Endpoint Index

Every route the proxy serves. X in the auth column means "API key (Authorization: Bearer sk-… or x-api-key: sk-…)"; J means console JWT (Authorization: Bearer <jwt>); — means public.

LLM endpoints

Method	Path	Auth	Protocol	Notes
POST	/v1/chat/completions	X	OpenAI	Chat Completions
POST	/v1/responses	X	Responses	JSON or SSE; background: true requires Redis
POST	/v1/responses/compact	X	Responses	Native passthrough first, local lossless packing fallback
GET	/v1/responses/{response_id}	X	Responses	Retrieve a stored response (24 h TTL)
DELETE	/v1/responses/{response_id}	X	Responses	Delete a stored response
POST	/v1/responses/{id}/cancel	X	Responses	Cancel a background response (409 if not cancellable)
GET	/v1/responses/{id}/input_items	X	Responses	limit 1–100 (default 20), order asc/desc, after cursor
POST	/v1/messages	X	Anthropic	Anthropic Messages
POST	/v1/messages/count_tokens	X	Anthropic	Native upstream count when available, otherwise a local o200k_base estimate
POST	/v1/embeddings	X	OpenAI	Embeddings
POST	/v1/systemone	X	System One	TypeSafe Jev evaluation (also via OpenRouter and local Ollama models). The named model must carry the supports_systemone flag, otherwise 400 invalid_request_error
POST	/v1/decisions	X	Decisions	OpenAI's Decisions endpoint (gpt-6-luna), also served by System One models through the proxy's bridge. Gated by the same supports_systemone flag
POST	/v1/images/generations	X	OpenAI	JSON body
POST	/v1/images/edits	X	OpenAI	JSON or multipart (image, image[], mask)
POST	/v1/audio/speech	X	OpenAI	Binary audio response
POST	/v1/audio/transcriptions	X	OpenAI	Multipart; requires model and file
POST	/v1/audio/translations	X	OpenAI	Multipart; requires model and file
GET	/v1/models	X	—	Model list filtered by the key's allowlist
GET	/v1/protocols	X	—	Registered protocols with base paths

::: note Path aliases Protocol routes also answer on trailing-slash variants and on the alias forms used by different SDK versions — e.g. POST /chat/completions and POST /v1/v1/chat/completions reach the same handler as the canonical path. :::

Capability matrix

Provider support varies by capability — the table summarizes who can serve what.

Capability	Providers
Chat	All types
Embeddings	openai, the openai-compatible family (including DeepSeek, Kimi, MiniMax, Moonshot, Qwen, xAI, vLLM, SGLang, Chutes, Mistral, NanoGPT, OpenRouter), gemini, ollama
Images (generation + edits)	openai, the openai-compatible family, gemini, qwen
Audio (speech, STT)	openai, the openai-compatible family except qwen/qwen-intl (DashScope serves speech only on its native endpoints — see below), gemini (native TTS/STT), openrouter (STT)
Audio (translation)	openai, the openai-compatible family except qwen/qwen-intl, gemini. openrouter and qwen/qwen-intl have no upstream endpoint and reject the request with a 400
System One (evaluation)	typesafe (Jev, direct), openrouter (TypeSafe Jev resold), ollama (local models, v0.35+). Chat and embeddings against ollama keep using the native /api/* surface
Decisions (evaluation)	openai (native /v1/decisions), plus every System One provider above, which the proxy bridges onto a System One route. Only openai accepts inline images

The dedicated anthropic provider type is chat-only. The typesafe provider type is System One-only — assign it to models that serve /v1/systemone, and note that chat requests against such a model are rejected with a 400.

qwen/qwen-intl serve no /audio/* route: DashScope has no OpenAI-compatible audio surface. Speech synthesis (Qwen-TTS, CosyVoice) uses the native /api/v1/services/aigc/multimodal-generation/generation endpoint, and OpenAI-compatible speech recognition (qwen3-asr-flash) is called through /chat/completions with input_audio content parts. Audio requests against a qwen model are rejected with a 400 invalid_request_error naming the limitation.

Capability flags are additive: supports_systemone marks a model as a decision model, which gates both /v1/systemone and /v1/decisions — they are two envelopes over the same evaluation primitives, and the provider translates between them (see Decisions). The flag does not restrict the model's chat service, so a model can serve both when its providers support both. A request to either evaluation endpoint naming a model without the flag is rejected with a 400 invalid_request_error before any provider is called. The flag also gates being selected as the Routing Judge.

WebSocket endpoints

Path	Auth	Notes
WS /v1/responses	X	Same events as the SSE stream; send {"type":"response.create", …}
WS /v1/realtime	X	OpenAI Realtime relay; ?model= must resolve to a supports_realtime model on an openai/openai-compatible provider

MCP

Method	Path	Auth	Notes
POST	/servers/{server_name}/mcp	X	Streamable-HTTP MCP endpoint for one configured server (all MCP methods)

Health

Path	Auth	Notes
GET	/api/health	—
GET	/api/health/ready	—
GET	/api/health/live	—

Admin API

All under /api, all requiring a JWT unless noted. Groups:

Prefix	Contents
/api/auth/*	setup-status, setup, login, logout (logout accepts an optional token)
/api/api-keys*	Key CRUD, budget reset, per-key usage, spend summary
/api/config/providers*	Provider CRUD, provider types, key reveal, upstream model list
/api/config/models*	Model CRUD, pricing/metadata sync endpoints, model names
/api/config/server/*	All server settings sections (logging, web search, smart routing, provider selection, request policy, resilience, security, keepalive, rate limits, CORS, MCP security)
/api/config/circuit-breaker*	Breaker state and resets
/api/logs*	Log query/detail/stats, usage stats, audit verification, cleanup
/api/routing/*	Training-data export: feedback labels and judge verdicts as JSONL (/api/routing/export/dataset)
/api/mcp/*	MCP server CRUD, status, capabilities
/api/team/members*	Member CRUD, roles, model allowlists, account budgets, password resets
/api/me/*	Profile, budget, password, username, tracing, feedback
/api/catalog/models	Display catalog (any authenticated user)
/api/system/info	Version and update check (admin)

Static frontend

When frontend/dist exists, /assets/* serves compiled assets and any other GET path that is not /api/*, /v1/*, or /servers/* falls back to index.html (SPA routing). Without the build, those paths 404 and only the API is served.

Not available

• OpenAPI / Swagger / ReDoc at runtime — not registered (docs_url=None, redoc_url=None, openapi_url=None), so /docs, /redoc, and /openapi.json are never mounted. A GET to those paths therefore follows the SPA fallback from the Static frontend section: index.html (200) when frontend/dist exists, 404 without it. The generated /v1 contract is published as OpenAPI Specification inside this documentation instead.
• Log export endpoints — query GET /api/logs with pagination instead.

Related

• Authentication — headers and credentials
• Errors & Rate Limits — failure shapes