Agent Sandbox is a way to run AI Agents inside a MicroVM.
- Uses podman and krun
- Uses squid in a container to control what the container can access
- Supports opencode with Anthropic and Claude Code right now
- Uses meridian and opencode-with-claude
- Configurable sandbox home directory (~/.local/share/agent-sandbox/homes/$sandbox_name), -H overrides the default
- Configurable proxy (~/.config/agent-sandbox/squid.conf)
agent -s: open a shell in the sandbox. opencode and claude can be run from thereagent -o: run opencode in the sandboxagent -c: run claude code in the sandboxagent -n foo -c: run claude code in a sandbox named fooagent -H $HOME/foo -n foo: run claude code in sandbox named foo with its home directory as ~/foo
- Support more agents
- Support more endpoints with opencode
- Make container somewhat configurable