Hello nullers
I've been building Airlock, is an open-source MCP server. Every command your agent sends is parsed, weighed, asked to you in your client when it's risky, then run in a sandbox. The sandbox writes only to the project, can't read your keys, and reaches only the package registries on the network. The output comes back with secrets masked and is checked for prompt injection.
It works with Claude Code, Codex, VS Code, Cursor and Goose and other agents, on Linux and Apple Silicon Macs:
I'm looking for people who use a coding agent every day to try it on a real project. I'd like to hear where it asks too often, what it blocked wrongly, and what breaks.
Repo and quick start: https://github.com/hbenabdallah/airlock
https://youtu.be/ve3vKHaWW2k
bash
curl -fsSL https://raw.githubusercontent.com/hbenabdallah/airlock/main/install.sh | sh airlock connect claude-code
GitHubGitHub - hbenabdallah/airlock: An MCP server that lets any agent run commands in a sandbox, asking you first about the ones that matterAn MCP server that lets any agent run commands in a sandbox, asking you first about the ones that matter - hbenabdallah/airlock
commit c23dfe4