GET /v1/schemas/{name} returns the full JSON Schema of one of five request types. If you arrived here because a tool argument told you to describe_schema('AgentConfig') before writing it, that stub is the whole reason this endpoint exists — fetch the document, then write the object.
Two fields: the name you asked for, and the schema. Any other name is a 422.

​
The five

They compose, so a document you fetch may itself defer to another. Fetching AgentTeam gives you a 2.3k-character envelope whose members point at AgentConfig and AgentOverride rather than a document with two copies of the agent tree inside it.

​
Why they are fetched rather than inlined

Five of this API’s request types are not fields, they are whole product surfaces. An agent’s config alone renders as roughly 34,000 characters of JSON Schema. That is fine over HTTP, where a request body is a request body. It is not fine for a tool-calling surface. MCP has no way for one tool to share a definition with another, so every operation that accepts an agent config would have to carry the entire tree inside its own argument schema — create_agent, update_agent, create_outbound_call, calls_token, create_text_conversation, create_call_batch and the members of a team, each with a private copy. That was most of a tool surface that no longer fitted in a 200k-context client. So on the agent-facing surface those arguments carry a placeholder — an object whose description ends in describe_schema('AgentConfig') — and this endpoint serves the real thing on demand. You pay for the big document once, in the turn that is actually writing an agent, instead of in every turn.
Only the agent-facing view defers. The OpenAPI document carries every type in full, both generated SDKs carry the full types, and request validation on the server is unchanged. A config you send is checked against the same model whether you read its shape from here, from the SDK, or from the endpoint reference.
That is why this page matters mostly to two readers: someone driving the MCP server from a coding agent, and a CoPilot writing a config in the dashboard. If you are writing TypeScript or Python, you already have these types.

​
Fetching one

The name argument is a closed set in both SDKs — a typo is a type error before it is a 422.

​
Where each one is accepted

For what the fields inside those documents mean, rather than what shape they have, read the agent config, per-call configuration, agent teams on a call, task inputs and output and the operation tree. A JSON Schema tells you what will be accepted; those pages tell you what to put in it.