Add "meta" to call_tool method

[jeremyschulman]

Is your feature request related to a problem? Please describe.

I would like to be able to pass metadata via the "tools/call" invocation in the call_tool method. This would afford the Caller to use both the pre-defined client_id value anticipated by the Context.client_id attribute, as well as any other Caller specific values.

Describe the solution you'd like

Here is how I have modified my local copy to allow this feature functionality

    async def call_tool(
        self, name: str, arguments: dict[str, Any] | None = None, **meta
    ) -> types.CallToolResult:
        """Send a tools/call request."""
        return await self.send_request(
            types.ClientRequest(
                types.CallToolRequest(
                    method="tools/call",
                    params=types.CallToolRequestParams(name=name, arguments=arguments, _meta=meta)
                )
            ),
            types.CallToolResult,
        )

Could you please consider adding this feature as shown, or an equivalent implementation that would allow for the desired behavior?

Thank you!

[shiwenbin1617]

mark

[000-000-000-000-000]

curious, is this not already addressed? For some reason this: modelcontextprotocol/modelcontextprotocol#414 made me think it was.

[jeremyschulman]

@000-000-000-000-000 - thank you for the link. My read on that issue is the discussion about the meta parameter in the context of the protocol specification. My issue is about the implementation of the specification in the python SDK "call_tool" method that would expose that feature of the specification.

[pyrotank41]

I noticed this too @jeremyschulman , following is how I went about it:

Following is how i was calling it:

result = await session.call_tool(
                "echo",
                arguments={"name": "Hello, World!"},
                _meta={"custom_param": "custom_value"},
            )

but the _meta doesnt exist as a parameter for the call_tool method. so to get around this i just created a new class like:

from mcp import ClientSession, StdioServerParameters, types
class CustomClientSession(ClientSession):
    async def call_tool(
        self,
        name: str,
        arguments: dict[str, Any] | None = None,
        read_timeout_seconds: Any = None,
        progress_callback: Any = None,
        _meta: Dict[str, Any] | None = None,
    ) -> types.CallToolResult:
        """Send a tools/call request with optional progress callback and meta support."""

        # Properly create the Meta object if _meta is provided
        meta_obj = types.RequestParams.Meta(**_meta) if _meta else None

        # Pass meta as the 'meta' argument (not '_meta')
        params = types.CallToolRequestParams(
            name=name, arguments=arguments, _meta=meta_obj
        )

        # Build the request object
        request_obj = types.ClientRequest(
            types.CallToolRequest(
                method="tools/call",
                params=params,
            )
        )

        # Print the outgoing JSON-RPC request for debugging
        print("Outgoing JSON-RPC request:")
        print(json.dumps(request_obj.model_dump(mode="json", by_alias=True), indent=2))

        return await self.send_request(
            request_obj,
            types.CallToolResult,
            request_read_timeout_seconds=read_timeout_seconds,
            progress_callback=progress_callback,
        )
        
async def main():
    # Define server parameters
    server_params = StdioServerParameters(
        command="python",  # The command to run your server
        args=["server.py"],  # Arguments to the command
    )

    # Connect to the server
    async with stdio_client(server_params) as (read_stream, write_stream):
        # Use our custom session class
        async with CustomClientSession(read_stream, write_stream) as session:
            # Initialize the connection
            await session.initialize()

            # List available tools
            tools_result = await session.list_tools()
            print("Available tools:")
            for tool in tools_result.tools:
                print(f"  - {tool.name}: {tool.description}")

            result = await session.call_tool(
                "echo",
                arguments={"name": "Hello, World!"},
                _meta={"user_id": "1234567890"},
            )
            print(f"Stdio Echo: {result.content[0].text}")


if __name__ == "__main__":
    asyncio.run(main())

and in my server, used it like so:

def extract_meta_from_context(ctx: Context) -> dict:
    """
    Extract meta information from the MCP context.

    Args:
        ctx: The MCP context object

    Returns:
        dict: Dictionary containing custom meta data (excluding progressToken)
    """
    if not hasattr(ctx, "request_context") or ctx.request_context is None:
        return {}

    if not hasattr(ctx.request_context, "meta") or ctx.request_context.meta is None:
        return {}

    # Convert meta object to dictionary
    meta_dict = (
        ctx.request_context.meta.model_dump()
        if hasattr(ctx.request_context.meta, "model_dump")
        else {}
    )

    return meta_dict


# Run the server
if __name__ == "__main__":
    # parser for the type of message transport channel to run the server
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--transport",
        type=str,
        default="stdio",
        choices=["stdio", "streamable-http", "websockets"],
    )
    parser.add_argument(
        "--host",
        type=str,
        default="127.0.0.1",
        help="Host to bind to (only used for SSE transport)",
    )
    parser.add_argument(
        "--port",
        type=int,
        default=8050,
        help="Port to bind to (only used for SSE transport)",
    )
    args = parser.parse_args()

    # Create an MCP server with host and port from arguments
    mcp = FastMCP(name="DemoServer", host=args.host, port=args.port)

    # Simple tool
    @mcp.tool()
    def echo(name: str, ctx: Context) -> str:
        """Echos the input back to the client

        Args:
            name: The input to echo back
        """
        # Use stderr for debugging output (won't interfere with stdio communication)
        print(f"Context: {ctx}", file=sys.stderr)
        print(f"Context type: {type(ctx)}", file=sys.stderr)
        print(f"Context content: {ctx.request_context.meta}", file=sys.stderr)

        # Use the helper function to get meta information
        meta = extract_meta_from_context(ctx)

        # Return the echo with meta information
        return f"{name}! ({meta})"

    # run the server
    if args.transport == "stdio":
        mcp.run()
    elif args.transport == "streamable-http":
        mcp.run(transport="streamable-http")
    elif args.transport == "websockets":
        raise NotImplementedError("Websockets are not supported yet")
    else:
        raise ValueError(f"Invalid transport: {args.transport}")

When i run the client.py i get the following output:

[06/22/25 09:35:45] INFO     Processing request of type ListToolsRequest                                                     server.py:523
Available tools:
  - echo: Echos the input back to the client

Args:
    name: The input to echo back

Outgoing JSON-RPC request:
{
  "method": "tools/call",
  "params": {
    "_meta": {
      "progressToken": null,
      "user_id": "1234567890"
    },
    "name": "echo",
    "arguments": {
      "name": "Hello, World!"
    }
  }
}
                    INFO     Processing request of type CallToolRequest                                                      server.py:523
Context: 
Context type: <class 'mcp.server.fastmcp.server.Context'>
Context content: progressToken=None user_id='1234567890'
Stdio Echo: Hello, World!! ({'progressToken': None, 'user_id': '1234567890'})

As you may notice, the _meta is making it to the server and we are able to get it in the server.

My only concern is that this is not supported by the tool_call method of the ClientSession, hence the CustomClientSession. its an easy fix. maybe the team will get to it eventually.

thanks

[pyrotank41]

@jeremyschulman, i created a PR for this as well #1003

[felixweinberger]

See closure in #1003, closing this associated issue.