vscode doesn't recognize 'available' mcp commands

[loqstrz]

Im using DeepSeek combined with Roo Code and I installed everything manually and everything is running. But when I try to make it use the MCP commands, it just can't do it...

Is this a problem related to the model I use?

[loqstrz]

and to also clarify: even the mcp server is running. I started it through IDA and I got a confirmation that its running. Roo Code can also interact with the LLM and the config for the MCP server is also set accordingly like explained in the instructions. Yet it can't use the MCP commands for some reason...

[mrexodia]

Can you share your MCP configuration for Roo Code?

[loqstrz]

I ended up doing the automatic approach, and even then it couldn't get the MCP commands. It always ends up trying to do something with my opened VSCode files or something similar.

This is interesting, do you have any idea why that could be? Refreshing the MCP server also works, everything is enabled and there are no other errors.

[loqstrz]

And while we are at it: What models would you recommend to use? I am trying to run everything locally using Ollama but it somehow is slow within Roo Code and Cline, yet if I chat with the LLM through powershell, it is super super fast...

[loqstrz]

It turns out that this indeed was an issue with the model. Use the qwen2.5-coder-tools model if you work locally. Not only is it super fast, it also recognizes the MCP commands.

All the other models I've tried were insanely slow and did not recognize the MCP commands.

[mrexodia]

Thanks for the information, I will add a local setup qwen2.5-coder-tools to the README (and close the issue when it's done)!

[sumansuhag]

Ensure Protocol Compliance across Layers: Verify LLM, Roo Code, and VSCode extension precisely adhere to MCP specifications and versioning for seamless communication.

[sumansuhag]

AttributeError signals a protocol handshake failure within the AI development environment, where the VSCode integration isn't correctly interpreting MCP commands from the LLM via Roo Code.

[loqstrz]

Looks like that model is not so good after all.

"Roo tried to use list_functions with an invalid JSON argument. Retrying..." and thats a loop. It won't get further either.

The arguments it passed to it are: "0< /offset > < count >100< /count >" (had to put spaces here or else github makes this look invisible)

Will be experimenting some more because I want this to work.

[sambow23]

I'm having the same issue with Cursor, server is running and I can connect to it, but either claude sonnet 4 or gemini pro 2.5 do not recognize it

Edit: asking the model what mcp tools are available seems to make it recognize it? Not sure why that works

[loqstrz]

yeah @sambow23 make sure to mention that MCP tools are available and it should pick it up