lamp-the-djinn

Appearance

Getting Started ​

By the end of this page you'll have run a coding agent inside the cage: full autonomy for the agent, no risk to the rest of your system.

Prerequisites ​

  • Docker installed, running, and accessible by your user (docker info succeeds).
  • uv (for uvx) — https://docs.astral.sh/uv/. pipx works too.

Run it ​

From the project directory you want the agent to work in:

bash
uvx --from git+https://github.com/thekevinbot/lamp-the-djinn lamp-the-djinn

The first run pulls the container image (or builds it locally if the pull fails), starts the cage, configures the firewall, and drops you into Claude Code with --dangerously-skip-permissions — inside the cage, that's safe.

ltd is the short alias; both entry points are the same program.

Run a different agent ​

Everything after ltd's own options is the command that runs in the cage, passed through untouched:

bash
ltd claude -p "fix the bug"   # the -p goes to claude, not ltd
ltd npx @anthropic/claude     # an agent straight from npm
ltd aider                     # any harness on PATH
ltd pi -c                     # harness flags pass through
ltd --safe-mode               # bare claude with permission prompts on

What just happened ​

  • Your project directory — and nothing else — was mounted into the cage, at its real host path.
  • A default-deny egress firewall came up: only allowlisted domains are reachable from inside.
  • The agent ran as an unprivileged user with no access to your credentials.
  • When the command exited, the cage was destroyed. The diff it left in your project is yours to keep or git reset --hard away.

Next steps ​

  • Deep Dive — the two seams (isolation, provider), the trust model, and how credentials stay out of the cage.
  • Troubleshooting — GPG signing and mounting extra folders.
  • Dedicated GitHub identity — recommended, to distinguish agent commits from your own.