add-mcp-tool

Expose one or more existing @task methods as Model Context Protocol tools. Once enabled, AI assistants (Claude Desktop, Claude Code, Cursor) can discover and call your tasks directly.

How MCP Works in the SDK

When ENABLE_MCP=true, the handler mounts an MCP server at /mcp. At startup, it scans the TaskRegistry for methods decorated with @mcp_tool and registers them as tools. Each tool's parameters are derived from the task's Input Pydantic model — fields are automatically converted to JSON Schema.

Steps

1. Install the mcp extra

uv add "atlan-application-sdk[mcp]"

Or if the project uses dependency groups:

uv add --group mcp "atlan-application-sdk[mcp]"

Verify pyproject.toml has the mcp dependency.

2. Add @mcp_tool to the chosen task methods

from application_sdk.server.mcp.decorators import mcp_tool
from application_sdk.app import App, task

class MyConnector(App):
    @task(timeout_seconds=3600)
    @mcp_tool(
        name="fetch_metadata",
        description="Fetch database schemas and tables from the connected source",
    )
    async def fetch_metadata(self, input: FetchInput) -> FetchOutput:
        ...

Order matters: @task must be the outer decorator, @mcp_tool must be inner (closer to the function).

Decorator parameters:

• name (optional): Tool name shown to AI clients. Defaults to the function name.
• description (optional): Tool description. Defaults to the docstring.
• visible (bool, default True): Set False to register but hide from AI discovery.

3. Set ENABLE_MCP=true

Add to .env (local) and your Kubernetes ConfigMap (production):

ENABLE_MCP=true

4. Restart and verify

Start the app and check the MCP endpoint:

# Start the app
ENABLE_MCP=true uv run application-sdk --mode combined --app app.connector:MyApp

# List registered tools
curl http://localhost:8000/mcp

The response should list all @mcp_tool-decorated tasks.

5. Connect an AI client

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "my-connector": {
      "url": "http://localhost:8000/mcp",
      "transport": "http"
    }
  }
}

Claude Code — run in the project directory:

claude mcp add my-connector http://localhost:8000/mcp

What Pydantic Models Look Like to AI Clients

Given:

class FetchInput(Input):
    database: str
    schema_filter: str = ".*"
    max_tables: Annotated[int, Field(le=1000)] = 100

The MCP tool parameters become:

{
  "database": { "type": "string", "required": true },
  "schema_filter": { "type": "string", "default": ".*" },
  "max_tables": { "type": "integer", "maximum": 1000, "default": 100 }
}

Security Note

The MCP endpoint (/mcp) is unauthenticated by default. In production, ensure it is behind a network policy or auth proxy that restricts access to trusted AI clients only. Do not expose MCP to the public internet without authentication.

Verification Checklist

• uv run python -c "from application_sdk.server.mcp.decorators import mcp_tool" — no import error
• ENABLE_MCP=true in environment
• curl http://localhost:8000/mcp returns tool list
• AI client can discover and call the tool