Hosted Letta MCP server

Let your other agents work with your Letta agents. Any MCP client, such as Claude Code, Cursor, or another agent, can connect to https://api.letta.com/mcp to create stateful Letta agents, delegate work to them, and read their replies. Letta agents keep their memory across every conversation, so the calling agent does not need to resend context.

To give a Letta agent access to external MCP tools instead, see MCP tools in the Agent SDK.

Connect a client

Create an API key in Letta Platform and send it as a bearer token:

Cursor: .cursor/mcp.json

{
  "mcpServers": {
    "letta": {
      "url": "https://api.letta.com/mcp",
      "headers": { "Authorization": "Bearer <LETTA_API_KEY>" }
    }
  }
}

VS Code: .vscode/mcp.json

{
  "inputs": [
    {
      "id": "letta-api-key",
      "type": "promptString",
      "description": "Letta API key",
      "password": true
    }
  ],
  "servers": {
    "letta": {
      "type": "http",
      "url": "https://api.letta.com/mcp",
      "headers": {
        "Authorization": "Bearer ${input:letta-api-key}"
      }
    }
  }
}

The client must support custom bearer headers; OAuth-only clients are not supported yet.

Tools

Tool	Purpose
list_agents	Find agents by name, tags, or query.
list_models	List models available to your project.
create_agent	Create a stateful agent with its own memory.
send_agent_message	Message an agent and optionally wait for its reply.
get_run	Check the status of a message’s run.
get_reply	Fetch the final reply to a sent message.

Full input schemas are available from tools/list.

Agents and conversations

An agent keeps its memory across conversations; each conversation has its own message history. When calling send_agent_message:

• Pass agent_id to start a new conversation.
• Pass a conv-... conversation_id to continue one.
• Pass agent_id with conversation_id: "default" for the agent’s default conversation.

Messages run on a Cloud computer in unrestricted permission mode: the agent executes tools without approval prompts.

Replies

A call waits up to 110 seconds for a reply. Longer work keeps running; use get_reply with the returned IDs rather than resending.

Status	Meaning
completed	The reply is included.
queued	Accepted; the call did not wait.
wait_failed	Accepted, but waiting timed out. Call get_reply.
acceptance_unknown	The message may have been accepted. Check the conversation before resending.
submission_failed	Not accepted.