AI

• 1: AI agents inside Lima
• 2: AI agents outside Lima (MCP)
• 2.1: Gemini

1 - AI agents inside Lima

Lima is useful for running AI agents (e.g., Codex, Claude, Gemini) so as to prevent them from directly reading, writing, or executing the host files.

Lima v2.0 is planned to be released with built-in templates for well-known AI agents.

For Lima v1.x, you can install AI agents in Lima manually.

e.g.,

lima sudo apt install -y npm
lima sudo npm install -g @google/gemini-cli
lima gemini

2 - AI agents outside Lima (MCP)

Starting with Lima v2.0, Lima provides Model Context Protocol (MCP) tools for reading, writing, and executing local files using a VM sandbox.

2.1 - Gemini

This page describes how to use Lima as an sandbox for Google Gemini CLI.

Prerequisite

In addition to Gemini and Lima, make sure that limactl mcp plugin is installed:

$ limactl mcp -v
limactl-mcp version 2.0.0-alpha.1

The limactl mcp plugin is bundled in Lima since v2.0, however, it may not be installed depending on the method of the installation.

Configuration

1. Run the default Lima instance, with a mount of your project directory:

limactl start --mount-only "$(pwd):w" default

Drop the :w suffix if you do not want to allow writing to the mounted directory.

2. Create .gemini/extensions/lima/gemini-extension.json as follows:

{
  "name": "lima",
  "version": "2.0.0",
  "mcpServers": {
    "lima": {
      "command": "limactl",
      "args": [
        "mcp",
        "serve",
        "default"
      ]
    }
  }
}

3. Modify .gemini/settings.json so as to disable Gemini CLI’s built-in tools except ones that do not relate to local command execution and file I/O:

{
  "coreTools": ["WebFetchTool", "WebSearchTool", "MemoryTool"]
}

Usage

Just run gemini in your project directory.

Gemini automatically recognizes the MCP tools provided by Lima.