AI agents

Lima is useful for running AI agents inside a VM, so as to prevent agents from directly reading, writing, or executing the host files.

For running AI agents, it is highly recommended to only mount your project directory (current directory) into the VM:

limactl start --mount-only .:w

Drop :w for read-only mode.

limactl start --set ".mounts=[{\"location\":\"$(pwd)\", \"writable\":true}]"

Set writable to false for read-only mode.

lima sudo apt install -y pipx
lima pipx install aider-install
lima sh -c 'echo "export PATH=$PATH:$HOME/.local/bin" >>~/.bash_profile'
lima aider-install
lima aider

Follow the guide shown in the first session for authentication.

Alternatively, you can set environmental variables via:

lima vi "/home/${USER}.linux/.bash_profile"

See also https://github.com/Aider-AI/aider.

lima sudo snap install node --classic
lima sudo npm install -g @anthropic-ai/claude-code
lima claude

Follow the guide shown in the first session for authentication.

Alternatively, you can set export ANTHROPIC_API_KEY... via:

lima vi "/home/${USER}.linux/.bash_profile"

See also https://github.com/anthropics/claude-code.

lima sudo snap install node --classic
lima sudo npm install -g @openai/codex
lima codex

Follow the guide shown in the first session for authentication.

Alternatively, you can set export OPENAI_API_KEY... via:

lima vi "/home/${USER}.linux/.bash_profile"

See also https://github.com/openai/codex.

lima sudo snap install node --classic
lima sudo npm install -g @google/gemini-cli
lima gemini

Follow the guide shown in the first session for authentication.

Alternatively, you can set export GEMINI_API_KEY... via:

lima vi "/home/${USER}.linux/.bash_profile"

See also https://github.com/google-gemini/gemini-cli.

lima sudo snap install node --classic
lima sudo npm install -g @github/copilot
lima copilot

Type /login in the first session for authentication.

Alternatively, you can set export GH_TOKEN=... via:

lima vi "/home/${USER}.linux/.bash_profile"

See also https://github.com/github/copilot-cli.

lima sudo snap install node --classic
lima sudo npm install -g opencode-ai
lima opencode

Type /connect in the first session for authentication. Unlike other agents, this step is not necessary for OpenCode.

See also https://github.com/anomalyco/opencode.

Syncing Working Directory

⚡ RequirementLima >= 2.1

The --sync flag for limactl shell enables bidirectional synchronization of your host working directory with the guest VM. This is particularly useful when running AI agents (like Claude, Copilot, or Gemini) inside VMs to prevent them from accidentally modifying or breaking files on your host system.

Comparison with mount

FeatureMounts (--mount/--mount-only)Sync (--sync)
PurposeMake host directories visible inside guest (bidirectional if write mode is enabled)Temporary bidirectional sync of a working directory (guest changes merged back on accept)
Live updatesYesNo
SafetyLower (AI agents can access host files directly)Higher (changes are reviewed before being applied to host)
Requires rsyncNoYes

Usecase - Running AI Code Assistants Safely

  1. Create an isolated instance for AI agents which must be started without host mounts for --sync to work:
limactl start --mount-none template://default
  1. Navigate to your project
cd ~/my-project
  1. Run an AI agent that modifies code:
limactl shell --sync . default claude "Add error handling to all functions"

Or simply shell into the instance and make changes:

limactl shell --sync . default
  1. After running commands, you’ll see an interactive prompt:
⚠️ Accept the changes?
→ Yes
  No
  View the changed contents
  • Yes: Syncs changes back to your host and cleans up guest directory
  • No: Discards changes and cleans up guest directory
  • View the changed contents: Shows a diff of changes made by the agent

Requirements

  • rsync must be installed on both host and guest
  • The host working directory must be at least 4 levels deep (e.g., /Users/username/projects/myproject)
  • The instance must not have any host mounts configured (use --mount-none when creating)

See also