HTTP compatibility

Boundary statement

Compatibility is an adapter contract, not provider ownership or total upstream parity. All three HTTP entrances resolve a configured model, call the shared UAR provider/runtime path, and return the vocabulary selected by the client-facing adapter.

Routes

Adapter	Route	Principal behavior
UAR chat	POST /api/chat/completion	OpenAI-shaped request with UAR session and stream extensions
OpenAI compatibility	POST /v1/chat/completions	same chat handler under the OpenAI route
OpenAI model discovery	GET /v1/models, GET /v1/models/{model_id}	configured-provider model views
Anthropic compatibility	POST /v1/messages	Anthropic message/tool input translated into UAR execution and translated back

The legacy /api/chat and /api/sessions families are disabled routes. New clients use the routes above or the governed run API.

Model addressing

A bare model name resolves against the configured default provider. A provider/model value selects the provider explicitly. Unknown provider or model references fail rather than silently selecting a different provider. OpenAI-shaped input therefore does not mean the request uses OpenAI.

Provider configuration and credentials are separate. See Configure providers and Credentials.

Authentication and sessions

When JWT authentication is required, protected calls carry Authorization: Bearer <token>. UAR creates trusted user and tenant context only after verification. Health and discovery exceptions are documented in Authentication.

For chat, X-UAR-Session-ID is the preferred continuity header. The body and cookie forms are also accepted by the current handler. An omitted id creates a new session; an invalid non-UUID session id fails with a client error. Returning an id to the next request is what preserves conversation continuity.

Streaming selection

stream: false returns one adapter-shaped response. stream: true returns SSE. The UAR/OpenAI chat request also accepts:

• openai — OpenAI chunks, the default;
• agui_spec — the uar.agui/1 profile using official AG-UI event names;
• agui — deprecated legacy agui.* events;
• dual — OpenAI chunks plus the current AG-UI mapping.

Tool results, activated skills, and context actions can appear as documented UAR extensions in OpenAI chunks. Clients that require only upstream fields must ignore unknown extension members.

Anthropic input supports the message and tool forms implemented by the current adapter, including its translated streaming output and prompt-cache markers. It still uses UAR routing and does not bypass runtime governance.

Errors

Bad session identifiers and malformed requests return client errors. Unknown models return 404. Missing or invalid required credentials return 401. Provider startup, timeout, and stream failures remain distinguishable from UAR validation or governance outcomes; consult the returned adapter error and runtime logs together.

Compatibility limit

These routes implement the documented subset in current source. They do not promise every OpenAI or Anthropic beta header, object field, transport, error code, or future extension. Generated OpenAPI is a summary, not a conformance certificate. Verify the exact client workflow against the deployed version.

Profile limits

minimal and server-full include the HTTP adapters. server-full adds the broader governed release composition, but the wire shape alone does not prove those features ran. embedded-mobile exposes no HTTP routes; its host calls shared services directly.

Next: Events, AG-UI, and A2UI.