How it works
When an MCP client connects to Agent Manager, it receives a list of available tools. Each registered agent appears as a callable tool namedrun_{agentId}. Clients invoke agents using standard JSON-RPC 2.0 calls over an SSE connection.
Connecting an MCP client
1
Obtain a Bearer token
Authenticate via
POST /api/auth/login and copy the returned token value. See Authenticating Requests for details.2
Configure your MCP client
Point your client at the Agent Manager SSE endpoint and include your token in the headers.Claude Desktop (Cursor (MCP server settings):
claude_desktop_config.json):3
Verify the connection
After saving the configuration, your MCP client should show Agent Manager in its server list. You’ll see tools like
list_agents and run_finance_agent (one run_ tool per registered agent).MCP endpoints
Handshake
When a client connects toGET /mcp/sse, Agent Manager sends an endpoint event containing the session-scoped message URL:
POST /mcp/messages calls within the session.
JSON-RPC methods
Agent Manager handles the following MCP JSON-RPC methods:Available tools
Once connected, your client sees two categories of tools:list_agents
list_agents
A system tool that returns the list of agents registered in Agent Manager. Useful for discovery.
run_{agentId}
run_{agentId}
One tool per registered agent. Each tool accepts a
message parameter — the prompt to send to the agent — and returns the agent’s response as text.Authentication for MCP
MCP connections follow the same authentication model as the REST API. Include yourAuthorization: Bearer {token} header in the MCP client’s header configuration. Connections without a valid token are rejected.
Protocol version
Agent Manager implements MCP protocol version0.1.0. It declares tools as its capability set — each registered agent is exposed as a tool that any compliant client can discover and invoke.