Connect an MCP client

AXL serves an MCP endpoint at /mcp. Clients can discover agents, run them, inspect tasks, and - when enabled per agent - call an agent through a dedicated tool.

Expose an agent as a tool

Add this to the agent's agent.toml:

[interop]
mcp_tool = true
public_description = "Answers partner support questions."

The MCP tool name is derived from the agent key. An agent with key = "support" appears as agent_support.

Connect

Point an MCP-compatible client at:

https://axl.example.com/mcp

Use the authentication credentials issued by the AXL operator. After connecting, list tools and resources rather than assuming a fixed catalog; the published agents depend on deployment configuration.

Choose the right operation

• Call an agent_<key> tool for the simplest way to send an agent a task.
• Use the general agent operations when the client needs to select an agent at runtime.
• Use task operations when the client needs to list, inspect, or cancel asynchronous work.
• Hosts that declare the io.modelcontextprotocol/tasks extension get a native task back from run operations instead of a {task_id, session_id} payload, and drive it with tasks/get / tasks/cancel - no tool-based polling needed.
• Hosts that also declare form elicitation see input_required when a run parks on a question or an approval, and answer it with tasks/update - the host becomes a human-in-the-loop surface for the runs it started.
• Read agent information resources when building a discovery or selection UI.

For protocol method names and schemas, use the HTTP and protocol reference.

Next

• Expose an agent - the interop block in full.
• Connect an MCP server - the opposite direction, AXL as client.
• MCP server reference - methods and schemas.