Errors

Every /v1/* gateway error, chat completions, embeddings, threads, uses the same envelope:

{
  "error": {
    "message": "Model gpt-4 is not served on this gateway",
    "type": "invalid_request_error",
    "code": "model_not_allowed"
  }
}

type is "invalid_request_error" for anything the caller can fix by changing the request, and "server_error" for upstream failures.

Codes

Code	Status	When
`missing_key`	401	No `Authorization` header
`bad_key`	401	Key is unknown, revoked, or its project is inactive
`model_not_allowed`	400	Model isn’t served by this gateway at all
`model_not_allowed`	403	Model is served, but this key’s `allowed_models` doesn’t include it
`empty_messages`	400	`messages` is an empty array
`empty_input`	400	`/v1/embeddings` `input` resolves to an empty list
`thread_messages_invalid`	400	With `thread_id`, `messages` wasn’t exactly one `user` message plus any `system` messages
`thread_multimodal_not_supported`	400	Image content sent alongside a `thread_id`
`context_length_exceeded`	400	Prompt (or system block, or a single embeddings input) exceeds the model’s context window
`thread_not_found`	404	Thread doesn’t exist, or belongs to another project
`nothing_to_continue`	409	The thread’s last assistant turn already finished normally
`rate_limit_exceeded`	429	This key’s requests-per-minute bucket is full
`budget_exceeded`	429	This project is at or over its monthly token budget
`upstream_error`	502 (or the upstream’s own status, e.g. 503 with no reachable upstream)	Ollama or OpenRouter failed or was unreachable

See Usage, rate limits & budgets for exactly what the two 429s mean in practice, and Local vs. cloud routing for why a model id might land in the 400 vs. the 403 case above.

The admin/console surface

/keys and /projects (minting and revoking API keys, managing projects) are a separate, console-only, cookie-authenticated surface. They return plain FastAPI {"detail": "..."} errors, not the envelope above, see Authentication & API keys.