run coding agents in localised airlocked microVMs
i built code-airlock, an open-source tool for running AI coding agents inside disposable microVMs
the problem I kept running into was the tradeoff between approving every command manually and losing most of the time savings, or disabling command prompts and trusting the agent with my actual machine. I tried the second path too often and regretted it.
code-airlock runs the agent inside a disposable microVM using Docker Sandboxes. the agent works on a clone of the repo, so it can install dependencies, run builds, start containers, and modify files inside the sandbox without direct access to my host filesystem or credentials.
when the agent finishes, I review the diff and pull back only the changes I want!
the design choice was to isolate the whole environment instead of maintaining a long list of deny rules. Deny rules are brittle with coding agents because agents are goal-driven and will often route around constraints when trying to finish a task. A microVM gives a clearer blast radius.
Current support:
- Claude Code
- Codex
- OpenCode
- few others
Requirements:
- Docker Sandboxes CLI
- Hardware virtualization: Apple Silicon or KVM