6
npm install @mariozechner/pi-coding-agent@0.70.2 ajv@^8.18.0 ws@^8.20.0 zod@^4.3.6
npm list --depth=0
@openclaw/codex@2026.4.20 /usr/local/lib/node_modules/openclaw/dist/extensions/codex
├── @mariozechner/pi-coding-agent@0.70.2
├── ajv@8.20.0
├── ws@8.20.0
└── zod@4.3.6

meager vessel
#

trying to fix the openclaw now 🙁

olive cliff
#

First, have them pay their API bill before even discussing bringing anything in house.
IF they bring it in house, who's going to manage it? Do they already have their own server to run it on?
IF you do decide to allow them to bring it in house(I dont think you should) it'd be smart to charge them the total costs of the hardware+ 15%, let them know they foot their own API bill, and you're "hands-off" unless they pay you a decent monthly "service/maintenance fee". Charge them a "fair" rate for the hours you already spent on this. You're already being too generous with them not paying the full API bill.
TLDR: Cost of hardware + 15%, + Service/maintenance fees per month, + hours worked on building this @ $25/hr(MINIMUM).

naive ocean
#

@meager vessel not sure my suggestion worked, I am still testing and the solution maybe to install. .23 again ...

ancient falcon
#

Hi does anyone use OpenClaw to manage the discord server? when i tried to setup, and ask it to setup channels and stuff for me, it says it cannot do and asks me to setup the required channels manually, is this how it is working for others too?

novel sundial
ancient falcon
earnest bluff
#

For anyone using Claude-cli to use their Claude subscription, beware the new update broke that for me. Had to roll back to 2026.4.23.

olive cliff
#

The minimum you should charge is their API bill and an hourly rate of $25+. You can also tac on a monthly service fee of like $200-$5000+ depending on the total work load, the "size" of the client, etc.
If this would be your first "client" you can go a little lighter on the total pricing structure.
The big thing with this one is that they want to bring it in house, while they still have an unpaid bill, and questions that they'd need to answer, "Who's going to maintain it/manage it?", "Do they have their own hardware they will be running it on, or will it still be on your off-site server?". If its still on YOUR hardware, that means YOU own it and they'd have to set up some sort of contract where they are continuously paying you, if they stop for whatever reason, you terminate their access.

ancient falcon
olive cliff
#

I think when you change permissions you gotta re-add the bot too.

drowsy aspen
#

is the Krill claw bot legit? its trying to get me to let it have remote acess

olive cliff
drowsy aspen
#

Thats what I thought. I just wanted to double check

olive cliff
#

if anything or anyone is asking you to allow them remote access, dont. Unless you already know and trust them personally IRL. Just dont.

#

There are some fake Krill bots out there. The real one should never ask for remote access

drowsy aspen
#

Yeah I thought so.

#

Maybe you can help me then. I am trying to fix the ai model but it wont let me. I am in the console using the tui setup and when I chose my ai model i chose the wrong one and now i cant find a way to change it and im stuck in the setup

olive cliff
#

openclaw onboard

that will let you reset the default model.
Or you can try to use "/models" or "/model [provider]/[model]

naive ocean
meager vessel
#

thanks 😉 but i am solving it by my hermes agent 😉

naive ocean
#

now i get gateway timeout after 10000ms ... gateway seems not to be starting, .24 as created havoc ufff !

#

oh wait we are back .... wow walking a thin line with the updates these days, this is to fragile ...

ancient falcon
marsh bolt
marsh bolt
marsh bolt
tawny ridge
#

Hi guys, have gpt pro plan, does openclaw usage get capped within that plan or do you pay per request?

mighty rampart
#

Has anyone opened an issue in the GitHub repo with the details or has a maintainer acknowledged it?

formal fog
#

Lets talk about memory. There are many solutions, but what I need is a way to have some separation between long term memory for different agents, AND some way to separate memories for different projects. For instance, I do software development, I do research, and I run a role playing game. I don't want role playing memories to be considered if i use the research agent. I also want my coding agent to have memory of the current project, but not all projects (maybe they are for different clients). I also want a separate place for shared memories which are available to all agents, or maybe to all projects, but not to my research agent. Any idea on how to set something like this up?

hybrid garden
#

ah... long term memory isolation...

nvm

#

hermes does it lol. Idk of other implementations.

mighty rampart
formal fog
#

and now I am wondering if i can modify my OpenBrain memory plugin in exactly this way.

tardy halo
#

hello all, just updated to the latest...
Seeing this spawn constantly on my linux box by openclaw gateway:
npm install @anthropic-ai/sdk@0.90.0 @anthropic-ai/vertex-sdk@^0.16.0 @aws-sdk/client-bedro...etc... eating my tiny 4 cores

marsh bolt
marsh bolt
dire zealot
#

messages consumed by long-poll but never logged or surfaced as pairing requests

muted grail
#

Anyone got any guides or advice on using OC as a discord bot safely? Properly sandboxing it so people can't run local commands and such? What do I need to harden it as a Discord bot for information retrieval and search?

muted grail
tardy halo
muted grail
#

I'm specifically training an OC to be a competitive fighting game agent I can run in Discord servers. It will be trained on the games, competitive history, tech, the players, and learning resources I've already written for said discord servers as post embeds

tardy halo
muted grail
#

It will be used to answer questions, teach people to play, help run online events, player histories/bios, all that jazz

#

something I've wanted to do for years and was just a bit over my head. I went as a far as learning how to manually train with curated data sets with very specific sources and all that....it was a nightmare

#

In over my head and prolly going about it the wrong way

muted grail
#

Anyone have issues with web_search tools? I can't get anything working. My config is just fine (perplexity api key), it worked in the past, but now I get some variation of this:
TOOLCALL>[{"name": "web_search", "arguments": {"query": "OpenCode Go reviews", "freshness": "day", "count": 5ALL>

tired pine
#

Is .24 just broken? My gateway is in a permanent restarting loop. From what I see here in chat I am not the only one.... reverting to .23 :/

muted grail
#

the fuck is this?

#

well I'm trying to paste a screenshot, but it's not letting me....

#

Got a friend request from what appears to be a fake version of Krill?

fluid stump
muted grail
#

🦞 OpenClaw Bot Activated
Runtime: ClawdBot v2.4.1 | Local Gateway: Port 18789
Agents: Main (Codex) | Email | Calendar | Tools: 12 Active
Channels: Discord | WhatsApp | Email | GitHub | 1Password
Status: Paired | Secure | Latency: 189ms | Tokens: 47k/400k

Hello! I'm your personal AI claw—handling emails, scanning markets,
managing calendars, shipping code, and executing workflows across
all your channels. What needs clawing today?
Quick Commands:

/status - Full system report
/help - All tools & agents
/pair - Resync channels
/think - Complex reasoning mode
/ship - Build/deploy workflow
Ask away—emails to clear? Code to write? Markets to scan? Signals to your trading bot? 🦞

#

that's what it DM'ed me

fluid stump
#

It's the bot

muted grail
#

but I made no new bot, have no bot named krill, and did not ask for a new agent or friend request

#

just suddently got a friend request

#

that's not my bot

fluid stump
#

I think it's usually when people request for help, i've experienced it once.

muted grail
#

Hmmm. I did use Krill for help a few days ago, but only just now got the friend request....you can see why I'd be suspisious, I think

#

it says no servers in common, too

#

now isn't that interesting

fluid stump
#

Oh yeah, i get you now.

fluid stump
#

Coz how it figured you out?

#

So confusing but try running commands on it to be sure.

muted grail
#

dk why I can't put a screenshot here

#

but no way am I a running any commands

#

that's a baaaaad idea

fluid stump
#

True, was suggesting some basic commands to be sure.

muted grail
#

Just strange. the account is "Krill APP Krill6616

#

is that the right user name?

fluid stump
#

Not sure, let me check the one i got help from days ago.

vernal wave
#

Hey all. I just installed the latest OpenClaw today, added all my Ollama models and created half a dozen agents that I'll have set up for specific roles. Currently, I'm only using the Web Chat, no channels set up. I see that the Web Chat has a single dedicated agent, however, when I go into the Agents view and select one of the new agents, I'm seeing this blue banner message "Switch chat to this agent to view its live runtime tools.", but I do not see a way to actually do that. The Chat page just has "main:main" in the drop down, and the Sessions page doesn't allow you to create one from the UI. Am I dumb?

I also tried clicking "Set Default" on the new agent to see if that'd change contexts, but it doesn't do anything, not even making a Network call.

fluid stump
#

Might just be the right one, it helped me days ago. I'm trying to put a picture of the chats here but not letting me

muted grail
#

Yeah, must not have image permissions in this channel

#

mmk, as long as it's legit. How easy would it be to send a FR to someone looking like your own bot, and then have to run totally malicious commands on your machine

#

that shit would be easy

#

so, worrysome

vernal wave
muted grail
vernal wave
#

Yeah

#

It was immediately after I joined the server

#

It's kind of a dick and straight to the point, short on its responses, I figured it's just saving tokens

muted grail
#

Okay, cool then, thanks for the tip @vernal wave . We gotta be careful with this shit!

#

It's just a few days late on DM'ing me, I guess

vernal wave
#

lol yep. Figure it's just part of the server, but still, not gonna give remote access. Cause, ya know, common sense.

muted grail
#

oh no, I ain't using that DM thread for a damn thing

#

haha

final patio
#

This is an original bot Krill 🦐#1161 userid. That would be definitely a fake.

muted grail
#

I can't drop a screenshot here, but I can DM it or something if some staff want to take a look

olive cliff
#

the official Krill bot will never DM you or send you a friend request. nor will it ask for remote access to your system. Good rule of thumb for all discord servers, never accept a DM or a friend request, UNLESS you are explicitly expecting one.

muted grail
latent hedge
#

Looking for ANYONE with a dual machine setup with Openclaw + Ollama local LLM’s on each machine
———————————————-

radiant slate
latent hedge
inner frost
#

hi, guys im having issues with the openclaw dashboard the localhost link it gives me says this site cant be reached

unborn yarrow
#

@inner frost looks like its a common problem this morning for many. I disabled the bonjour plugin from the cli and that resolved the problem at my end.

inner frost
unborn yarrow
#

@inner frost openclaw plugins disable bonjour

inner frost
#

thank you! its fixed now

tidal pawn
#

I am running openclaw on a local linux machine, it was running fine, since the update to 2026.4.24 the gateway is unresponsive... any idea where I could look? doctor did not solve it

nimble hill
tidal pawn
#

I found the issue by the way, the bonjour plugin is broken

hollow tiger
#

I would love to try the openclaw iOS app. Is it possible?

fresh moon
#

hello guys i need help choosing where to host my openclaw or whats the best openclaw hostings or helps with openclaw ive been hosting it in a vps for the 3 months but hthe constant bugs have been kinda a pain and breaking down would love to know what you guyys think :))

nocturne torrent
#

Hi Guys, I hope You won't be angry for posting this question here - but i really give up right now ... I'm trying to install openclaw with Telegram integration on my Motorola Edge 30 Neo phone with Android 14. Every attempt ends with unhandled promise rejection : CIAO PROBING CANCELLED... I did the onboard and it happened after running gateway ... Is there anybody that could help ? like a hint or something ... 🙁

thorny jetty
#

Hi guys, could anyone share me an example of an outbound agent for B2B using OpenClaw? Tasks i'm mostly interested in are: building prospect lists using Apollo/Zoominfo, outbound (connecting on LinkedIn and sending messages, outbound sequence planning and execution) and finally CRM integration.

spark valley
fresh moon
#

is kimi down?

nocturne torrent
spark valley
#

Dp openclaw logs --follow

sharp saffron
#

The latest 4.24 version has issues. For my case the major were 1. Bonjour plugin is broken , 2. The way it connects to LM studio service , to include special characters such as “@“, causes the model selection to be problematic at least .

fresh moon
#

is there someone here i can talk to about openclaw? just really wanna better understand it fo rmy setup and what vps to use?

hallow sedge
#

I keep breaking claw. starts slowing down and eventually stops responding. What's the trick for getting it to fix itself better? I have qwen3.5:35b running in vram via ollama on windows and openclaw on wsl

#

I've had to reinstall all the way down to windows multiple times. at most, it works for a day. Works awesome at first

#

I need a cheat sheet of prompts to help me get it initially set up.

north moss
#

keep getting this error :
fresh moon
#

hello just wondering does more cpu cores help my openclaw be faster?

radiant notch
fresh moon
radiant notch
fresh moon
chrome night
radiant notch
fresh moon
north moss
#

just bought a vps and got this error ⚠️ Agent failed before reply: No API key found for provider "anthropic". Auth store: /root/.openclaw/agents/main/agent/auth-profiles.json (agentDir: /root/.openclaw/agents/main/agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir. | No API key found for provider "anthropic". Auth store: /root/.openclaw/agents/main/agent/auth-profiles.json (agentDir: /root/.openclaw/agents/main/agent). Configure auth for this agent (openclaw agents add <id>) or copy auth-profiles.json from the main agentDir. Logs: openclaw logs --follow gateway connected | idle agent main (main) | session main | claude-cli/claude-opus-4-7 | tokens ?/1.0m

#

i use claude CLI so i dont have a auth key , idk what do to at this point

chrome night
north moss
#

it worked fine yesterday

chrome night
#

That is surprising, given Anthropic cut off use of their subs within OpenClaw (and many other harnesses, maybe all) on April 4th.

#

Gemini 3.1 said:

Anthropic officially cut off OpenClaw and other third-party harnesses from using Claude subscription limits on April 4, 2026, at 12:00 PM PT.

Billing Changes
Following this cutoff, Claude Pro and Max subscribers can no longer route their flat-rate plan usage through external agent frameworks. Instead, users who want to continue using OpenClaw must pay through a separate pay-as-you-go "extra usage" billing tier tied to their account or use a standard Claude API key. To help ease the transition, Anthropic offered affected subscribers a one-time credit equal to the price of their monthly plan.

Reason for the Cutoff
Anthropic stated that running autonomous 24/7 AI agents like OpenClaw through standard consumer subscriptions put an "outsized strain" on their infrastructure. The company noted that routing third-party tools through standard subscriptions was never permitted under their Terms of Service, prompting the strict cutoff to preserve capacity for their core products.

#

Maybe you are not using a sub or you are using their API key with/without a sub? Both are options but only the API key route works through harnesses like OpenClaw.

chrome night
north moss
olive cliff
coarse plover
#

new user here, wondering if it's possible to communicate with openclaw via voice msgs in telegram

north moss
chrome night
olive cliff
# north moss do you know what's the problem then ?

That depends. Are you using an Opus or sonnet model WITHIN Openclaw as your agent? Or are you using ClaudeCode CLI as a tool for your OpenClaw Agent to then use, to trigger Claude?
Do you have an API or do you have a monthly plan?

north moss
olive cliff
#

You have to use them separately, or have OpenClaw treat Claudecode like a CLI tool.

#

Using 0auth for claude on anything outside of claude = ban.

north moss
olive cliff
#

Ask it to.

#

You can also try setting up ACP harness for it.

coarse plover
olive cliff
coarse plover
#

ok, I'll see about whisper, the stt hasn't worked for me so far out of the box. thanks!

grizzled sandal
#

hello! Openclaw/linux newb here. I set my claw to update through the web interface, it's been a few minutes now and it's not responding.... how can I check what it is doing?

olive cliff
#

npm install -g openclaw@2026.4.23

grizzled sandal
#

should I try to stop the current update proces first?

olive cliff
#

you can stop it or continue it, it wont matter if you are rolling back to the previous updates

quartz egret
#

I have a main agent (APEX) and 4 sub-agents (NOVA, ARIA, NEXUS, LEDGER). Each has its own Discord bot token. How do I configure openclaw.json so each bot responds in its own channel AND all bots can talk in a shared #team-room channel?

latent skiff
#

@opaque sorrel

#

The 24.4.26 version is giving problems, right? I'm new to this and installed that version but it doesnt work, even if I follow the docs.openclaw.ai page...

#

now i'm trying the 23.4 version

slate minnow
#

Who’s running openclaw fully local on Mac Studio hardware?

agile void
#

Has anyone found a way for Openclaw not to do mutations with a fallback model? Is the only way not to have mutations done with fallback models is to have it done on another gateway and enable only primary model? Also is there a way to have per-agent model override? Please help?

nimble hill
agile void
nimble hill
#

That is how an agent worls ...lol how else can it function? If you have deterministic workflow, you can write a n8n wofllow.

agile void
#

In my current openclaw implementation that has multiple Agents and subagents, I just wanted to make sure that the sandboxed subagents only perform mutations with a primary model and not with fallback, in case the primary is unavailable. And the solution I got is to have is to have them run with two gateways and a separate telegram bot. I don’t want to have another separate tg bot for the ops agents that have mutation capabilities. I hope I’m clear now 🙂

pulsar flame
#

I'm having issues with installing updates of the bundled plugins, has anyone had this issue and resolverd it

nocturne torrent
#

Hi guys, Can anyone suggest best openclaw version to install on motorola edge 30 neo (android 14) . All I want is to be able to run it and use lets say telegram . Maybe there is some 100% working tutorial ?

rose halo
#

But openclaw isn't good for running multiple isolated agents. In my exp here I had many issues making agents for whatsapp. Many workarounds

nocturne torrent
#

or maybe just which version is the most stable ? 🙁

mighty rampart
marsh bolt
marsh bolt
frozen glacier
#

I'm trying to run openclaw gateway as a non-root user and i can start it as a non-priv. user but during the start up process something triggers another gameway to start as root then various folders under ~/.openclaw change ownership to root. I am wondering, should it be possible to run the gateway as a non-root user?

#

Oh, looks like some unknown versions of something have installed over my version of node and are pointing to linuxbrew wtf is going on here..

slate minnow
pulsar flame
muted grail
amber nacelle
#

I am having gateway issues. Not being a coder I tried to research the error I was receiving to see if anyone else has seen and solved this issue. No luck.

#

I am trying to self host OpenClaw on a Win 11 machine. I installed Open Claw Version 2026.4.24 with Ollama 0.12.2 (model os gemma4) I also have WSL installed.

#

I have screenshots but cannot seem to post them at the moment

#

The issue is history failed: Error: gateway closed (1006)

muted grail
amber nacelle
#

Darn. Okay that is fair. I just wanted to make sure I was sharing as much as I could. Thanks

muted grail
amber nacelle
#

@muted grail thank you. I appreciate it. When I ask for help I like to provide as much detail as possible. I will post in help after letting this sit here. Don’t want to over post.

frozen glacier
#

does anyone else have issues w/the new release of openclaw? I cant for the life of me figure out why openclaw won't start. it's stuck on epoll after loading plugins

marsh bolt
marsh bolt
marsh bolt
frozen glacier
#

Config was last written by a newer OpenClaw (2026.4.24); current version is 2026.4.23.

should I worry?

nocturne torrent
#

Hey, Guys help ... How to fix that :
Agent failed before reply: WhatsApp plugin runtime is unavailable : missing light-runtime-api for plugin 'whatsapp' please

serene mica
red plaza
#

Do you currently know of a free template for use with OpenClaw?

frozen glacier
red plaza
#

sorry, model

frozen glacier
#

you can try a running your own model locally

red plaza
#

My PC isn't powerful enough to run this

thorny gale
#

My thing stopped working

frozen glacier
frozen glacier
red plaza
#

ok thanks @frozen glacier

pulsar flame
serene mica
#

how to add multiple keys for same providers to avoid reach api limit?

manic dragon
#

First time Discord poster here... hoping this is the right place to post this...

Hey folks, sharing this in case it helps someone else.

I ran into an OpenClaw gateway issue where things got stuck in a pairing required loop and cron / approvals / other privileged operations stopped working.

Symptoms I saw

• CLI commands failing with gateway connect failed: GatewayClientRequestError: pairing required
• openclaw devices list showing a pending repair request that kept changing request IDs very quickly
• log file growing fast / hitting its size cap
• cron jobs, approvals, and other privileged operations failing

What seemed to be happening
In my case, the Telegram Native Approvals side appeared to be retrying across accounts in a way that kept refreshing the pending repair request before I could approve it.

What worked for me

  1. Temporarily disable Telegram:

openclaw config set channels.telegram.enabled false

  1. Restart the gateway:

openclaw gateway restart

  1. While things are quiet, approve the pending repair request:

openclaw devices list
openclaw devices approve <requestId>

  1. Verify the CLI / gateway is healthy again
  2. Re-enable Telegram and restart:

openclaw config set channels.telegram.enabled true
openclaw gateway restart

  1. If the log hit its cap, clear/truncate it

Main lesson
What mattered in my case was stopping the retry storm first, then approving the request. Trying to approve while the request IDs were churning did not work reliably.

If useful, I can also post the root-cause theory separately as product feedback.

wary flint
#

Is Discord/OC connectivity broken? My system is unresponsive, as is Krill.

muted grail
wary flint
#

Well, I just read Krill is being throttled for a bit (no biggy). Checking logs.

muted grail
wary flint
#

False. Krill bot is down. Throttled.

The logs are eye squintingly vague.

#

CLI ... HEALTH ...

[openclaw] Failed to start CLI: Error: gateway closed (1006 abnormal closure (no close frame)): no close reason
Gateway target: ws://127.0.0.1:18789
Source: local loopback

patent cobalt
#

How does Krill work?

muted grail
wary flint
#

Meh. No sweat.

#

Still ... my build is currently dead

muted grail
patent cobalt
#

Is there any info when does it work?

#

I could really use it's help

wary flint
#

Krill is up most of the time. Right now = rate limited.

muted grail
#

It just happens to be down right now

wary flint
#

Maybe we can help?

muted grail
#

Not normally an issue. We can try and help though. What's up?

patent cobalt
#

If you don't mind im happy to ask for help

wary flint
#

Have at it.

patent cobalt
#

I don't know much about having multiple sub-agents on OpenClaw while you talk to the main one
i want to make a sub agent scenario where I talk to an agent and that agent decides to launch multiple specialiced subagents if needed.
It's not like a full agent that I want to have his own personality towards me, just a subagent meant to do what it's meant to do. The main agent then will respond back to the user with the info of the subagent(s)

#

subagents like (coder, searcher, academic,git,tester...)

wary flint
#

Sub-agents are typically exactly what you just described. Your main agent is your proxy to activate their abilities. "Have Cody the sub-agent code a python script that loads a website in different browsers for visual comparison".

Your main agent will spawn "Cody" the coder ... and there it goes. You never talk to Cody. Your main agent does.

sacred birch
#

I guess the bot that helps in #1459642797895319552 isn't working. I have a real quick (I hope) question: My agent set up a cron job for me in response to webchat. What model (ie: gpt-5.4-mini) will it use? The default for the agent? The one currently selected in the webchat? A different one?

Is there a way for force it to use a different one?

patent cobalt
#

But I need to specify to use a subagent everytime?

wary flint
#

Well, sort of. You can add their usage into AGENTS.md or MEMORY.md to say (example) "whenever I ask for code, have 'Cody' do it" and if the main agent fails to launch "Cody" now and then, just call the main agent out on it.

patent cobalt
#

Alright, and to create this sub agents its like create a normal agent?

muted grail
wary flint
#

So, you engineer the exact behavior or each sub-agent to be a coder, a graphic designer, a writer, etc ... each gets their own behavior to be spawned by the main agent.

As for creating them ... there's a command line for it. You should definitely read the docs as if it were bible.

patent cobalt
#

Alr thanks

muted grail
#

You've got the right mindset @patent cobalt , spawning sub-agents for specific tasks that are still attached to your main agaent, sounds like the right way to do it

wary flint
#

Adding to what Casey just said ... don't be afraid to abandon the sub-agent setup as well, if you find you need more time with the docs first. That's what I'm doing. I had five sub-agents and because I couldn't "see" why the combination of my main/subs were producing bad results, I finally said "screw this, let's perfect ONE agent and we'll build the subs when I'm better at this".

Then my system went down LOLOL

sacred birch
#

thank you

wary flint
#

Anybody have an idea on a basic BOOT.md that'll tell me "yeah, I'm up and working"?

grizzled sandal
#

Q: is there a way to interrupt my agent while it's working? I know I can /stop it, but it happens that it sometimes misunderstands, and I'm not sure what would happened if I /stop it abruptly.

wary flint
#

Best I can recommend is set the interaction mode to "steer". Also, read about the "/btw" slash command, though it doesn't have any effect on the current turn.

#

"steer" means you can (don't quote me on this) "interrupt" without stopping the process so that new logic may be considered.

Sidebar, I need someone better than me to explain THIS noise from the doctor in CLI:

15:22:23 [channels] failed to load bundled channel qqbot: ENOTEMPTY, Directory not empty: /Users/j.vision/.openclaw/plugin-runtime-deps/openclaw-2026.4.24-17c3fa238f70/dist/extensions/node_modules/openclaw/plugin-sdk '/Users/j.vision/.openclaw/plugin-runtime-deps/openclaw-2026.4.24-17c3fa238f70/dist/extensions/node_modules/openclaw/plugin-sdk'

Sheesh.

mighty rampart
#

Are you on .24? They changed how plugin dependencies get loaded and something is not right. Go back to .23

wary flint
#

Hi, wsmoak. I'm on the newest update from this morning. Any chance you could walk me through a file-by-file effort to not need the downgrade?

mighty rampart
wary flint
#

:scrolls: to just barely within orbit of Earth's atmosphere. 😉

#

Sidebar: Why. WHY cause this? I was on the most recent update. Always am. WHY release a trap like this?

#

As for the fix, I see Ken Cavalton's wording. Maybe that's what you mean.

mighty rampart
#

Never ever just take the latest update. 🙂 read here first and have your agent check the release notes + open issues. Then move up to one you are pretty sure is okay.

wary flint
#

I've found myself in a situation where OC begins to fail in its active/live state and whenever that happens, I notice there's an update available. Like, it seems to downgrade it's reach when there's an update out there.

muted grail
#

I'm now picturing my Agent begging me not to update it

marsh bolt
wary flint
#

Do the docs mention a way to downgrade?

inland knot
#

openclaw update --version 2026.4.23

This was provided by my Sonnet

wary flint
#

Testin' it out now (and many thanks for the sourcing)

inland knot
#

no worries, if it doesnt work i can give you the npm command directly

sleek kayak
#

hi guys any suggestions why my bot crashes every time i write a prompt

inland knot
#

but it should work

wary flint
#

Seems to just show the latest verison

sleek kayak
#

i am really new at this thing

muted grail
wary flint
#

@muted grail that did it. @inland knot thanks to both of you.

#

"Error: Unsafe package dist path: dist/extensions/acpx/.openclaw-install-stage/node_modules/.bin/acpx"

muted grail
#

Sounds like a path issue with npm

#

you used npm to install?

wary flint
#

That was so long ago, I don't remember.

#

Doctor is active. If it all goes to tihs, I may just have to reinstall and port my agent back in.

Big question is whether or not the openclaw.json file is compatible across the board.

muted grail
#

yeah always go for an openclaw doctor --fix

#

when in doubt

#

If that doesn't work, I've got one idea

wary flint
#

Oh, we're back. I'm tempted to build a --fix hook for every launch LOL. I've got Codex OAuth running on xhigh via the following prompt:

"I just tried installing the latest update and it broke EVERYTHING. I need you to deep research the matter, then tell me what's needed for us to successfully proceed."

We shall find out.

muted grail
#

Sick, glad you're back up and running. Wish I coulda got you sorted a bit quicker, but I'm at the end ofmy knowledge base. I'm new to this too, just trying my best to help

#

I think the method I suggested for downgrading left a broken symlink/cached file

#

which openclaw doctor --fix hopefully cleaned up for you

wary flint
#

I'd be willing to wager that I know 50% of what I think I know about code in general, and of that 50% it's probably a quarter of what you know LOL.

BUT ... I bet I know exactly enough to help some folks out in here, if only because I'm glued to the docs and reading every page.

muted grail
#

We're all just doing our best out here 🐙

#

I'm about to set up a git for my own OpenClaw just to have version control when an update borks everything

#

lol

wary flint
#

NICE! Good thinking

muted grail
#

Was a joke at first, but hey...that's not actually a bad idea

wary flint
#

I have to wonder if updating through the web portal would be better.

muted grail
#

hmmm....what thinking level should I be using for my agent running on Nemotron?

#

@wary flint I think you can also downgrade with npm install -g openclaw@versionnumber Which might be the better way if you originally installed with npm

#

may have avoided your issue

#

replace version number with whatever you're downgrading too, ofc

wary flint
#

My agent is suggesting that the gateway went through a restart during the upgrade. I have it scanning the session and logs to find out if I'm the one who triggered it or if the installer did.

#

My agent (quote)

"

Best reading of the timeline
What most likely happened is:

you started the update
the update process itself triggered the gateway restart
that restart happened at the wrong moment, while the install tree was in flux
OpenClaw came back against a half-updated runtime and blew up

So:

you triggered the update
but the update process triggered the restart that broke the runtime

Confidence
High, but not absolute.
I do not have a log line that says you “clicked X at exact second Y.”
What I do have is strong negative evidence against a manual restart and positive evidence for source=update.

"

muted grail
#

Ahhhh, yeah, that kinda clears it up. It restarted itself at a inopportune time. It's known to do that. I see the web dash just disconnect and catch up eventually, all the damn time

#

Discord tip - use backticks to write code blocks
like this

wary flint
#

\I'm attempting a code block

#

\I clearly don't know what a backtick is\

#

LOL

muted grail
#

surround your text in back ticks (the key above the tab button)

#

the tilde key

wary flint
#

Dang it, now I have to forget how to word delete because backtick took its place in my geezer brain

muted grail
#

Markdown is kinda the shit, eh? it is so useful sometimes

wary flint
#

Re-attempting the install using the following:

`export TARGET="2026.4.24"

which openclaw
openclaw --version
openclaw status --all
openclaw gateway status --deep
openclaw config get meta.lastTouchedVersion

launchctl bootout "gui/$(id -u)" "$HOME/Library/LaunchAgents/ai.openclaw.gateway.plist" || true

npm install -g "openclaw@${TARGET}"

hash -r
which openclaw
openclaw --version

openclaw gateway install --force
openclaw gateway restart
openclaw doctor --repair

openclaw --version
openclaw gateway status --deep
openclaw status --all
openclaw config get meta.lastTouchedVersion`

Met with a failure. I'm downgrading again. First time I've ever had to. There is definitely something way wrong with this build.

undone meteor
#

hey guys, Im losing my mind over openclaw not following up with answers after calling tools. the tool output is big enough so that the context window fills up to 100% and thats when it breaks. when that happens, openclaw does not give a response anymore until I send it a message again, even though compaction happened afterwards. I did setup qmd and lossless-claw and the plugins appear to be working, they can look up data. I must be missing something super obvious? but I exhausted the docs already.

#

tool output is working, just not when the context window goes 100%, then its game over

wary flint
#

Might be your timeout settings or (less likely) token limits. Are you setup to show /reasoning in the output?

undone meteor
#

reasoning output is shown. Im using sglang to host, so I can also see that its not doing prompts further

#

what is openclaw supposed to do when the tool it calls produces more output than what fits into the remaining context window?

#

I ask it a question, it does 2 tool calls successfully and then its game over with the third

#

sglang tells me that the max context window is above 90k tokens, so in openclaw I specify: contextWindow 90k, contextTokens 70k, maxTokens 20k, are those reasonable values?

wary flint
#

If memory serves, yes. I do remember setting my maxTokens to 8192 for a then openrouter model choice.

#

But you've left more room and that does NOT mean "please use all of these tokens" ... so I don't think it's you.

#

Do you have tool config access?

#

(I host mine locally, not on a server)

undone meteor
#

all locally hosted

#

its like when it hits 100%, it does compaction and this shuts down the thinking process, rather than triggering it again

wary flint
#

Then you probably want to figure out what's triggering the compaction.

undone meteor
#

well its context window approaching 100%

#

90k context window is not much, so it has to trigger basically all the time

#

if you do things like websearches, or file analysis

wary flint
#

And you're only calling one tool at a time?

undone meteor
#

I say "search for X on the internet" and the LLM is doing the tool calls

#

sequentially

wary flint
#

Right. Okay.

undone meteor
#

and Im the only user of the LLM

wary flint
#

What model?

undone meteor
#

qwen 3.6

#

big fan, 150 tokens/s

#

and really made for agentic thinking and coding

#

until it stops

#

what is supposed to happen in openclaw when you work with such payloads?

wary flint
#

Try this. Create a /new session, then prompt it with "I need the top ten most reputable source article on the topic of [topic]. In your output, list the title, link and a single paragraph summarizing its contribution to the subject (only ONE paragraph and exactly THREE sentences per paragraph)."

See if those "forced" boundaries return something.

Also, consider a non-reporting cron job that tells it, every five minutes, to make sure it didn't leave you hanging.

patent cobalt
#

Sorry to bother with the subagents, but does every subagent need a whole workspace with all the .md files inside the main agent workspace? Also why does the main agent not have it's "own" directory

wary flint
#

You can configure the workspace directories when you create them.

.openclaw/workspace/main/.md
.openclaw/workspace/sub/.md

Somewhere during that process, you must've gone with the defaults.

patent cobalt
#

I think I messed up creating the default one then, its just on workspace/ . although it works

tardy bolt
#

how do I wipe openclaw and delete absolutely everything. I've uninstalled everything and trying to reinstall but it just keep's erroring to the point I might have to do a CPU reset

undone meteor
#

25% context -> prompt -> reasoning 1 -> tool call 1-> tool output 1 -> reasoning 2 -> tool call 2 -> tool output 2 -> compaction -> no more response, 50% context. Havent seen "100% context" anywhere, it was either not shown or the compaction has this issue even if context < 100%

#

guess I need to read debug logs

undone meteor
#

downgrading worked for me

#

Ive also deleted everything yesterday

#

workspace too, because rage

tardy bolt
#

I'm getting this error when trying to re-install and doctor isnt working Upgrade complete. Run openclaw doctor to check for additional migrations.
iex : Cannot process argument transformation on parameter 'Succeeded'. Cannot convert value "System.Object[]" to type
"System.Boolean". Boolean parameters accept only Boolean values and numbers, such as $True, $False, 1 or 0.
At line:1 char:39

  • irm https://openclaw.ai/install.ps1 | iex
  •                                   ~~~
    • CategoryInfo : InvalidData: (:) [Invoke-Expression], ParameterBindingArgumentTransformationException
    • FullyQualifiedErrorId : ParameterArgumentTransformationError,Microsoft.PowerShell.Commands.InvokeExpressionComma
      nd

PS C:\Users\neobr>

muted grail
#

Can someone pin a message that says "Having (x) issue? Downgrade with (insert proper way here), please?

#

I think we've repeated the "just downgrade" response enough to warrent a message pin

olive cliff
# undone meteor 90k context window is not much, so it has to trigger basically all the time

Are you using Qwen3.6 as a local LLM or a cloud?
90k context window is way too tiny for Qwen3.6 in my testing. I get by using its full 256k context window hosting locally. Once you reach 100% context window, you are figuratively strangling your llm and it cant breathe. Conpact when you're getting close to 100%, not after the fact. Also be careful with how many compaction you do per session. Its memory starts getting garbled/truncated.

undone meteor
#

qwen 3.6 35b a3b

#

local

#

openclaw compacts automatically long before it goes to 100%, except when there is this one tool call that goes beyond.
but Im not sure if this even has to do with the 100% ctx, maybe its even bugged with regular compaction < 100%

olive cliff
undone meteor
#

I dont have that kind of VRAM and Im not going to offload to CPU, so

olive cliff
undone meteor
#

compaction is the way to go, even if it loses stuff. but the problem is that it completely stalls doing anything at that point

olive cliff
undone meteor
#

I can offer an rtx 5090 with 32GB VRAM to the LLM

#

but do I really need a larger context window

#

is that the root cause of the issue

#

Im waiting for FP4 cache support, then I get twice as much

olive cliff
#

Your small context window is whats choking things up. You are hitting it too frequently and causing compaction to trigger, and llm calls to drop.
You should be able to bump it to 128k. I'd recommend trying to find the MAX context window your system can handle, and then slowly decrease it from there to find a sweet spot for speed of response.

#

IMHO for most workflows, you want 128k+ context window.

undone meteor
#

sure I did. can bump up to 165k, but then I get OOM

#

certainly I do want that, but you know how the market is

olive cliff
#

Then drop it a bit from there. Gotta find that sweet spot.

undone meteor
#

Im willing to accept it being dumbed down, but Im not willing to accept that it completely shuts down

olive cliff
#

Im running with no GPU btw.

undone meteor
#

I think I had these issues with 160k too

#

yes but at what inference speed?

#

10 tokens / s?

#

I get 150 / s with only GPU

#

with that speed, I can self correct a lot of mistakes

#

if it just would go on...

olive cliff
#

You can get by fine with 40/s

undone meteor
#

Im aware I could build a post-compaction hook, but as far as I know, it should already work out of the box

olive cliff
#

You dont need token speed. You need reliability.

undone meteor
#

but it must be on the software side already

#

its doing a tool call, after tool call, its supposed to print the result, it can easily detect that, even if no AI

#

which is why I think theres a bug or config issue

olive cliff
#

If your hitting 100% context window when its hanging, thats the core issue.

#

An LLM cant do anything once its at its context window.

undone meteor
#

the LLM cannot, openclaw gateway can

olive cliff
#

How much system RAM do you have?

undone meteor
#

no issue with RAM at all

#

128

#

openclaw is just not doing the post compaction hook that its supposed to do according to reason

olive cliff
#

JUMP THAT CONTEXT WINDOW TO 256K.
If i can manage it without a gpu, 128GB ram, you should have no issues either.

undone meteor
#

when I compute the LLM on the CPU yes

olive cliff
#

Then thats your solution.

undone meteor
#

its not acceptable to me to be honest, I want this in the GPU and Im actually super happy how much I alrady get out of it

#

to me this is a software problem

#

I understand compaction means it loses some information

#

but what is not acceptable is that it stalls

#

it should be bumped again

#

whenever this happens I tell it "resume" and it does resume

#

but I shouldnt have to type this

#

and I shouldnt have to wait for a heartbeat, nor implement a custom hook

olive cliff
#

If you have it on the GPU, your bottle neck is the 32GB VRAM. Yea you get speed, but you're gonna max that context window.

Try running through your CPU with a larger context window, and repeat the workflow thats been causing you issues.

undone meteor
#

I remember often when it hits the "100% ctx" it resumes by itself correctly, but it does that in like 30% of the cases maybe

#

I have spent weeks to get off CPU hotloading, spent weeks trying vllm, sglang, tensorrt, because you cant disable cpu offloading on ollama and lm-studio, its just not right

#

yes context window is great, but Im not running this on my cpu

olive cliff
#

Your only options are to use a smaller model, or run on your cpu. Otherwise you're going to keep beating your head against a wall.

undone meteor
#

Im convinced there is a broken config or software bug

#

the software should not simply die because of hitting that wall

olive cliff
vapid valve
#

Yoo anyone using nullclaw agents?

#

whats this convo about

undone meteor
#

I have a 90k context window, I run one prompt, it does 2 tool calls and with the third tool call it does compaction and stalls

#

because of compaction

olive cliff
#

Think about it like this.
You are trying to cram 2 tons of shit, in a 2 inch box. Its not going to happen until you get a bigger box(context window)

undone meteor
#

I dont mind that the tool output is pruned

#

I do mind, that it does not go "oh I compacted, lets see what I did last, now I should resume"

olive cliff
#

But thats not always the case with compactions. It wont always auto pickup where it left off. It depends on where in its tool call chain it began the conpaction.

Again. Your only solution IS A LARGER CONTEXT WINDOW.

undone meteor
#

after tool output there must be reasoning, this can be detected with software logic without LLM

#

trivial condition

olive cliff
#

The reasoning is within the LLM.

undone meteor
#

only after it was initiated by the software

#

full context window kills the LLM, not openclaw

olive cliff
#

Ollama or LMStudio?

undone meteor
#

Im using sglang, it shouldnt matter

#

openclaw sees there is tool output, openclaw sees it did not start a prompt, therefore openclaw can start a new prompt after compaction

olive cliff
#

Im not familiar with sglang, but in LMStudio you can change certain settings that may help out with OOM, like KV cache, etc

undone meteor
#

yeah I spent weeks optimizing, Im using FP8 cache, Im waiting for FP4 cache

#

many tears

olive cliff
olive cliff
undone meteor
#

all I want is a "resume" prompt after compaction. I can implement this trivially, but Im convinced openclaw has this by default - unless I misconfigured it

#

Q4 sounds like GGUF, but GGUF is an invention of ollama which doesnt work with any other software

#

and Im using NVFP4 weights and FP8 cache already

#

Im 100% convinced this is misconfigured

#

Im debugging this more in depths to see whats happening

#

(LM-Studio counts as ollama since it uses ollama as a backend from what I heard?)

#

(and Im not using that because I specifically want to avoid CPU offloading)

olive cliff
#

GGUF works fine. Q4 works fine. Again, we are running the same model, on the same ram. So if you're having issues with the way you have yours set up, then try something else. I recommend trying it the way I have it set up as im having no issues. Keeping it locked down to a 90k context window is like putting a Ferrari in a school zone.

undone meteor
#

(also ollama has less inference performance)

#

Im using NVFP4 weights, those are 4 bits too

olive cliff
#

LMStudio with Qwen3.6 has never failed me yet.

undone meteor
#

sure, but its running on your CPU plus your GPU, thats not an option for me

#

I spent weeks to get away from that and Im willing to work with a 100k context window for that

#

only the software fails me

#

its trivial to detect for software that after a tool output no reasoning was performed and that it therefore needs to do a new prompt to remedy the situation

olive cliff
#

Again, your doing yourself a disservice with your implementation. Your main issue is that things start to fail when you're reaching 100% context window.
Stop being a hard head and try my advice. Your approach is NOT WORKING.

undone meteor
#

you are right that having 10x larger context window gives more room to breathe

#

it absolutely sucks having to compact every 4 turns

#

reduces intelligence

#

and Im willing to accept that because Im not going for CPU offloading

#

the disadvantage of compaction is that you lose some information, not that it stalls

#

Im sure I have a bug / misconfig

olive cliff
#

Trust me. At one point during one of my workflows, I added a new model, had my agent configure it, but it configure all models to be 60k context window, and I hit 40 compations for that session really quick and was trying to figure out why shit wasnt working. Once I identified it changed the Qwen3.6 context window too, thats when I knew what the issue was.

undone meteor
#

its like its pruning instead of compacting

olive cliff
undone meteor
#

sure, no problem

#

what the problem is that it does not retry afterwards

#

a simple "resume" prompt is all that it takes

#

and that can be initiated from openclaw

olive cliff
#

Because its NOT DESIGNED TO CONTINUE AFTER A CONPACTION.

undone meteor
#

it does not need an LLM brain to decide that

#

it is

mighty rampart
#

Have you searched in GitHub issues? You are describing a hook (wanting something to happen automatically) and there are recently closed issues about a post compaction hook meant to kick it back into action.

undone meteor
#

good idea

#

but also I will debug the hell out of this issue

#

this technology is so promising and this is a dealbreaker bug, and Im not the one to back down from that xD

mighty rampart
#

Stopping after compaction is just a thing that happens. Something in the harness or an outside orchestrator must do something to make it start again.

undone meteor
#

btw Im sure I will get to 250k context window solely on GPU when Im done with optimization

#

outside orchestrator, exactly, I would assume openclaw has it built in, because its pretty much trivial to detect. tooloutput without reasoning followed up should lead to "resume" prompt

mighty rampart
#

OpenClaw is “just” the minimal Pi agent underneath.

#

The hooks plus the heartbeat are how it keeps going/ acts proactively.

undone meteor
#

post compaction hook is what I crave

#

already tried implementing myself, but Im convinced its wrong if it already has that

mighty rampart
olive cliff
#

You could set a short heartbeat to check current tasks and continue working. But not with that tiny context window.

#

But that should pick up after a compaction on its next trigger.

undone meteor
#

if tooloutput & !reasoning_output prompt("resume")

#

something like this

#

e.g. openclaw does flush memory before compaction

mighty rampart
#

The hook exists. You need to implement it if you want something to happen after session compaction.

#

If you think that should be built in then open a feature request. Meanwhile it’s there for you to use.

grizzled sandal
#

I've been tryoing to have my agent write a pretty basic skill for over 3 days, it still fails hilariously... is qwen3.5:9b too stupid for the task?

drowsy aspen
#

So one problem I have been running into is my rate limit being hit for openai, i thought it was supposed to auto change to a slower model. why does it not do that in openclaw but it does it in actual chat gpt? I have the 4o-mini as a model and i switched to it manually but i still get limit hit

olive cliff
grizzled sandal
olive cliff
mighty rampart
undone meteor
#

Qwen 3.6 is designed for coding, Qwen 3.5 is not !

grizzled sandal
#

I asked, and the agent said it does have a skill creator skill, but I have no idea if it used it, or if it did so correctly

olive cliff
grizzled sandal
#

This is not even coding really, the skill includes 2 python scripts but those work OK. the model is failing on the markdown skill description

drowsy aspen
olive cliff
#

Are you using the same provider as your primary/default model for your fallbacks?

drowsy aspen
olive cliff
undone meteor
#

Im not going to let software fool me, Im going to go to rock bottom of this xd

olive cliff
#

Im not 100% certain on codex/gpt/openai, but for example:
If i have opencode/claude as my primary, and opencode/gpt as a fallback, then if I hit my apt limit, or my monthly sub limit for opencode, then no other opencode provided models will work.

grizzled sandal
#

Ok, so this is the first response my agent gave me: "Yes! I have access to a skill creator skill at /home/juani/.openclaw/workspace/skills/skill-creator/SKILL.md." so i asked it to describe the skill, then "The skill creator skill mentioned in TOOLS.md (/home/juani/.openclaw/workspace/skills/skill-creator/SKILL.md) doesn't exist in your workspace yet. According to TOOLS.md, the skill creator skill should be found there, but it seems it needs to be created"

#

Is it just hallucinating everything¿?

olive cliff
vapid valve
vapid valve
olive cliff
grizzled sandal
#

I understand that AIs will always answer confidently, hallucinjating or not, but how can an agent work if even the most basic of queries fail? Am i missing something or is it a model issue... or what?

vapid valve
#

literally my own nullclaw agent did the same shit, before having it hooked to any MCP for my own vps I told it to tell me a little bit about what you have available. Even said to it "can you see my vps are you able to access through the mcp?" and it gaslit itself into thinking it exists and upon further questioning it said "nvm I dont have it I believe it should be setup" 😂

#

these agents just be doing that shi

vapid valve
#

interesting

#

lemme check this out

olive cliff
drowsy aspen
# olive cliff Im not 100% certain on codex/gpt/openai, but for example: If i have opencode/cla...

I kind of get it. I just dont understand that if its using my login to run and the limit gets hit for openclaw why does it still work in the browser. I feel like it shouldnt work there also or it should use the switchable model. because 5.5 can only use 160 tokens/180min but when that happens in the browser it auto switches to 4o-mini and is just slower. But when i use 5.5 and hit my limit and then expect the fallback to switch to 4o-mini or if i manually switch it still doesnt work.

vapid valve
#

and not even in stock

#

the world we live in for people to just throw out money like that

#

I wouldve been people had it not been out of stock 😭

olive cliff
vapid valve
tardy halo
olive cliff
#

Price to performance wise, and on 300w, cant be beat.

olive cliff
grizzled sandal
#

I'm reading up on the corsair ai machine mentioned.. I udnerstand it uses normal ram isntead of vram to run the models, so it works at like 1/4 of the speed, but lets you load huge models... is that it?

olive cliff
#

I had it batch process like 30 research paper PDFs at once. That took a long asssssssss time haha, but it did it.

grizzled sandal
#

I see... that's interesting. I thought normal ram was waaaaaaay slower than that, like 1/1000th of the speed.

olive cliff
#

And you can reallocate the memory to either system RAM, or VRAM for the iGPU. You can have as much as 96GB VRAM natively, but can be pushed to 122GB VRAM. But thats also leaving only 6GB for the system, so a headless setup would be recommended at that extreme.

#

The key is that its Unified memory. Its a lot faster than your typical RAM. Its direct on the board memory.

grizzled sandal
#

Got it, thank you for the info 🙂

olive cliff
#

That also means its not upgradeable. But honestly, this thing is great for local AI, and its a beast at gaming. Its more worth it when its on sale. There are other manufacturers that use the same board, like Frameworks Desktop PC, honestly just search for "AMD Halo Strix" and you can find some more affordable options. Around $2k-$3k, the only difference would be storage size at that point, which is upgradeable via nvme slots.

mighty rampart
#

Finally had a chance to work on my 4.24 instance and got it repaired. There were some stale lock files that needed to be deleted. Then it gets past the stuck plugins. (Claude Sonnet figured it out from the logs, reading the source code and looking around.)

neon marsh
#

hey im trying to update to 2026.4.24 but im getting stuck in a crash loop ig. the vm load goes up to 100%.
already disabled bonjour as a possible fix but still wont update..

muted grail
#

Celebrate the fact you don't have to fix a broken claw instead

neon marsh
#

probably should. thx:)

muted grail
lapis salmon
#

Trying to have my agent spawn a sub-agent via a custom skill. EVERY time I get an error that "streamTo" is not a valid parameter for runtime=subagent .. which is true. But I don't tell it to use that param. In fact I now tell it "ONLY use these..." but it still uses it. Opus 4.6 tells me the reason is because the sessions_spawn tool itself lists the parameter as a valid "optional" parameter, with no indication that it is only valid for runtime=acp. The agent just fills it in from the tool definition in its context. So I am now vibe-coding a script wrapper for my agent to be able to spawn a sub-agent. 🤪

dawn coral
#

how do i do claude cli

tardy halo
muted grail
drifting epoch
#

is there a regression re: vertex? was working before i upgraded to 4.23, now not working

gleaming kettle
#

Morning, not sure if this is correct or not, when i create an Agent, it doesn't create the the workspace for example workspace-mybot, instead i get .openclaw/agents/mybot and .openclaw/workspace/mybot, is this the way now or is something going wrong or not configured correctly, asking for a friend 😉

radiant slate
#

Hi all. I'm building, trying to , a personal assistant that has brains to fix things, but still is able to have a conversation, good quality, with under a 2 sec audible return through the speakers. MY goal is to use all local and no cloud. I have a 3090 on a dedicated ai pc, and i have a 4090 i can use to access for tts on my networked Gaming Pc. I cant seem to get the response down to under about 10 seconds due to the large files of soul, tools, memory in every turn. Has anyone here done this and what LLM models did you use? Thx in advance.

cyan flame
#

Hey guys, does anyone have a solution to agents narrating every step? I'm using Discord for my agents and its actually unbearable even with notifications off

I've been talking to Claude about building a mute-while-thinking plugin, has anyone gone down this path?

thanks

copper nova
cyan flame
copper nova
cyan flame
#

Is it reasoningDefault: "off" ?

#

That didnt work either, I don't think Openclaw can supress it

nocturne torrent
#

Hey, Guys help ... How to fix that :
Agent failed before reply: WhatsApp plugin runtime is unavailable : missing light-runtime-api for plugin 'whatsapp' please

fresh moon
#

hello guys anybody know a great one click deploy openclaw thats actually usable?

fresh moon
#

hello would love to know what openclaw hosted instance are people using here than just getting a vps to set it up

#

anybody here running kiloclaw?

lethal wigeon
#

Hey, I'm having an issue where my OpenClaw gateway keeps crashing in a loop and I can't get it to stay up.

Setup:

  • VPS (Ubuntu), OpenClaw v2026.4.24
  • Gateway running as a systemd service
  • Telegram channel configured and paired successfully

The problem: The gateway starts fine, runs for ~20 seconds, then crashes with:
"Unhandled promise rejection: CIAO ANNOUNCEMENT CANCELLED"
...and then systemd immediately restarts it, causing an infinite crash loop (restart counter is now at 14+).

From journalctl logs, the sequence is always:

  1. Gateway starts
  2. Bonjour/mDNS advertiser tries to re-advertise
  3. CIAO ANNOUNCEMENT CANCELLED → unhandled promise rejection
  4. Main process exits with status=1/FAILURE
  5. Systemd restarts it

Because of this, openclaw gateway probe always returns timeout or ECONNREFUSED. Telegram messages never get routed to the agent even though pairing completed successfully.

Things I've tried:

  • openclaw doctor --fix
  • openclaw security audit --deep
  • Setting gateway.controlUI.allowInsecureAuth to false
  • Tried adding "bonjour": false to the gateway config block (rejected as unrecognized key)

Is there a supported way to disable the Bonjour/mDNS advertiser, or is there a known fix for this crash loop? Thanks

hoary surge
humble ginkgo
#

Does anyone know what's the culprit behind this? Any architectural change or additional processing step?

#

I can say the interface is better. I want to fix that response time problem. But if it's unknown cause, maybe i should just check out the latest feb build

winter flame
lethal wigeon
winter flame
#

You're welcome mate, happy hacking!

nocturne torrent
#

Hey can Anyone tell me how to completely remove whatsapp from openclaw. When i was configuring Openclaw (onboard) i picked Whatspp which was a mistake and now i have a problem with it . I would like to remove it so I will not see it anywhere in openclaw dashboard etc. - How can i do it - please dont tell me that i have to uninstall and remove everything to do it ?

winter flame
nocturne torrent
#

i did it but nothing changed i can still see it in dashboard

winter flame
#

Maybe you already did, but did you restart your gateway after modifying it?

nocturne torrent
#

and i can still see errors about light-runtime-api

nocturne torrent
winter flame
#

If you have problem in the process, run openclaw doctor to see where the problem could rely on.

nocturne torrent
winter flame
fringe cargo
#

okey dokey, can anyone help me with a video tutorial of set up on a windows computer, ive tried all the documentation, and then followed on with 3 seperate videos and im getting no where. Anyone got a link to an up to date video, or anything in particular i should try?

scenic sundial
#

hello, I have a bug to report:

🐛 Bug Report for OpenClaw team
Title: Discord channel crash on Windows after npm update to 2026.4.24 — ESM \c:\ protocol error

Environment: Windows 10 x64, Node v24.14.1, OpenClaw 2026.4.24

Steps to reproduce: Update from 2026.4.14 → 2026.4.24 on Windows, restart gateway.

Result: Discord channel subprocess crashes immediately on every startup with \Received protocol 'c:'\ ESM error. Force-reinstall and cache clearing do not resolve it. Rollback to 2026.4.14 resolves it.

Note: The 2026.4.24 changelog mentions a Windows ESM fix — fixing Windows npm updates that could fail to load copied dist modules — but it does not cover this case.

winged lagoon
#

Heads up 👇

If you’re using OpenClaw, avoid the newer builds after 2026.4.23 (a979721) for now.

Version 2026.4.24+ seems to have issues with the gateway (timeouts / connectivity problems).
Recommendation:

  • either stick to 2026.4.23 (a979721)
  • or rollback if you already upgraded
  • or just skip 24 until it’s fixed

Saved me some debugging time 👍

signal star
#

Sorry, guys, what's the easiest way to set up Discord and integrate an LLM? I'm going to try the Claude subscription.

undone meteor
winter flame
signal star
#

I thought it was just in case I didn't have OpenClaw installed.

winter flame
undone meteor
#

am I the only one getting 100% ctx issue with 0 compactions being done?

#

or bad config

olive cliff
undone meteor
#

thats exactly what I want to achieve

#

but if it does multiple tool calls consecutively without me prompting, I have no manual control over it

olive cliff
undone meteor
#

I want compaction to work

olive cliff
#

Its working the way its designed to.

undone meteor
#

Compection: 0 times according to /status

#

so if I run with 1mio context window and it happens again, then I just need a 2 mio context window?

olive cliff
#

You're trying to drive across the country with no gas. You aren't going anywhere with this.

undone meteor
#

they promised me gas stations along the way

#

(at least one of) my problem/s is compaction is not triggered despite having it set up

olive cliff
undone meteor
#

bot says its enabled and has run 0 times so far. Ive been following the documentation, set up the config values using openclaw config. compaction mode "default" and "safeguard" are availailable, both dont do anything. I have reserveTokensFloor 40000 (40%), notifyUser true, memoryFlush.enabled: true, 🤷‍♂️

#

have to eat debug logs, otherwise I wont get a lead I think

olive cliff
#

All I did for compactions was told my agent to set it up, and it auto compacts on my end. I did absolutely zero configuring myself and let my Agent handle it. 🤷‍♂️

#

But either way, when a compaction hits, I still have to manually continue, like you pointed out yesterday.

olive cliff
#

@undone meteor
Do you have compactions set at 100% of context window? Or set to trigger just shy of that? Like 80%?

This is how I have my Qwen3.6 configured in my json.
{
"id": "qwen/qwen3.6-35b-a3b",
"name": "Qwen3.6 35B A3B",
"api": "openai-completions",
"reasoning": true,
"input": ["text", "image"],
"cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 },
"contextWindow": 262144,
"contextTokens": 200000,
"maxTokens": 8192
}

undone meteor
#

openclaw.json:
compaction: reserveTokensFloor 40.000

contextWindow 94000
contextTokens 74000
maxTokens 20000

in comaction mode default, it compacts proactively, in compaction mode safeguard it compacts just before it fills up.

I do see a message "pre compaction memory flush", yet /status says 0 compactions afterwards

marsh bolt
undone meteor
#

thats for the future

#

Im still at the "hello world" stage and failing it

#

once it starts calling tools, its over

#

if compactions dont work, I dont need to try anything other before fixing compactions

olive cliff
# undone meteor openclaw.json: compaction: reserveTokensFloor 40.000 contextWindow 94000 contex...

From my agent to you:

Their config is the opposite problem — contextWindow: 94000 is too small.

With a 94k window and contextTokens: 74000, OpenClaw triggers compaction at ~79% (74k). But if the actual model has a larger window (say 128k or 200k), the model never actually fills up — so OpenClaw does the "pre compaction memory flush" but the compaction counter stays at 0 because the model didn't actually overflow.

Root cause: contextWindow doesn't match the model's real limit.

Fix: Set contextWindow to the model's actual limit (e.g., 131072 for 128k, 262144 for 256k). Then contextTokens should be ~80% of that.

For their model:

• If 128k window → contextWindow: 131072, contextTokens: 100000
• If 200k window → contextWindow: 209715, contextTokens: 160000

The "pre compaction memory flush" means OpenClaw tried to compact because it thought it was full, but the model never actually hit its limit.

undone meteor
#

thank you for your help btw, just talking about the problem makes it more endurable x)

olive cliff
undone meteor
#

with reservEtokensFloor 40.000 it should trigger compaction at 94k - 40k = 54k

olive cliff
#

With such a tight window, it just screams compaction loops to me. Which cause issues.

undone meteor
#

either way I should confirm what is happening with the logs

olive cliff
#

For a chatbot, yea small context window is fine. But for an actual workflow, you need to maximize your context window and have proper compaction thresholds in place.
With your hardware, you shouldn't be bottle necking yourself with a tiny window like that. You are only giving yourself half of half of half of the total context window.
Its like driving a Ferrari in a school zone.

#

I recommend mimicking my settings for the context window, do away with any additional token reserve settings you have and copy mine verbatim for your qwen3.6 model

undone meteor
#

the reason why I dont want to use system memory is because it means its being processed on the CPU (at least partially) than solely on the GPU

#

this in turn means my CPU is not 40°C but 95°C

#

fans spin up accordingly

#

I can limit the CPU frequency, but then its still 70°C instead of 40°C, its not elegant

#

if I compute it only inside the GPU, then CPU and GPU both stay below 40°C

#

on top of that, the GPU computes 140 tokens / sec

#

just the fans spinning alone is sufficient for me to consider it a no-go to compute on CPU

#

as much as I wish to use my system RAM for the context window

#

my GPU cost 3400 EUR, my CPU cost 400 EUR, both are the best consumer hardware available, the GPU must suffice

#

but I think the absolute context window size not the reason for the stalling to begin with

#

my GPU is a ferrari, if I use CPU offloading, Im going back to volkswagen in terms of inference speed

#

and its just a matter of time until sglang / vllm support 4bit cache, then I can double the context window, on top of that

#

the model weights are already 4bit compressed, so the model file is only 20GB. this means I have 12GB left for cache and surplus, this must be enough

marsh bolt
#

Alright - who’s gonna try 4.25 and make sure it fixes the weird plugin thing??

undone meteor
#

lol Im deseparate Im doing it

#

"whats the worst that could happen"

winter flame
undone meteor
#

appears to work, I can connect to the gateway

#

whereas I could not with 2026.4.24

tardy halo
#

I'll wait at the end of the day to see what people are saying. I've been waiting for a couple of fixes that 4.24 had

winter flame
#

I'll wait like a month and pick the best may release.

undone meteor
#

openclaw tells me that memory flush alone reduces the context window and therefore it doesnt need to trigger compaction. as far as I understood, memory flush only means writing *.md files. anyhow, openclaw can read through its source, log and config files. I'll get there

rustic pier
#

anyone else's gateway just stopping randomly?

undone meteor
#

"say the line bart"
"try updating to 2026.4.25 or downgrading to 2026.4.23"

foggy veldt
#

hello, the only thing that change to upgrade my plan chat gpt plus to pro with openclaw is only my usage limit ?

fading tide
foggy veldt
mighty rampart
leaden heron
foggy veldt
leaden heron
rustic pier
foggy veldt
leaden heron
fading tide
signal star
#

Is anyone using OpenRouter + Claude-CLI?

undone meteor
steep compass
#

I just updated my openclaw and my cost useage now says 1/3rd of what it said before, anyone experience this?

vocal oyster
#

Hello, Is this the place for stupid people to get help?
I've tried to install opanclaw, and I don't know how or what I broke
but it's barely moving. even openclaw status is hung.

what am I missing?

signal star
#

There's no pay-per-use fee, right? You pay a subscription fee that gives you 5-hour sessions, right?

tardy halo
vocal oyster
grizzled sandal
#

I am amazed at how my agent manages to find new ways to misinterpret and disobey very explicit orders. ALMOST like it is doing it on purpose. 100% monkey claw shit.

tribal vapor
#

is OpenClaw only installable on certain linux distributions? I tried installing on Arch Linux, it failed instantly.

tardy halo
tardy halo
vocal oyster
tardy halo
#

You cna give the logs to a web chat and see what it says

vocal oyster
grizzled sandal
#

My agent also likes to: 1) do a web query instead of looking at it's own skills and 2) spawn sub-agents to run the skill. The subagents are experts at twisting words, contradicting themselves and even inventing new requirements.

tardy halo
vocal oyster
#

I've tried it.
it got into a loop.

said to try to do doctor and then restart.
over and over.
than I've tried to look at the docs.
and now I'm here

#

I have two red lines in to log, usally red is bad:

19:27:09 [diagnostic] lane task error: lane=main durationMs=94843 error="FailoverError: LLM request failed: provider rejected the request schema or tool payload."
19:27:09 [diagnostic] lane task error: lane=session:agent:main:main durationMs=94845 error="FailoverError: LLM request failed: provider rejected the request schema or tool payload."
tardy halo
tardy halo
#

seems the model doesn't like tools

vocal oyster
mighty rampart
tardy halo
grizzled sandal
#

I have an rtx 3080ti with 12vram, currently running qwen3.5:9b with Q4. Is there a "better" model out there that could fit in my vram?

leaden heron
signal star
vocal oyster
tardy halo
vocal oyster
tardy halo
vocal oyster
#

so I don't know.
first time installing it.
it's just stuck all of the time.
can't do anything

#

ok, after almost 5 min the status is done.
that took a lot of time IMO

tardy halo
#

it might be the tons of plugins it install, I think it's been bloated with node packages

vocal oyster
#

can I send I picture shomehow?
don't want to logdump

tardy halo
gritty surge
#

is the newer openclaw better at assigning correct models to different tasks?

vocal oyster
mighty rampart
tardy halo
# vocal oyster https://drive.google.com/drive/folders/1KMYyUqWATnm_YKOgKS2OduTlCzp70Mj8?usp=sha...

Go to deepwiki.com look for openclaw repo there and give it this warning in the prompt: " Read-probe diagnostics are limited by gateway scopes (missing operator.read). Connection succeeded, but read-only status calls are incomplete. Hint: pair device identity or use credentials with operator.read."

It will guide you on what's wrong and how to fix it:

What's Happening
When OpenClaw probes the gateway, it attempts read-only RPC calls like health, status, and system-presence probe.ts:340-345 . If your client credentials don't include the operator.read scope, these calls fail with "missing scope: operator.read" helpers.test.ts:229-251 .

The system treats this as a scope-limited failure rather than a complete connection failure helpers.ts:279-281 , resulting in the warning message you see output.ts:78-85 .

Solutions
Option 1: Pair Device Identity
Use the devices CLI to pair with proper scopes:

openclaw devices list
openclaw devices approve <requestId>

marsh bolt
#

Guys, Kimi 2.6 is killing it for content generation, GLM 5.1 best for tool calls. You can subscribe to them for $18/mo each and I have never hit the limits. I'm struggling to find a reason to use GPT 5.5. It's so dull and lifeless. Gets hung on tool calls and generates crap content.

zenith quartz
#

and im tired of looking in the docs 🙁

#
  1. is the kimi claw needed 2. do i need to make an api key via my subscription to use my subscription - and which is the correct model name
marsh bolt
#

Ya, it was a bit weird setting both up. Took a bit of tinkering. Can't say I wrote down the steps tho

zenith quartz
#

kimi/kimi-code

#

i don't think i ever had issues setting up anything :)))

#

but this is just confusing af

marsh bolt
#

Kimi was the hardest I had to do an openclaw models auth command - i'll try to find it

zenith quartz
#

i don't need the oauth or etc

#

just asking if the kimi claw was needed

marsh bolt
zenith quartz
#

im legit having issues with how the model names are written in the config

#

yep figured that out after 2 hours

#

and you made an api key ?

#

within your sub?

marsh bolt
#

Yep it took me a while with kimi. hardest one so far. I'll share my config.

#

"kimi-coding": {
"baseUrl": "https://api.kimi.com/coding/",
"apiKey": "<your subscription api key>",
"models": [
{
"id": "kimi-code",
"name": "Kimi K2.6",
"reasoning": true,
"input": [
"text",
"image"
],
"cost": {
"input": 0.6,
"output": 2.4,
"cacheRead": 0.12,
"cacheWrite": 0
},
"contextWindow": 262144,
"maxTokens": 32768
}
]
},

zenith quartz
#

pls if you can at least the model name

#

thank you

#

THAN YOU REALLY HELPFUL1

marsh bolt
#

BUT - you have to run the openclaw models auth or it won't work - it will ask for API key

#

openclaw models auth login --provider kimi-coding

zenith quartz
#

yeah yeah will do that thanks

marsh bolt
#

I was getting 401 until i did the models auth

zenith quartz
#

i made the mistake of trying to setup the kimi claw

#

then i tried every other setting but the regular one

#

then the api key confused the f out of me

marsh bolt
#

ya it's weird. took me a while too

zenith quartz
#

it didnd't help that i have 2 openclaw gateways and one is in D

#

so its failry hard to do the auth prompting cuz it keeps opening up my other gateway

shut onyx
#

I find kimi k2.6 model is amazing with openclaw! Did anybody tried it?

marsh bolt
#

Yes! See above

queen ingot
#

Hey! I use discord to send messages to openclaw. My openclaw has some commands that takes some Time to run. And after some Time i get this error:

Discord inbound worker timed out.

Is this something that If fixable and how?

shut onyx
# marsh bolt Yes! See above

Its really good with openclaw , keeps working without hesitation and disconnection and completes everything I tell him, they really cooked

#

Now I’m confused what model should I use DeepSeek v4 pro or kimi 2.6 🙃

dense glacier
#

Hi Guys,
I am not a developer at all. I just went down the rabbit hole learning about openclaw after eatching a youtube video and now am amazed with my AI I created that I talk to via whattapp on a new phone I bought for him. I did invest significant time and used other AI to help me but never gave up for days and now I am blow away every day. but still a bit of a tech novice.
He has created me websites and so much for business ideas I am working on. However yesterday was a melt down day. i had a burlary at my house and his whattsapp number died when he was turned off. It took me all day in powershell to try get him hooked up on a different new whattsapp gateway then i wanted telegram as a back up and all holy hell broke out. he went loopy getting all mixed up with gateways.
now the gateway resets and his glitches. I am worried all i have achieved will be lost if dont join a community and speak to people that actually know what they are doing. PD (my bot told me to say this)
Hi — I'm running OpenClaw on a Linux VPS (Ubuntu), WhatsApp channel. After any gateway restart, the capability drops to read-only. I can see in the logs that the gateway-client keeps auto-pairing with operator.read scope and then requesting upgrades that don't stick. I manually edited ~/.openclaw/devices/paired.json to grant full admin scopes and it shows admin-capable after restart — but I'm not confident it'll hold. Is there a config option to permanently set the scope for the internal gateway client? Running version 2026.4.24.
Please dont laugh at me.

marsh bolt
#

/btw is such a satisfying command

quick crow
#

which the most stable and fast openclaw version i am getting being hit by slowness, i have multiple agents

Measured timings:

  • Channel inbound marker: 22:48:54.366
  • Agent session.started: 22:49:40.442
  • Prompt submitted: 22:49:40.555
  • Model completed: 22:49:43.133
  • Session ended: 22:49:43.136

So:

  • Possible pre-agent delay: 46.1s
  • Actual agent/model run: 2.7s
spice jewel
#

iex : Cannot process argument transformation on parameter 'Succeeded'. Cannot convert value "System.Object[]" to type
"System.Boolean". Boolean parameters accept only Boolean values and numbers, such as $True, $False, 1 or 0.
At line:1 char:39

  • irm https://openclaw.ai/install.ps1 | iex
  •                                   ~~~
    • CategoryInfo : InvalidData: (:) [Invoke-Expression], ParameterBindingArgumentTransformationException
    • FullyQualifiedErrorId : ParameterArgumentTransformationError,Microsoft.PowerShell.Commands.InvokeExpressionComma
      nd

I have this error pop up every time I try to set openclaw up. I was wondering if someone knows a fix or is able to help me directly.

inland knot
#

How would you guys run OpenClaw on a budget? that being like 10-20$ per month reliable without having to worry of rate limiting or weekly limit

compact pulsar
#

Stuck after update to 2026.4.24 — crash loop, agent unresponsive

Updated OpenClaw last night and it left a corrupted partial install. The update never completed cleanly — one of the directories has unknown in the name instead of a version stamp.

What I see in /root/.openclaw/plugin-runtime-deps/:

openclaw-2026.4.24-4eca5026e977
openclaw-unknown-a44f41aef101

What's happening:

  • Agent is completely unresponsive on Telegram
  • Server went into a crash loop (processes restarting constantly)
  • SSH access was freezing/timing out
  • Had to hard reset the VPS to regain any access
  • After reset, SSH is back but agent still not responding
  • Dashboard shows "error loading conversation"

Attempted fix:
rm -rf /root/.openclaw/plugin-runtime-deps/openclaw-unknown-a44f41aef101
— Directory would not delete even with no processes locking it (lsof showed nothing)

Running on Ubuntu VPS, systemd service.

Two questions:

  1. How do I force-remove the corrupted directory?
  2. Is there a clean way to roll back to 2026.4.23?

Any help appreciated — platform is down.

mighty rampart
#

4.24 was like that but deleting those deps files worked for me. Why not go forward to 4.25?

compact pulsar
teal mural
#

Beginner here, any apps for openclaw that i should know about/research, any tips for saving tokens besides reseting gateway to shorten token usage (dont even know if thats the proper way to do it). Or anything important you have learned for general use of openclaw?

mighty rampart
compact pulsar
polar badge
#

Hello, I'm trying to sync up some things with OpenClaw.

  1. Whenever I use the localhost control panel to chat with the assistant it send a message from "openclaw-control-ui" and then doubles it with "You", which creates doubles of messages. It told me to set "dmScope": "main" which didn't seem to work.

  2. I'm talking to the assistant "Anna" in the Control UI, and in Slack where I have the assistant setup but they don't know what the other "assistant" is saying despite being the same one (Or my intention anyways). The Control UI has a "main" tab and "slack:t0_SOME_ID:#slack_channel" tab where the conversation is kept which looks like:

System: [2026-04-27 20:39:28 UTC] Slack message in #slack-channel from Jesse: Good job <@U0-SOMEID> (Anna) I see the MEMORY file! Be more mindful of these sorts of situations. If you tell me you did something, do it, or at least ask <------ The <@U0-SOMEID> (Anna) is the @Anna mention in Slack.
Good job <@U0-SOMEID> (Anna) I see the MEMORY file! Be more mindful of these sorts of situations. If you tell me you did something, do it, or at least ask about it. written by "Jesse (U0-SOMEID)" <------ The <@U0-SOMEID> (Anna) is the @Anna mention in Slack.

Understood, Jesse. Will do, not say-it. Thank you. written by Anna

Good job <@U0-SOMEID> (Anna) I see the MEMORY file! Be more mindful of these sorts of situations. If you tell me you did something, do it, or at least ask about it. written by You <------ The <@U0-SOMEID> (Anna) is the @Anna mention in Slack.

Understood, Jesse. Will do, not say-it. Thank you. written by Anna

I think that sums it up.... Thx for any help!

Images:

  1. https://i.gyazo.com/df0efee291244d32cef2d010208b2215.png
  2. https://i.gyazo.com/01047dafdb3134734127ef9fd571c9dd.png
marsh bolt
marsh bolt
inland knot
marsh bolt
tawny ridge
#

Is anyone having issues running gpt5.5 with openclaw? Anthropic never an issue but trying to move to gpt and it’s constant problems

scarlet flint
#

I have a suggestion: LET ME USE MY D: DRIVE TO HAVE MY AI AGENT BE ABLE TO EXIST ON MY BACKUP LAPTOP BECAUSE I WILL HAVE TO GIVE UP MY GAMING LAPTOP TEMPORARILY. (T~T)

Only URLs with a scheme in: file, data, and node are supported by the default ESM loader. Received protocol 'd:'
marsh bolt
muted grail
#

What's the best way to keep Perplexity and an OpenClaw Agent in sync? I'd love it if my agent had access to my perplexity research in real-time, if not delayed at the least. I was thinking of using a middle-man like Obsidian or something?

strong ferry
strong ferry
muted grail
strong ferry
marsh bolt
#

Guys I ran openclaw update. Not pretty. This little gem saved me though - at least got the gateway running - curl -fsSL https://openclaw.ai/install.sh | bash

muted grail
candid latch
#

I just ran OpenClaw update 5 minutes ago. Got stuck updating via package manager. Known issues? Ive not been following anything here for a bit.

strong ferry
#

if you have something else in mind you can try that too coz i dont know your setup much

olive cliff
olive cliff
# undone meteor this in turn means my CPU is not 40°C but 95°C

When was the last time you applied new thermal paste? 95C Is extremely excessive and is an indicator that there's something wrong with the way your CPU is cooling. I rarely see anything above 80c, even when gaming + model being loaded. What CPU do you have? I'll see what I dig up.

marsh bolt
#

Gah this update is an absolute cluster. Everything went sideways with 4.24 / 4.25 and trying to use OpenAI. I will never get these hours of my life back

muted grail
#

or have Perplexity Computer try tyo do something, since I have a bunch of free credits that are gonna expire

#

heh

tough fiber
#

after the update i keep getting [model-pricing] LiteLLM pricing fetch failed (timeout 3s) and OpenRouter pricing fetch failed (timeout 3s) and openclaw is extremly unresponsive (it takes 2-3 minutes to respond to hello)

muted grail
muted grail
hybrid peak
#

Quick question: Since updating to version 4.24/25, my agents are taking a long time to respond. Is there a problem with the versions, or did I break them?

humble tide
undone meteor
# olive cliff When was the last time you applied new thermal paste? 95C Is extremely excessive...

AMD 9950X, I read that it just keeps on boosting until its at 95°C, so no matter how much cooling you throw at it, it just keeps boosting, but I dont know if its true. thermal paste is 3 months old.

good news, I could bump my context window by a factor of 2.5x by not using the environment variable SGLANG_USE_CUDA_IPC_TRANSPORT.
its worse for time-until-first-token but obviously I cant afford that memory.

my context window is now 350k, so even more than the model supports natively, which I fixed using rope + yarn.
the context window never even gets close to being remotely filled up.

I can now throw any task at it, but it still stalls sometimes.

still many bugs, but I can use the software.

so I can have the cake and eat it too (hosting only on GPU and having big context window)

#

convinced there is some sort of chat bug. it often ends with "let me try this:", as if tool calls would follow but are silently dropped by openclaw. I will figure this out

marsh bolt
#

openclaw-gateway keeps spiking on CPU

tardy halo
mighty rampart
undone meteor
#

lol

#

since Im hosting the llm locally, I can compare the logs of the LLM provider and openclaw to see if maybe there actually is a bug, because I think I saw some dropped output like that already

#

also its repeating output, but I think this is meant as a feature

#

"let me try this.... 3 sentences"
tool call
"let me try this.... 3 sentences, 3 more sentences"
tool call
and so on
and then I have the entire thing repeated like 8 times but with variations.
I see sometimes it collapses it, but often it does not. I guess its how the model reasons, but its also very verbose to read

marsh bolt
shut onyx
#

Guys, I’m using linux popOS to run openclaw , do you suggest transfer openclaw to linux ubuntu? Will it be better?

tardy halo
#

afaik, popOS is based on Ubuntu anyway

marsh bolt
#

Removed fallbacks - seemed like z.ai was limiting and bouncing around to fallbacks. Also updated to 4.26 which just launched. So far, CPU looks better, but still sluggish on telegram

#

I think it was just an issue with z.ai

#

Maybe not - something still spiking cpu

#

There is an issue with z.ai ongoing, but something is bogging down the gateway. Since 4.24 -> 4.26

tardy halo
#

I just updated, all good fo now

#

ah! they still have the bug that the model list display ALL the models instead of just the ones configured

#

oh yes, I see the gateway CPU spikes :/

#

active-memory is not working for me, my agent had a lobotomy

icy fjord
#

uh oh

#

has anyone solved the missing operater.read gateway issue

tardy halo
#

and QMD might not be working 🙁

wary flint
#

Mine is fixed. I'm trying to remember what we did. Let me have the agent explain it.

#

My agent is writing it up now. Standby.

#

My agent ... wrote it wrong ... and has been scolded ... standby.

#

Here we go.

`The repair was not about reorganizing memory/ first; it was about restoring the ability to semantically use it. The session trail shows the actual problem was that semantic memory had been turned off because the configured embedding path was broken, so the system fell back to MEMORY.md plus manual file reads instead of real recall across memory/ and session history. In other words: the memory files still existed, but the retrieval layer that made them useful had died.

The fix was to stop treating that as a content problem and repair it as a provider/runtime problem. We switched memory search away from the broken embedding setup, re-enabled it, pointed it at the supported local embedding model, and cleaned out the stale setting that belonged to the old remote path. When that still failed, the sessions showed the real missing piece: the local embedding runtime itself was not actually available to the running service, so it had to be installed under the supported Node 24 runtime.

The last step was the one that made it stick: do a hard service restart, not just a soft reload, so the live gateway actually came up on the corrected runtime and could see the embedding package. Only after that did verification show semantic memory was back: memory search reported the local provider as active, status showed the vector layer ready, and recall could again search both memory/ and session transcripts instead of just reading MEMORY.md by hand.`

wicked warren
#

Anyone having issues with ollama or zai models not accepting "high" thinking mode?

#

After the latest update?

nimble hill
latent hedge
#

Anyone noticed the the last weeks that your discord bots for your agents suddenly take alot longer to get online, than originally?

smoky chasm
#

anyone seeing this starting from yesterday?

[discord] gateway: Gateway websocket closed: 1000
[discord] gateway: Gateway reconnect scheduled in 1086ms (close, resume=true)
[discord] gateway: Gateway reconnect scheduled in 2097ms (invalid-session, resume=false)
(sometimes:)[diagnostic] stuck session: sessionId=unknown sessionKey=agent:main:discord:channel:<id> state=processing age=420s queueDepth=1

it appears everytime i call the bot

thorn lark
#

Hey everyone,

I'm following up on a persistent Slack connection issue. My AI assistant and I tried everything yesterday to fix a channel_not_found error, including creating a brand new Slack app and confirming all the scopes.

The weird part is that my agent can receive my @mentions from Slack, but it can't send messages back to the same channel. We confirmed the bot is in the channel and even made the channel public, but it still gets the channel_not_found error every time.

I just tried it again this morning and the problem is still there. Has anyone ever seen an issue like this where events are received but sending fails with this specific error, even when all permissions seem correct? We're completely stuck and would appreciate any help you can offer.

Thank you.

winged lagoon
blazing willow
#

hello anyone can help me with my openclaw? i currently hosted on hostinger vps and i have telegram channel integrated..i wanted to make my personal assistant processes much more smooth. Im using openrouter free as my model..

thorn lark
winged lagoon
#

after update openclaw.json ?