Endpoints

hlid listens on bind_addr (default 127.0.0.1:8080) and exposes:

Inference

Method	Path	Dialect
POST	/v1/messages	Anthropic Messages
POST	/v1/messages/count_tokens	Anthropic token count
POST	/v1/chat/completions	OpenAI Chat Completions
POST	/v1/responses	OpenAI Responses
POST	/v1/completions	OpenAI Completions (legacy)

The path tells hlid what dialect the agent speaks. The model in the body tells it where to go. Streaming is requested the way each dialect already does it ("stream": true).

Discovery

Method	Path	Purpose
GET	/v1/models	Model list

Observability

Method	Path	Purpose
GET	/hlid/requests	All recent request captures
GET	/hlid/requests/<id>	A single capture
GET	/health	Health check — {"status": "ok"}

Auth

When auth_key is configured, every endpoint requires it, sent either way:

Authorization: Bearer <key>
x-api-key: <key>

Failures return a 401 in the dialect of the endpoint — an OpenAI-shaped error body on OpenAI paths, an Anthropic-shaped one on Anthropic paths — so SDKs surface the error cleanly instead of choking on an unexpected shape.

Errors

All errors are dialect-native for the endpoint that received the request:

• Unknown model (no registry match) → 404-family error naming the model
• Upstream failure → the backend's status and message, passed through
• Timeout → 504-family error after request_timeout_secs