Hey early adopters! 👋

[jamubc]

Since the gemini-mcp-tool launched a couple days ago, I'm super
curious to hear about your experiences so far!

What I'd love to know:

• What did you build with it / Did it change how your workflow?
• Any unexpected use cases you discovered?
• Any problems with failed tool calling? Claude not following gemini-mcp-tool related instructions?
• What was your first impression?
• Did the tool work as expected?

No experience is too small!

Let's learn together

The tool is new, so we're all figuring this out together. Your
discoveries today become tomorrow's best practices.

Drop a comment and let's see what this community can build! 🚀

[maker95]

I see that most of the times I use the gemini-mcp-tool the tool tries to use 'replace' without success. ¿I am doing something wrong? It seems that Gemini CLI tries to use 'replace' from within the CLI wich it seems is not found as a tool in the mcp. For reproduce it just had to ask Claude to use gemini for try to find out why a bug is happening, the response from Gemini is received but trying to use 'replace' without success every time it tries.

I installed Gemini CLI from "npm install -g @google/gemini-cli" and succesfully installed "gemini-mcp-tool" into Claude, it works. The problem is the 'conflicting' intent to use replace tool.

[jamubc]

Thanks for the feedback!

I hear your concern and I'm attempting to tackle this issue now. Hopefully we see an update in a few hours.

I have a couple ideas:

1. "Claude does the edits": Force Gemini (via prompt) to return structured responses with specific line numbers/sections, then let Claude handle the actual file modifications using its native tools (multi edit, find and replace, etc.). This seems like the cleaner approach but requires careful context forwarding.

2. "Gemini does the edits": Handle Gemini's permission prompts programmatically. My concern here is security - we'd need to be careful about auto-approving edits. A sand-boxing approach could work: let Gemini make edits in a sandbox, generate diffs, and manually approve before applying to real files.

3. ???

I will let you know how it goes!

[jamubc]

Update: claude edits, gemini reads, control via claude's HIL UI with diff view.

Hey @maker95,

I've pushed a significant update to the update branch that addresses the replace tool issue you reported.

Key changes:

• Claude for handling edits. Gemini now formats suggestions, and Claude applies them, eliminating the tool-not-found error while allowing Claude to make edits without ever reading the file.
• Introduced experimental deterministic hook system to improve tool usage accuracy. The current MCP-based approach works for most tested cases, but hooks should significantly enhance reliability. Not fully integrated yet, may not work as expected. Builds upon current system, no missing behavior without hooks yet.

Please test it with your original bug-finding prompt and let me know if it performs as expected. Thanks!

Test Details: See the Claude code log for recent testing examples.

[jamubc]

Just refactored Gemini-mcp-tool for the Newley released QWEN cli, allowing you to connect any OpenAI type api to a clone of Gemini cli.

In essence you can use it to run any AI you want as a "AI-mcp-tool" allowing you to orchestrate complex tasks with a variety of models.

I was able to use qwen cli running gpt-4o with no issues, same performance as Gemini-mcp-tool.

Hope it's useful:

qwen-mcp-tool