run

Start the server synchronously.

Once running, the host exposes POST /invocations. The default request body is:

.. code-block:: json

{"message": "Hello!", "stream": false}

message is either user text or a list containing a structured HITL response item. stream is optional and defaults to false. Non-streaming requests return JSON:

.. code-block:: json

{"response": "Assistant text"}

A pending LangGraph interrupt adds Responses-style function_call and mcp_approval_request items under output. Resume it by sending the matching function_call_output or mcp_approval_response as the next request's message list.

Streaming requests return text/event-stream with token payloads:

.. code-block:: text

data: {"token": "..."}

event: done
data: {}

With options.resilient_background=True, a request containing {"message": "Hello!", "background": true} returns 202. Retrieve or cancel it through GET /invocations/{invocation_id} and POST /invocations/{invocation_id}/cancel respectively.

Multi-turn callers should reuse the x-agent-session-id response header as the next request's agent_session_id query parameter.