#Gateway becoming very slow . CPU 100% - Versions 4.24 and higher. [76382]

1 messages · Page 1 of 1 (latest)

gentle nacelle
#

After upgrading from 4.23 to 4.25 (and later 4.26), the WebUI (Control UI) and gateway became very slow/laggy. Commands from the web interface take a long time to reach LM Studio, even though the model itself is healthy. Rolling back to 4.23 immediately restores normal fast performance on the exact same hardware and setup.
Setup

OpenClaw version: currently pinned at 2026.4.23 (stable)
Model server: LM Studio 0.4.12 running locally (mainly qwen3.6-27b-uncensored-hauhaucs-balanced + 35B variant)
Hosting: Ubuntu server, VM inside proxmox with 8xCPU + 12GB RAM) accessed via Tailscale WebUI
Connection: OpenClaw → LM Studio OpenAI-compatible endpoint
No cloud providers as primary (OpenRouter was removed)

What happens

On 4.23 → instant response, very snappy
On 4.25/4.26 → noticeable gateway/WebUI delay (sometimes 30-90+ seconds before the model even starts processing)
LM Studio logs show normal prompt processing, but the delay is clearly before the request reaches LM Studio.
Htop inside openclaw gateway, shows one of the 8 CPUs at 100% for a while.

Question
Has the WebUI/gateway latency regression introduced in 4.25 been addressed in the latest betas? Or is there a known heavy component (new plugin registry, TTS, telemetry, etc.) that is causing this?
Any recommended config tweaks or flags to make newer versions as fast as 4.23 while keeping the new features would be greatly appreciated.
Thanks!

low pondBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

digital flumeBOT
#

This does sound like a real regression class, not LM Studio.

I checked current release state: npm latest and beta are both 2026.4.26 right now, so there isn’t a newer published beta beyond the one you already tested. The next relevant fixes appear to be on main / unreleased, not in a newer beta yet. Release channel docs note that npm dist-tags are the source of truth, and dev tracks moving main/experimental builds. (docs.openclaw.ai)

What matches your symptoms best:

  • Likely culprit: plugin/runtime metadata, provider discovery, config normalization, or memory hooks before the LM Studio request is dispatched.
  • Less likely: TTS, unless you explicitly enabled auto-TTS. Auto-TTS is off by default. (docs.openclaw.ai)
  • Not impossible: telemetry/OTEL, but I’d suspect plugin/provider discovery first.

There are several relevant follow-up fixes already in the changelog after 4.26: repeated bundled runtime scans before chat turns, plugin/config CPU pegging before agent turns, default memory recall delaying turn startup, and new /readyz event-loop diagnostics. (raw.githubusercontent.com)

For a 4.25/4.26 test, I’d try these low-risk tweaks:

openclaw config set models.pricing.enabled false
OpenClaw
Release channels - OpenClaw
# 
openclaw config set memory.qmd.update.startup off
openclaw gateway restart

If you only use WebUI + LM Studio, you can also test narrowing plugins, but don’t set plugins.enabled: false because LM Studio is plugin-backed. Use an allowlist carefully, e.g. include lmstudio plus anything else you actually use. Plugin docs cover allow/deny behavior. (docs.openclaw.ai)

Good diagnostics to capture during the slow window:

openclaw --version
openclaw status --all
openclaw logs --limit 500 --plain
curl -fsS http://127.0.0.1:18789/readyz

If 4.26 still shows 30–90s pre-LM-Studio delay, I’d stay on 2026.4.23 for now and file a GitHub issue with those logs + your exact config shape. This is exactly the kind of regression maintainers need a repro for: https://github.com/openclaw/openclaw/issues

sudden dock
#

i thing im seeing the same thing here too.. its getting too slow

gentle nacelle
fiery bolt
#

👍

#

me too

tidal granite
#

is it still slow for you guys?

gentle nacelle
#

According to the developers, it will be fixed shortly, hopefully in the next update.

dense valley
#

it's still slow on 2026.4.29

nimble dune
#

updated yesterday from 4.20 to 4.29 and now have 1 minute+ lags on every commands

dense valley
#

yes! I'm builidng local-router and I just wait for 1+ minute for OC to hit my proxy

nimble dune
#

I am hesitant to downgrade, as it means reinstalling from npm, and I am afraid it might break my config.

nocturne furnace
gentle nacelle
#

Currently, I m trying to install 4.30.

nimble dune
gentle nacelle
#

4.30 has been merged to the main repository, but it's not out yet. The developers have to make it available in npm. Patience for the time being.

spark tusk
#

been monitoring the bugs around performance today, and they're all slowly getting closed, so several fixes are in main already, but not on a named release yet

cosmic ocean
#

possible to update to main but then back to release?

spark tusk
#

should be -

If you want an exact trunk commit instead of a moving branch:

npm install -g github:openclaw/openclaw#<commit-sha> --no-fund --no-audit

Return to a release tag
If the git tag is v2026.4.20:

systemctl --user stop openclaw-gateway

npm install -g github:openclaw/openclaw#v2026.4.20 --no-fund --no-audit

gentle nacelle
#

"2026.5.2-beta.1"
Is out.. Heading to give it a go

gentle nacelle
#

Gateway becoming very slow . CPU 100% - Versions 4.24 and higher.

#

Problem remains in 2026.5.2 beta 1 - tested.

nocturne furnace
#

I reverted to .23 and things are back to normal again.

#

I am glad you're taking one for the team and testing out the Betas @gentle nacelle

tardy flicker
#

its been super super slow for me too, and its node.exe using 100% cpu. its stuck in a loop reading AppData\Roaming\npm\node_modules\openclaw\dist\extensions* over and over

gentle nacelle
#

what hardware are you using?

gentle nacelle
#

5.2.beta 2 still laggy. I got probably good dev response in the #general : Tool caching (a — refresh on gateway restart)

The problem: createOpenClawCodingTools() is called fresh every run. It builds 40+ tool instances by calling createCodingTools(workspaceRoot) from @mariozechner/pi-coding-agent (the pi 0.68 regression), wrapping each with workspace guards, policy resolution, sandbox checks, etc. This is the core-plugin-tools: 4s stage.

What's cacheable: Most of the tool construction is deterministic given a stable (workspaceDir, config, agentId, modelProvider) tuple. The per-run varying inputs are:

  • sessionKey, sessionId, runId, trace — used in tool closures for session routing
  • abortSignal — a new controller per run
  • messageTo, messageThreadId, senderId, etc. — message context for reply routing
  • hasRepliedRef — mutable ref

What that means: The tool definitions (schemas, names, policy decisions) are stable per (workspaceDir, config, agentId, model). The tool closures bind per-run ephemeral state. You can't cache the full tool objects because they close over abortSignal and message routing context.

But you can cache the expensive part: The real cost is createCodingTools(workspaceRoot) calling pi's internal tool factories 40+ times. That returns bare tool objects before OpenClaw wraps them. You could cache that result at the process level keyed on workspaceRoot, then re-wrap the cached base tools with per-run context each time — skipping the expensive factory calls.

This is an extension-level intervention — it means patching createOpenClawCodingTools behavior, which is internal to the core. The cleanest approach is an upstream fix (cache createCodingTools result in pi itself or in the tool-prep path), or pinning pi to 0.67.68.

snow lark
#

Ugh. I’ve been thinking it was a misconfig when I reinstalled today on a beefier VPS. Spent hours debugging this slowness. Can I downgrade to 4.23 after having fresh installed 4.29? What are the steps?

gentle nacelle
#

Yes, it's possible. But my advise will be to get some rest, and the problem may be fixed any moment now.
"2026.5.2-beta.1",
"2026.5.2-beta.2"
]
moltbot@Ubuntu-Clonebot-local:~$ npm view openclaw versions --json

gentle nacelle
#

2026.5.2 looks really well -built, but does not fix the issue. Still 100% CPU, Still very slow response from the LLM models regardless local or cloud.

gentle nacelle
#

Minimal config test result (plugins = 0)
Applied extreme minimal config:

plugins.allow=["lmstudio"]
All channels disabled
No memory, no skills, no browser, tools.profile="minimal", tools.deny=["*"]
Ran openclaw doctor --fix

Results:

Gateway starts with 0 plugins in 6.2s
CPU still ~103–105% constantly (idle + simple chat turn)
Simple agent command: WebSocket closed, fell back to embedded
No improvement in main-thread load

This confirms the regression is in core gateway / embedded runner code (likely createOpenClawCodingTools() or similar), not in plugins or channels.

snow silo
#

thank you @gentle nacelle for keeping us updated.

i had the same problem (running hella slow) on 4.29. I just reverted to 4.23 and speed is much better

pallid swallow
haughty lagoon
#

This has been escalated, thanks for flagging and for the GH issues, keep this thread updated if anything changes for either of you please. Thanks!

gentle nacelle
#

@haughty lagoon We appreciate your work and help. Can’t wait to test the next beta/fix.

haughty lagoon
#

Gateway becoming very slow . CPU 100% - Versions 4.24 and higher. [76382]

#

My pleasure - all props here go to our awesome maintainers 🙂

gentle nacelle
gentle nacelle
sand mason
#

Please dump OPENCLAW_GATEWAY_STARTUP_TRACE=1 openclaw gateway run openclaw gateway stability --json openclaw gateway diagnostics export --json in your issues, will help with fixes

haughty lagoon
#

@gentle nacelle - yeah, what Vincent said. Thank you very much!

gentle nacelle
#

No luck yet. Tried everything.

ember compass
#

Has the slowness been solved yet. I am not upgrading now. 👍

marble dome
#

I’m experiencing the same issue as well. It started on 2026.4.24 and continued until 2026.5.2, and I’m still looking for a solution. I’ll keep an eye on this thread. Any suggestions would be appreciated. I’ve now rolled back to the 2026.4.23 build. By the way, I’m a native Windows user.

gentle nacelle
#

@marble dome I am finishing a migration, and I will try 5.3.-beta.1 or beta.2. In a bit.
In windows do you see 1 Core being stuck at 100%?

tardy flicker
#

it was hard to spot at first, because its just 3.1%, but thats 100% of 1 thread of my 7950x3d

pallid swallow
tardy flicker
#

24

pallid swallow
gentle nacelle
#

Partly fixed. Now, openclaw gateway is workable. 100% CPU not persistent . The delay is at ~6 seconds.

#

2026.5.3.beta.2

pallid swallow
#

My debugger agent pinned it down to a session list call issue in the openclaw mac app / dashboard, and I just updated it to the 5.2 version, to see if that fixes it.

I thought I had seen somewhere else where that was related to the token cost usage feature in the dashboard, but not sure where I saw that.

gentle nacelle
marble dome
#

Hi， I updated to 2026.5.3.beta.2 and it seems fixed my problem. I will keep trying. Thank you!!

junior pine
#

im waiting for 2026.5.3 to release before updating my openclaw instance. Hoping it comes out soon.

tardy flicker
#

so 5.3 does fix the 100% cpu bug in node, but its still 2 minutes between when i send a message, and my gpus start doing anything. no idea why its hanging for so long

torn meadow
#

yes, it's still very slow...

sand mason
#

dump your /status output

pallid wind
#

4.26 works great, everything above that - which i think started with 4.29 is not good. I installed 5.2 and had to revert back (gateway issues, continuity, etc), i have not tried 5.3 due to trust issues... cadance should be less updates but better quality. updates feel rushed. i use codex only which has been a huige positive , no claude - claude is a disaster with claw now. however, seems to be working find for devs so likely user error or how my current architecture is set up. I can install 5.3 and just debug but thats not ideal

torn meadow
#

@sand mason

OpenClaw 2026.5.4-beta.1 (9cc3ae1)

Version            2026.5.4-beta.1
OS                 linux 6.8.0-110-generic (x64)
Node               22.22.2
Config             ~/.openclaw/openclaw.json
Dashboard          http://127.0.0.1:18789/
Tailscale exposure off
Channel            beta (installed version)
Update             pnpm · npm update 2026.5.4
Gateway            local · ws://127.0.0.1:18789 (local loopback) · reachable 625ms · auth token · vmi3151434 (94.136.184.173) app 2026.5.4-beta.1 linux 6.8.0-110-generic
Gateway service    systemd installed · enabled · running (pid 492796, state active)
Node service       systemd not installed
Agents             2 total · 0 bootstrapping · 0 active · 85 sessions
Secrets            none

CHANNELS
Discord  ON  OK  token env+config (sha256:87cfc5bd · len 72) · accounts 2/2

DISCORD ACCOUNTS
default  OK  token:env · credential available in gateway runtime · allow:227359709965975552
krill    OK  token:config · credential available in gateway runtime · allow:227359709965975552

AGENTS
main          ABSENT  66 sessions  active 7h ago   ~/.openclaw/agents/main/sessions/sessions.json
krill (krill) ABSENT  19 sessions  active 9h ago   ~/.openclaw/agents/krill/sessions/sessions.json

DIAGNOSIS
✓ Config: /root/.openclaw/openclaw.json
✓ Secret diagnostics (0)
✓ Restart sentinel: none
! Port 18789 already in use — gateway already running locally (pid 492796)
! Tailscale exposure: off · daemon unknown
✓ Skills: 13 eligible · 0 missing · /root/.openclaw/workspace
✓ Plugin compatibility (none)
✓ Channel issues (none)
sand mason
torn meadow
# 
10:33:01.360  Bootstrap: checkpoint hit (273 stored msgs)       597ms
10:33:01.957  Assemble: live context, 6 msgs (lossless-claw fallback)
10:33:02.003  Message ingested
10:33:02.099  Assemble for LLM: 3 ctx items, 7 msgs            142ms
              ↓
              LLM generation                                   5.56s
              ↓
10:33:07.655  Response delivered
10:33:07.947  Post-cleanup (context-engine queued)             292ms
#

gpt-5.4

sand mason
#

i need the full /status inside of chat

torn meadow
# 

Phase    Time
Bootstrap + ingestion    739ms
LLM generation    5.56s
Post-cleanup    292ms
Total gateway    6.59s
Why 5.56s on a 7-message context? The session was idle for ~9 hours (last turn was 01:35 AM). Provider-side prompt cache is completely cold — gpt-5.4 had to process everything fresh. Still within normal range for a cold start.

Compared to yesterday's first ping (5.2s): essentially the same — cold cache, new session. No tool-stages trace in the beta to confirm pdf-tool status. The dedup-oversized issue from the post-restart session (storedCount=273) is still present but not blocking delivery.

One thing to watch: assemble: falling back to live context on every turn means lossless-claw still hasn't built an LCM graph for this session. Once more messages accumulate past the leaf trigger threshold, it'll compact and subsequent turns will use the graph path.```
sand mason
#

i dont need any of this

torn meadow
#

ok

sand mason
#

i just need /status

#

i think i know the issue, just seeing if your setup is what i expect

torn meadow
#

it's odd too that /status doesn't show my primary model anymore... 🤷‍♂️

sand mason
#

not sure where you got that output from

torn meadow
#

openclaw status --all

pallid wind
#

5.4 just released

torn meadow
sand mason
#

@torn meadow you able to run /status inside a chat app

pallid wind
torn meadow
# 
Krill
APP
 — 10:40
🦞 OpenClaw 2026.5.3 (06d46f7)
🧠 Model: openai-codex/gpt-5.4 · 🔑 oauth (openai-codex:default)
🔄 Fallbacks: openrouter/stepfun/step-3.5-flash, openrouter/z-ai/glm-5.1, google/gemini-flash-latest
🧮 Tokens: 15k in / 20 out
🗄️ Cache: 19% hit · 3.6k cached, 0 new
📚 Context: 19k/272k (7%) · 🧹 Compactions: 0
📊 Usage: 5h 98% left ⏱4h 18m · Week 69% left ⏱16h 19m
🧵 Session: agent:krill:discord:channel:1500938660189831188 • updated 8m ago
⚙️ Execution: direct · Runtime: OpenClaw Pi Default · Think: medium · Text: low · verbose · trace · elevated
👥 Activation: mention · 🪢 Queue: steer (depth 0)```
sand mason
#

yea i know the issue

#

openai-codex/gpt-5.4 -> openai/gpt-5.4
but you also have many models so it goes through then all till it lands

#

this is a hangover from another change, i replicated the issue today. will get this fixed tomorrow

torn meadow
#

it's also odd that /status in discord shows i am on 2026.5.3 but in terminal status show 2026.5.4-beta.1

#

even after i refreshed the context with /new

torn meadow
sand mason
#

this one took me ageees to fine

#

find*

#

but was able to reproduce few hours ago

torn meadow
#

i also tried to set up codex harness with openai/gpt-5.4 but the doc is very confusing, i think it's also incorrect.

  • refers to 'codex' plugin which is supposed to be bundled, but actually is not
  • when i try to install 'codex' plugin from clawhub, gateway blocked it saying it is dangerous 😓
sand mason
#

hence why not in today but will be in tomorrow

sand mason
sand mason
torn meadow
#

hopefully it will be bundled too

#

thank you spaceman, appreciate all your effort

sly pasture
#

I’m facing the same issue. My CPU usage keeps hitting 100%, causing my agents to respond very slowly. Initially, I thought the problem was with my server or the AI model I was using. Has OpenClaw already fixed this issue?

gentle nacelle
#

Hey VincentK, I am slightly lost with the 100% CPU, the non responsive gateway, and the delayed in response gateway. My latest test was at '2026.5.3-1'. Despite the improvements, there was an annoying >6 second delay, between openclaw and LM Studio logs, or openai-codex model response. Have you spotted the causal factor?
I know that the whole issue is tough and work demanding.

gentle nacelle
#

I have updated to the latest 2026.5.4 and post the results.

gentle nacelle
#

2026.5.4 , is a really good version. My suggestion is to update from 4.23, with the drawback of slight lag.

sand mason
#

5.5 shipped the fix for the oauth model fallback causing huge delays

#

doctor fix and it should be done

#

this for people on model in /status where it shows "openai-codex" as the model

#

huge bunch of fixes to how we read files and memory issues in many other spots too

sand mason
#

and run doctor --fix

sand mason
pallid wind
#

Here is the patch I made with codex to ensure the openclaw updates don’t crash the gateway and agent continuity persists;

The patch was to disable the bad local one-shot updater instead of letting it keep
rerunning.

Changed: .openclaw/update-backups/update-55-one-shot.completed.20260506-060201.sh

What it now does:

  • Exits immediately with 0.
  • Logs that OpenClaw 2026.5.5 is already installed.
  • Explicitly does not run a package update.
  • Explicitly does not restart or stop the gateway.
  • Explicitly does not send another Telegram completion message.

Why this matters: the earlier launchctl-submitted updater had keepalive-style
behavior, so even though OpenClaw was already updated, it kept rerunning the update
path. Each rerun stopped/restarted the managed gateway, which created the outage/
debug loop and repeated Telegram notices.

Verification:

  • The manual update LaunchAgent is no longer loaded.
  • openclaw status reports OpenClaw 2026.5.5 up to date.
  • Gateway is loaded, running, and reachable at ws://127.0.0.1:18789.
  • openclaw gateway probe succeeds.

Welcome review @sand mason and thx for all your hard work G!

fiery bolt
#

I have a problem. Previously, Web_fetch had an SSRF blocking issue with fake IPs in TUN proxy mode, but it worked fine after enabling the allowSSRF configuration. However, I found that the Firecrawl plugin also blocks fake IP network segments internally, and I can't find a direct configuration option to allow fake IP network segments. Currently, I can only whitelist specific URLs within the proxy software.

nocturne furnace
gentle nacelle
#

I am on it. Its stable. No issues during installation.

Had gpt5.5 fix the error of missing openai-codex from models.

Took it a while, but fixed. Roughly the same minor delay between the two versions.

nocturne furnace
gentle nacelle
#

You just could not choose it. Only in API.

nocturne furnace
#

I see, maybe I should wait a few more days for things to settle then. .23 is still working okay for me.

I am a bit cautious after the whole updating debacle and am not in the mood to spend another day trying to fix things.

I've read discord is still a bit brittle, and there are still some plugins bug. I have sort of weak hardware so it's extra sensitive.

pallid wind
nocturne furnace
shy trellis
gentle nacelle
#

@nocturne furnace Take your time, grab a coffee. Update issues are fixable. Reconfigure what ever does not run. Past version files may be left, so is config. try to run "openclaw doctor --fix" before and after each update. Try using the wizard, for services or channels that do not run. You can do it.

pallid wind
fiery gale
#

For Me 2026.5.4 fixed the slowness

nocturne furnace
pallid wind
sly pasture
nocturne furnace