Gemini Editor MCP Server

gemini-editor exposes an alternative MCP tool to edit files using Gemini. This reduces burden on main model by not enforcing tight schemas and using a smaller editor model to apply the changes as described by the main model. Essentially following the very effective architect/editor structure from aider.

Why use this alternative Gemini Editor?

The built-in replace tool in Gemini CLI is very strict and requires very specific input structure. This results in gemini model frequently stumbling with it. This is due to how LLMs are bad at strict schemas and too many tools which causes the performance to degrade very quickly.

• Simplified Change Description: We ask main model to instead just describe their changes however it makes sense to them. No tight schemas.
• Delegated Execution: Uses a smaller "editor" model (gemini-flash) to interpret the description and apply code modifications.
• Efficient (soon): Currently replaces the entire file. A granular search-and-replace feature is planned.

Prerequisites

This uses the Gemini API to perform file edits. Before running, you must set the GEMINI_API_KEY environment variable in your shell:

export GEMINI_API_KEY="YOUR_API_KEY"

Configuration with Gemini

To expose the gemini-editor tool to Gemini CLI, you need to configure your settings. You can do this either globally or for a specific project.

Global Configuration

Add the following to your global ~/.gemini/settings.json file:

{
  "mcpServers": {
    "gemini-editor": {
      "command": "npx",
      "args": ["gemini-editor"]
    }
  },
  "excludeTools": ["replace"]
}

Project-level Configuration

In your project's root directory, create or open the .gemini/settings.json file and add the following:

{
  "mcpServers": {
    "gemini-editor": {
      "command": "npx",
      "args": ["gemini-editor"]
    }
  },
  "excludeTools": ["replace"]
}

After this, Gemini will be able to use the replace_file_contents tool instead of the default built-in one.

Roadmap

• [] Search/Replace-based Edits
• [] Better Evals