Reference

HTTP endpoints

Every path pooled serve answers, and every status code it returns, in one place.

pooled serve listens on http://127.0.0.1:8080 by default. The OpenAI base URL is http://127.0.0.1:8080/v1; the Anthropic base URL is http://127.0.0.1:8080. For setup, see Serve API.

Paths

Method and path	API	What it does
`POST /v1/chat/completions`	OpenAI	Chat Completions, with tools and JSON output, streaming or not.
`POST /v1/responses`	OpenAI	Responses, with function and custom tools, `previous_response_id` and `text.format`, streaming or not.
`GET /v1/responses/{id}`	OpenAI	A response this process stored.
`DELETE /v1/responses/{id}`	OpenAI	Deletes a stored response.
`GET /v1/responses/{id}/input_items`	OpenAI	The input items of a stored response.
`POST /v1/messages`	Anthropic	Messages, with tools, `tool_choice` and thinking, streaming or not.
`GET /v1/models`	both	The room’s model as `pooled/<model>`, with `max_model_len` set to the room’s context. Empty until the model is ready.
`GET /v1/models/{id}`	both	One model. `404` for any id but the room’s.
`GET /health`		Status for scripts. Never needs the token.
`GET /`		The banner, as plain text.

/v1/models answers in Anthropic’s shape when the request has an anthropic-version header, or an x-api-key header without Authorization. Otherwise it answers in OpenAI’s shape.

Stored responses live in memory only: at most 256 of them or 64 MB, each dropped an hour after its last use.

Not served

These return 404 with a message that says so:

POST /v1/completions (the legacy API; use /v1/chat/completions)
POST /v1/embeddings
POST /v1/messages/count_tokens
POST /v1/responses/input_tokens
POST /v1/responses/compact

A known path with the wrong method gets 405 with an Allow header.

/health

curl http://127.0.0.1:8080/health

{
  "ok": true,
  "room": "4TKG9P",
  "connected": true,
  "ready": true,
  "model": "pooled/qwen3.6-35b-moe",
  "ctx": 32768,
  "queue": 0,
  "served": 12
}

Field	Meaning
`connected`	A link to the host is open and it said hello.
`ready`	The room’s model is loaded and can answer.
`model`	The id `/v1/models` lists, or `null`.
`ctx`	The room’s context in tokens, or `null` if the host did not say.
`queue`	Requests waiting or running, counted here.
`served`	Requests answered since start.
`closed`	Present only when the host disconnected this client or turned API clients off: the reason.

With a token set, a caller without it gets only {"ok": true}.

Status codes

Errors come in each API’s own shape: {"error": {...}} for OpenAI paths and {"type": "error", "error": {...}} for Anthropic paths.

Status (OpenAI / Anthropic)	When	Retry
`400` / `400`	Bad request, an unsupported feature, or a prompt over the room’s context (OpenAI code `context_length_exceeded`).	Fix the request.
`401` / `401`	Missing or wrong key, when the server has a token.
`403` / `403`	A `Host` other than `127.0.0.1:<port>` or `localhost:<port>`, or any `Origin` header.
`404` / `404`	Unknown path, a path listed above as not served, or an unknown model id.
`405` / `405`	Known path, wrong method. See the `Allow` header.
`413` / `413`	Body over 4 MiB, or the request is over 3.5 MB once packed for the room.	Send less.
`429` / `529`	The queue here (`--max-queue`) or in the room is full. Has `Retry-After: 5`.	Yes, after `Retry-After`.
`503` / `529`	Model not ready, host gone, this client disconnected by the host, or the host pressed Stop. While not connected or not ready, the response has `Retry-After: 5`.	Yes, after `Retry-After`.
`504` / `504`	The room sent nothing for this request for 5 minutes, or 30 minutes while the host had it queued.	Yes.
`500` / `500`	Generation failed in the room, or the room broke the protocol.	Yes.

Once a stream has started, an error arrives as an error chunk (OpenAI, with no [DONE] after it) or an error event (Anthropic).

Limits

Request size, queue and timeout limits are on Limits and context.