Skip to main content

Recipe: keep working when Claude is down

Have the local model stand in when Anthropic cannot answer: an outage, an overload, no network, or a rate limit. On a subscription, running out of usage arrives as a rate limit, so this covers that too.

Set it up​

You need a local model configured, as in the subagents recipe. Then name it as the stand-in on a rule that sends requests upstream:

"routing": [
  { "name": "everything", "route": "upstream", "ifDown": "janus-local" }
],
"features": { "route": true, "escalate": false, "compress": false, "guard": false }

To move some work local before the limit is reached, see pacing.

A rule with no class matches every request. Put it last if you have other rules; the first rule that matches decides.

What happens​

Each request goes to Claude as usual. If Claude says it is out of usage (a 429 that won't clear for a minute or more), the same request is sent to the local model at once. A shorter failure (an overload, a server error, no connection, a brief rate limit) is passed to Claude Code to retry; only if Claude fails three times within two minutes does the local model step in. Either way the local model's answer is returned. The response carries an X-Janus-Fallback header, the console line shows that Claude was tried first and why it failed, the ledger records it, and janus stats lists when it happened, under "When Claude could not answer":

When Claude could not answer
  from              to     cause         requests  stood in  limit      answering again
  2026-10-14 09:54  11:24  rate_limited  61        60        five_hour  11:30

Each row is a run of requests Claude did not answer, none more than 15 minutes after the one before. stood in is how many the local model answered; the rest reached Claude Code as errors, to retry. For a rate limit, limit is which of the subscription's limits was reached (five_hour, seven_day) and answering again is when Claude said it would answer again.

Errors about the request itself, such as an expired login or a bad request, are passed to Claude Code untouched. The local model would not fix them.

What to expect​

  • The answers are the local model's. It is standing in for Claude on work that was meant for Claude, including the main conversation. Expect slower and weaker; the aim is to keep going, not to match.

  • Claude Code is not told. It sees a normal response naming the model it asked for. Watch Janus's console to know which you are talking to.

  • Requests too large for the local model fail with Claude's original error, since there is nowhere else to send them. A long main conversation can outgrow a local context. A model LM Studio has not loaded is only sent requests up to 32,768 tokens unless contextTokens says otherwise. The console line's ↳ no local stand-in: ... and the ledger's reason say why the local model did not answer.

  • It only works before an answer starts. If Claude fails part-way through a streamed answer, what was already sent cannot be taken back.

  • Picking up again is safe. When Claude is back, the conversation goes upstream with the local model's turns in it. Janus removes the one thing Anthropic would reject, the local model's unsigned thinking.

  • Janus says when Claude takes over again. Claude reads the local model's turns as its own, so the first answer it gives the main conversation afterwards gets a note:

    ↳ Claude is answering again, after 12 turns from the local model since 09:54; it reads them as its own, so check what they did

    The ledger marks that answer with handback, and janus stats counts them. It is a good moment to ask Claude to review what was done in the meantime.

Trying it​

Point upstream.baseUrl at a port with nothing on it, restart Janus, and use Claude Code. Each request should be answered locally, with the console line showing the failed attempt on Claude first. Put the URL back afterwards.

With a local-only policy​

Content the local-only policy protects is never sent upstream by a fallback in the other direction: when LM Studio is the one that is down, a request that must stay local is refused rather than sent to Claude.

The reference has the details.