Summary of issues (internal)

[eyurtsev]

Contributors

2025-09-11: We are not looking for contributions for these issues at this time. MCP support will be improved over the next few weeks and handled by langchain maintainers.

The given list is a summary of issues for our own reference

Issues

• Propagate context (e.g., RunnableConfig) to a remote MCP tool
  • MCP Tools Cannot Receive RunnableConfig Parameters in LangGraph React Agent #271 (comment)
  • LangChain tools with injected arguments are not supported #259
• Customize tools How to customize Command return for tools fetched from MCP to update specific keys in graph state? #156
• Per tool header injection:
  • Support per-tool header injection (separate from MCP server connection headers) #297 — need to spec out potentially middleware
  • [Feat] Allow HTTP headers to be supplied dynamically #280
• Prefix tool name w/ server name due to mcp collisions (Allow adding a prefix to tool names to deal with tool name collisions #273)
• Document how to deal with stateful sesssions:
  • How to add persistence sessions when working with langgraph #189
  • Intermittent 400 Bad Request when connecting to MCP Server using streamable_http transport Description #180
• Integration with custom streaming in langgraph for MCP tool calling progress notifications etc? (how can MCP tools use "get_stream_writer" like langchain tool #291)
• human in the loop / elicitation https://modelcontextprotocol.io/specification/2025-06-18/client/elicitation with langgraph interrupts (requires API design / research)
• The load_mcp_tools function currently doesn’t allow passing custom error-handling parameters, forcing users to manually patch returned tools. Allowing it to forward handle_tool_error and handle_validation_error would enable more efficient, flexible error handling during tool execution. [Feature Request] Support handle_tool_error and handle_validation_error in load_mcp_tools #263
• retries (e.g., server was scaled down)
• Sampling Feature request: Add MCP Sampling support #232
• Tool annotations and metadata Support MCP tool annotations and metadata #138
• Handle structured outputs: Extend _convert_call_tool_result helper class in tools.py #272
• Document how to capture server logs (How to capture the logs returned by ctx.info from fastmcp server #217)

Lower priority:

• Specifying more than one server name Allow specifying more than one server names in the get_tools call of MultiServerMCPClient #200

Sync support:

• Consider adding built-in sync support? (e.g., mcp 工具sse调用失败，报错 ToolMessage(content="Error: NotImplementedError('StructuredTool does not support sync invocation.') #173) StructuredTool can excute "invoke" method,but ReactAgent can not excute "invoke" method #158

[sydney-runkle]

Updated summary here, Dec 1st

Issues

Most important

• Handle structured outputs: Extend _convert_call_tool_result helper class in tools.py #272
• Elicitation support w/ langgraph interrupts, MCP Elicitation support #268

Nice to haves, smaller scope

• Propagating context into the tool MCP Tools Cannot Receive RunnableConfig Parameters in LangGraph React Agent #271 (comment) and LangChain tools with injected arguments are not supported #259, might not be possible except via interceptors
• Prefix tool name w/ server name due to mcp collisions (Allow adding a prefix to tool names to deal with tool name collisions #273)
• Customize tool returns (confirm Command and ToolMessage work) How to customize Command return for tools fetched from MCP to update specific keys in graph state? #156

Lower prio, not enough demand for 1.2

• Error handling behavior [Feature Request] Support handle_tool_error and handle_validation_error in load_mcp_tools #263, unscoped
• Specifying more than one server name Allow specifying more than one server names in the get_tools call of MultiServerMCPClient #200
• Sampling Feature request: Add MCP Sampling support #232
• Tool annotations and metadata Support MCP tool annotations and metadata #138

Docs

• injecting context
• per tool header injection
• stateful sessions How to add persistence sessions when working with langgraph #189 and Intermittent 400 Bad Request when connecting to MCP Server using streamable_http transport Description #180
• interceptors
• callbacks (progress streaming, how can MCP tools use "get_stream_writer" like langchain tool #291)
• Capturing server logs How to capture the logs returned by ctx.info from fastmcp server #217

[sydney-runkle]

At this point, I'd say next steps are:

• Better support for prompts (like get_prompts), pagination
• Better support for resources, using standard content, maybe exposing as tools, pagination
• Adding in hooks for other callbacks exposed via the core sdk
• For tools, adding error handling, looking into tool annotations + metadata
• Haven't seen much demand yet for sampling or other advanced features, but we definitely want to nail the core ones above