Provide an in-process read-only MCP tool

Goal

Let the agent call a controlled application function without launching a separate MCP process.

Steps

1. Define a function with @tool(name, description, input_schema), validate input, and bind it to a read-only data source. [mcp]
2. Create the server with create_sdk_mcp_server(name=..., version=..., tools=[...]) and register it in ClaudeAgentOptions(mcp_servers={...}). [mcp]
3. Add only the mcp__<server>__<tool> name to allowed_tools, test with synthetic data, and inspect ToolUseBlock/ToolResultBlock. [mcp]

How to verify

The agent can call only the intended tool; invalid arguments are rejected and errors set `is_error`.

Troubleshooting

If the tool is missing, check the server key, tool name, and full allowed_tools name; return content text blocks for results.