There is documentation on how the agent can be launched as an MCP server over stdio/SSE/HTTP, but there is no documentation on how to connect with the agent and create a session.
For the HTTP/SSE server (I have not worked with the stdio server).
By experimenting with an agent and the MCP Inspector, I figured out that you need to send 2 POST requests with specific payloads and wait for a 200 OK followed by a 202 Accepted, and forward the requests/responses through a proxy server (or a backend server, in the case of an application).
This is standard to most MCP servers, but there may be some benefit in writing down some documentation on how to create a session to obtain a session ID, and then send a tool call to the agent.
If it makes sense to add documentation on this, I can work on a PR for this.