#Agent failed - unknown model: anthropic/claude-opus-4-6
1 messages · Page 1 of 1 (latest)
That error means your gateway is receiving the message, but the configured model id can’t be resolved (anthropic/claude-opus-4-6 isn’t a supported/known model name in your current setup), so the agent aborts before it can reply.
Good news: slash commands bypass the model, so you can usually recover without SSH.
- In the same Discord chat, run:
/model list(or just/model)
- Then pick one from the numbered list:
/model 1(etc)
If you instead get “Model … is not allowed”, then you’ve enabled a model allowlist (agents.defaults.models) and it currently doesn’t include any working model. In that case:
- If you previously enabled
/config(requirescommands.config: true), you can try:/config unset agents.defaults.models(removes the allowlist), then rerun/model list
- If
/configis not enabled and you truly can’t SSH/iCloud/tailscale in, you’ll need someone/something to regain access to the Mac mini to edit~/.openclaw/openclaw.jsonand restart the gateway.
Can you paste:
- what
/model listreturns, and
- what your SSH failure is (timeout vs “connection refused” vs auth failure)?
Docs refs:
- Unknown model troubleshooting: https://docs.openclaw.ai/gateway/troubleshooting (search for “Unknown model”)
/modeldirective: https://docs.openclaw.ai/tools/slash-commands#model-selection-model
Darn, I'm screwed since I didn't enable config, on vacation.
You’re probably not screwed — /config being disabled only means you can’t persist changes to disk from chat. But /model and /reset don’t require /config and don’t go through the LLM, so they can still recover you even while you’re away.
Try this in Discord:
- Clear any bad per-session model override
/reset(or/new)- This often drops you back to the default model if the “4-6” thing was only a session override.
- Force-pick a working model (no config needed)
/model list(or just/model)- then select a number:
/model 1
If /model list shows only that broken anthropic/claude-opus-4-6 entry (or everything else says “not allowed”), then yeah: you’ve effectively locked the bot into an invalid model via the model allowlist (agents.defaults.models), and without SSH / someone local / some other remote access (Screen Sharing/Tailscale/etc.) you can’t change that on disk.
If you paste what you see from:
/model list- and whether
/resetchanges anything
…I can tell you which situation you’re in (bad session override vs broken on-disk allowlist) and the minimal fix for when you regain access.
why isn't anthropic/claude-opus-4-6 a working model?
Because, as far as OpenClaw can tell, that exact model id doesn’t exist / isn’t supported, so it fails model resolution before it can run the agent.
A few common reasons this happens:
- The name is simply wrong. OpenClaw expects model ids to match what it knows for that provider. If your config (or a session override) says
anthropic/claude-opus-4-6but OpenClaw’s supported catalog doesn’t include it, you getunknown model. - “4.6” is getting mixed up. People often say “I upgraded to 4.6” meaning something else (a bot version / a model family rumor / a UI label), but the actual Anthropic model id OpenClaw needs is very specific.
- Less likely here: OpenClaw also intentionally rejects some older/insecure model names, but
opus-4-6looks more like “not a real/known id” than “blocked legacy id”.
What to do:
- Run
/model list(or just/model) and pick one of the models it shows. That list is the source of truth for your gateway config. - If you want the “most current Opus name” for Anthropic, ask in #models (the exact string changes over time and I don’t want to guess), but
/model listwill always show what your instance can actually use right now.
Docs: https://docs.openclaw.ai/gateway/troubleshooting (section “Agent failed… Unknown model”)
[clawdbot] Providers:
- anthropic (1)
Use: /models <provider>
Switch: /model <provider/model>
Perfect — that response means the bot is still alive and commands are working (they bypass the LLM), it just can’t run the current configured model.
Next step:
- List the actual Anthropic models your gateway thinks are available/allowed:
/models anthropic
- It should print a list. Then switch to one of the listed
provider/modelentries:
/model <paste-exact-provider/model-from-the-list>- (or if it shows numbers)
/model 1
- (or if it shows numbers)
If /models anthropic only shows anthropic/claude-opus-4-6 (or it errors), then your on-disk model allowlist/catalog is effectively pointing at a model id OpenClaw doesn’t recognize — and without SSH/physical access you won’t be able to fix the config file. But in most cases you’ll see other Anthropic entries there and you can just pick one.
Paste what /models anthropic returns and I’ll tell you exactly what to type for the switch.
[clawdbot] Models (anthropic) — showing 1-1 of 1 (page 1/1)
- anthropic/claude-opus-4-6
Switch: /model <provider/model>
All: /models anthropic all
That output is the key clue: your bot’s configured model catalog/allowlist contains exactly one entry, and it’s anthropic/claude-opus-4-6 — but OpenClaw doesn’t recognize that exact model id, so every normal message fails with unknown model.
So /models anthropic isn’t showing “what Anthropic offers”, it’s showing “what your config currently allows”, which right now is only the broken one.
The only ways out (while you’re away)
A) If you have /debug enabled (sometimes people do):
Try:
/debug unset agents.defaults.models
Then run:
/model list
…and pick a model from the picker.
B) If you have bash commands enabled (commands.bash: true + elevated allowlist):
Run:
/commands(see if! <cmd>or/bashis listed)
If yes, you can fix it entirely from Discord by running on the Mac mini:
! openclaw models list! openclaw models set <pick-one-from-the-list>! openclaw gateway restart
C) If neither /debug nor ! is enabled:
Then yeah—there’s no remote-write path left from Discord alone. You’ll need some way back onto the Mac mini (SSH/Tailscale/Screen Sharing, or someone local) to edit ~/.openclaw/openclaw.json and restart.
If you run /commands and paste what it shows, I can tell you immediately whether A or B is possible on your setup.
can I request that openclaw recognize this model id? please, hail mary
You can request it, but it won’t be a “hail mary” that fixes your currently-dead bot while you’re away.
What’s happening is: OpenClaw has a strict model-id resolver. If a provider/model string isn’t in the supported catalog (or can’t be resolved via provider config), it fails fast with unknown model rather than trying to send unknown ids upstream.
What you can do
- File a GitHub issue/feature request asking to add/support (or alias)
anthropic/claude-opus-4-6:
https://github.com/openclaw/openclaw/issues
Include:
- the exact error line (
unknown model: anthropic/claude-opus-4-6) - your
openclaw --version - your
agents.defaults.modelssnippet (the allowlist that currently only contains that one entry) - whether you expect that model id to exist on Anthropic (link/source if you have one)
- But to get unblocked right now, you still need some way to change your config/allowed models (SSH/local access, or
/debug/!if you had those enabled). Adding support in OpenClaw would require a code change + you upgrading again anyway.
If you paste your /commands output, I can tell you if there’s any remote-only escape hatch enabled on your setup (like /debug or ! openclaw …).
I’m back at my mac mini with clawdbot on it, how do i get the 4-6 opus model issue fixed?