The In Process MCP is not working at all

[aipha-capture-dev]

I tried the in process mcp just like mentioned in the documenation. Claude client (or query for that matter) cannot see it at all. There is definitely a bug in it. The server itself is correct as in: you can check its tools, you can call them. However, embedding the server in a client or query does not work.
The example https://github.com/anthropics/claude-code-sdk-python/blob/main/examples/mcp_calculator.py is wrong and will produce correct results with or without calculator. Try to place a break point or a log inside the tool add_numbers for example and you will see that function is never hit

[badassleo007]

+1

[saixiaohui]

+1, I’m not sure if this is happening because I have other MCP servers configured.

[aoezdTchibo]

Hi!

I did some digging with Claude Code and found some interesting details about what's happening with the in-process MCP servers.

What I observed

Claude Code was able to run the existing integration tests successfully, and MCP servers work perfectly when tested directly. However, when using the query() function with SDK MCP servers, the tools aren't being discovered by Claude during actual conversations.

Investigation findings

1. Import issue - The tool and create_sdk_mcp_server functions aren't available via star imports because they were removed from __all__ in commit e4feaf2 when the feature was marked unstable.
2. Initialization protocol - The initialize() method sends hooks configuration to the CLI but doesn't include SDK MCP server tool definitions. The SDKControlInitializeRequest type is missing an sdk_mcp_servers field.
3. Tool discovery vs execution gap - Claude recognizes SDK tool names (shows "Tool used: X") but then reports the tool "is not available", suggesting a disconnect between tool discovery and execution routing.

Local experiments

I tried adding the missing initialization fields and exports locally. The SDK-side changes allow proper star imports and tool extraction from SDK servers during initialization, with all existing tests continuing to pass.

However, even with Claude CLI v1.0.108, the streaming mode initialization encounters "ProcessTransport is not ready for writing" errors, and tools still aren't executable in practice.

Current observation

The behavior suggests that while the SDK can create and manage MCP servers internally, the communication protocol between the Python SDK and Claude CLI for in-process servers may need additional coordination. The CLI appears to receive some tool information (hence recognizing names) but can't complete the execution routing back to the SDK.

[arthurgousset]

Thanks for sharing @aoezdTchibo !

[aoezdTchibo]

Also regarding the SDK MCP server streaming mode requirement...

I discovered that SDK MCP servers have a streaming mode requirement that external MCP servers don't have:

1. External MCP servers work fine with string prompts (non-streaming mode) while SDK MCP servers don't.
2. SDK MCP servers require streaming mode due to control protocol:
   Looking at src/claude_code_sdk/_internal/query.py:283, there's a hard requirement:

   async def _send_control_request(self, request: dict[str, Any]) -> dict[str, Any]:
       if not self.is_streaming_mode:
           raise Exception("Control requests require streaming mode")

3. SDK MCP servers need control requests for initialization (to register tools with the CLI), while external servers are passed via CLI arguments.

Question: Is this streaming mode requirement by design for SDK MCP servers?

I am trying to wrap a FastAPI server around Claude Code (exposing CC via endpoints for specific business processes) where I don't have an interactive chat but rather handle one-off API requests. Would it be possible to support SDK MCP servers in non-streaming mode by passing SDK server tool definitions via CLI arguments (similar to external servers)?

[arthurgousset]

Cross-posted in the TypeScrip repo, because that's the library I'm using:

• [BUG] In-process MCP servers bug in Claude Code TypeScript SDK claude-code#7279

I came here to check if anyone had this issue and found this thread.

[aoezdTchibo]

With yesterday's commit 8393004, I was able to successfully test the in-process SDK MCP tool locally. In my opinion, this issue can be closed.

Many thanks @ashwin-ant ❤️

[ashwin-ant]

Thanks all for reporting! This should be fixed!