Airlock

Airlock, a runtime firewall for AI coding agents

[](https://pypi.org/project/airlock-agent/) [](https://github.com/cyberbobas/airlock/actions/workflows/tests.yml) [](https://pypi.org/project/airlock-agent/) [](LICENSE)

Gate every tool call, MCP call and skill your agent runs against a least-privilege policy. Static scanners check a skill once, before install. Airlock sits in the call path and decides **this call, right now**. That is where a skill that reads clean and behaves badly actually gets stopped.

Part of [Agentoffense](https://agentoffense.com/solutions/airlock_ai/).


 agent ──native tools──▶ [PreToolUse hook] ─┐
                                            ├─▶ policy ─▶ allow / ask / block
 agent ──MCP stdio──▶ [airlock-mcp] ──▶ server ─┘         │
                                                          ▼
                                          audit.jsonl (hash-chained, signable)

Install

**One command:**

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/cyberbobas/airlock/main/install.sh | sh

# Windows (PowerShell)
irm https://raw.githubusercontent.com/cyberbobas/airlock/main/install.ps1 | iex

**Or with your own package manager:**

pipx install airlock-agent      # or: uv tool install airlock-agent / pip install airlock-agent
brew tap cyberbobas/tap && brew install airlock   # macOS

Then:

airlock demo                    # watch it stop a key-theft (nothing of yours touched)
airlock init                    # wire the hook, wrap every agent's MCP servers
airlock doctor                  # confirm what is actually enforcing
`