Gateway planes

Every plane is served from https://intertrace.fly.dev and authenticated with your itr_rt_* runtime key. Your organization is taken from the key, never from headers the client sends. GET /v1/models reports which planes and routes are live.

Send A2A JSON-RPC calls to /a2a/<peer>. The gateway inspects message/send content, decides each call for the peer it reaches, and keeps an A2A conversation as one session: the first message starts a session that becomes the task's context id. Tasks are scoped to your organization. A refusal comes back as a JSON-RPC error with the call's id and a reason_code. Peer discovery is at /a2a/_discovery.

Route calls to your HTTP services through /service/*. Each organization keeps its own service registry, with approved and blocked services and a monitor or enforce mode, so agents reach only the services you allow.

OpenAI-compatible inference calls go to /inference/v1/..., for example /inference/v1/chat/completions, with the same checks as the LLM plane. The default backend is OpenAI. Where multiple model pools are configured, the x-intertrace-inference-pool header keeps a client on one pool.