Containers isolate processes.
Bunkerbox isolates consequences.¶
What it does¶
AI coding agents are powerful. They can edit code, run builds, and iterate on your
project. They can also rm -rf your home directory, exfiltrate your SSH keys, or
fill your disk with garbage — because they run as you.
Bunkerbox puts the agent inside a lightweight, immutable VM. From that sandbox:
- Your repo is an overlay copy-on-write workspace. The agent sees and edits files as usual, but your real repository stays untouched until you choose to sync changes back.
- Your build tools are available on-demand. A whitelist tells Bunkerbox which
host commands to proxy into the VM via vsock. The agent runs
cargo build,make test, orgo vetusing your real toolchain — inside the overlay, never on the host itself. - Your credentials stay encrypted at rest. The VM gets a clean home directory. Sensitive files are sealed with AES-256-GCM between sessions.
- Your network is firewalled. Only destinations you explicitly allow are reachable from inside the container.
- All changes auto-sync when the agent exits. Your repo has the diff. You review, accept, commit.
Documentation website¶
Build the documentation website with:
The static site is written to:
Where to go next¶
Start with the Tutorial to set up Bunkerbox and let an AI agent work on a real project. Read Concepts to understand the model. Read Packaging to understand how a tool becomes a normal command.