@victornanka/gemini-cli-mcp

MCP server for delegating tasks to Gemini CLI. This allows you to invoke Gemini CLI agent capabilities from other MCP-enabled tools.

Prerequisites

• Gemini CLI must be installed and configured
• The gemini command must be available in your PATH

Installation

For Gemini CLI

Add to your ~/.gemini/settings.json:

{
  "mcpServers": {
    "gemini-delegate": {
      "command": "bunx",
      "args": ["@victornanka/gemini-cli-mcp@latest"]
    }
  }
}

For Claude Code

Add to your Claude Code MCP configuration:

{
  "mcpServers": {
    "gemini-cli-mcp": {
      "command": "bunx",
      "args": ["@victornanka/gemini-cli-mcp@latest"]
    }
  }
}

Usage

Once installed, you can use the task tool to delegate tasks to Gemini CLI:

@gemini-delegate task("Create a Python script that prints hello world", cwd="/path/to/project")

Parameters

• task (required): The task description to delegate to Gemini CLI
• cwd (required): The working directory where Gemini CLI should run (must be an absolute path)
• historyId (optional): Continue from a previous session (use the session_id from a previous response)

How it Works

This MCP server wraps the Gemini CLI and allows other tools to programmatically invoke it. When you call the task tool:

1. The MCP server spawns the gemini CLI with your task
2. It uses --output-format stream-json to get structured responses
3. The response is streamed back through the MCP protocol
4. The conversation session ID is preserved for follow-up tasks

Response Format

The tool returns a JSON object with:

• result: The assistant's response text
• session_id: The session ID for continuing the conversation
• total_cost_usd: The total cost of the request (if available)

Development

# Install dependencies
bun install

# Build
bun run build

# Test locally
bun run dist/index.js