MCP Tools Cannot Receive RunnableConfig Parameters in LangGraph React Agent

[Jasper-zh]

Problem

When using langchain-mcp-adapters with LangGraph's create_react_agent, MCP tools cannot access RunnableConfig parameters that are automatically passed to local LangChain tools.

Current Behavior

Local LangChain Tools (Working):

@tool
def get_user_info(config: RunnableConfig) -> Dict[str, Any]:
    token = config.get("metadata").get("token")  # ✅ Works
    return {"user_id": "current_user"}

MCP Tools (Not Working):

@mcp.tool()
def event_query(bot_name=None, status: int=None):
    # ❌ Cannot access config.get("metadata").get("token")
    return {"event_query": {...}}

Impact

• Authentication blocked: MCP tools cannot access user tokens
• Multi-tenant impossible: No way to pass user context
• Production limitation: Forces developers to avoid MCP for authenticated scenarios

Root Cause

LangGraph's create_react_agent manages tool invocation internally, and langchain-mcp-adapters doesn't preserve the ability to pass RunnableConfig to MCP tools.

Environment

• langchain-mcp-adapters: 0.1.7
• langgraph: 0.2.20
• mcp: 1.9.4

Request

Add support for passing RunnableConfig parameters to MCP tools, similar to how local LangChain tools receive them.

Priority: High (blocks production use cases requiring authentication)

[wudstrand]

+1

[eyurtsev]

This is a feature request to propagate context variables from langgraph to the underlying MCP tool implementation on the server side.

[fernandocorreia-snorkel]

We're running into this issue. Even just passing the metadata: dict[str, Any] RunnableConfig property (serialized) would be very beneficial. Is there any way we could help?

[vaedprasad]

We're running into this issue. Even just passing the metadata: dict[str, Any] RunnableConfig property (serialized) would be very beneficial. Is there any way we could help?

➕

[jpfcabral]

Let me know if #313 solve your problems about authentication in production scenarios.

I was able to set some headers (like Authorization Bearer) in the tool request dynamically.

[jpfcabral]

Btw, it doesnt make sense to pass RunnableConfig to MCP because it is a specific langchain concept.

Check the way langchain run with RunnableConfig:

https://python.langchain.com/docs/how_to/tool_configure/#inferring-by-parameter-type