AI & Agent Dev Bug Sandbox Back to Radar

Feature Request: Pass `_Meta` Parameter Through To MCP SDK `Call_tool()`

langchain-mcp-adapters does not pass the MCP `_meta` field from tool arguments to the MCP SDK's `call_tool()`, preventing per-request metadata like session IDs, routing tags, or tenant context from reaching MCP servers.

mediumConfidence 85%Langchain

Origin Analysis

The MCPToolCallRequest dataclass lacks a `meta` field, the `_invoke` helper passes the full args dict to `call_tool` without extracting `_meta`, and the `_handler` never forwards metadata to the MCP SDK.

1. Create an MCP tool with a server that logs or validates `_meta`. 2. Invoke the LangChain wrapped tool with arguments `{'message': 'hello', '_meta': {'session_id': 'abc123'}}` using `tool.ainvoke`. 3. Observe that the MCP server does not receive any `_meta` metadata and may treat `_meta` as a normal tool argument.

Fixing Code Block

# In langchain_mcp_adapters/tools.py

# 1) Add meta field to dataclass (inside MCPToolCallRequest)
    meta: dict[str, Any] | None = None

# 2) In _invoke, before constructing MCPToolCallRequest:
    tool_args = dict(args)
    meta = tool_args.pop('_meta', None)

    call_request = MCPToolCallRequest(
        name=tool.name,
        args=tool_args,
        headers=headers,
        timeout=timeout,
        meta=meta,
        interceptors=interceptors,
    )

# 3) In _handler, when calling session.call_tool:
    response = await session.call_tool(
        request.name,
        arguments=request.args,
        read_timeout_seconds=request.timeout,
        headers=request.headers,
        meta=request.meta,
    )

Adds a `meta` field to `MCPToolCallRequest`, extracts `_meta` from the incoming tool arguments before constructing the request (avoiding mutation of the original args), and forwards `meta` to `ClientSession.call_tool()` in the low-level handler. Interceptors can also override `meta` via `request.override(meta={...})`.

Edge Case Audit

This change requires an MCP SDK version that supports the `meta` parameter in `ClientSession.call_tool()` (spec 2025-06-18 or later). On older SDK versions a `TypeError` will occur. Ensure the extracted `_meta` is a dict or `None`; non-dict values will be rejected by the SDK. If this patch is reverted, user code relying on `_meta` being stripped from tool arguments will break again; consider a deprecation period or keep extracting `_meta` even if not forwarded. Concurrent invocations using the same interceptor should not mutate the request object; the current `override` creates a new dataclass, which is safe.