AGENT SKILL

airlock

Gate an agent extension before you install it. Audits a skill, MCP server, or tool repo and returns allow, quarantine, or block with line-level evidence.

License 许可
MIT
所属仓库
ByJH/airlock
最近更新
2026-07-07
Stars
★ 0
Forks
0

01 前言 Front Matter

name
airlock
license
MIT
description
Gate an agent extension before you install it. Audits a skill, MCP server, or tool repo and returns allow, quarantine, or block with line-level evidence.
permissions
read-only

03 SKILL.md 正文

Airlock

A gate for the things an agent installs. Use it before adding any skill, MCP server, or tool repo that came from a source you have not fully vetted. It works out what the artifact is, scans it for dangerous shapes, reads it adversarially with a model, detonates it once in a sealed sandbox, and returns one verdict.

When to run it

Run Airlock on an extension before the install step, not after. Good moments:

  • Someone shares a SKILL.md folder or a repo URL and asks you to install it.
  • You are about to add an MCP server you found in a public registry.
  • A dependency wants to run a lifecycle hook on install.

How to run it

Airlock is a Python command line tool. Install it once, then point it at a path or a cloned repo:

pip install -e ".[sandbox]"     # from the Airlock repo
airlock audit path/to/extension

The exit code is the gate, so a script or a hook can branch on it:

  • 0 allow. No known-bad shape, a clean or skipped judgment, no canary touch.
  • 1 quarantine. Something needs a human to look before it goes in.
  • 2 block. A critical finding, a canary touch, or a judged exfiltration.

Add --json for a machine-readable scorecard, --no-llm to run the deterministic core with no API key, and --no-sandbox to skip the container trial. The deterministic core always produces a verdict, even offline.

How to act on the verdict

  • On block, do not install the extension. Read the cited findings and tell the user which line drove the decision.
  • On quarantine, hold the install for a human review. Summarize what the scan and the judge flagged.
  • On allow, it is safe to proceed. A clean scan means no known-bad shape matched, not a proof of safety, so still prefer trusted sources.

What it looks at

The scorecard lists every finding with a file and line: credential-file reads, environment exfiltration, reverse shells, remote installers that pipe a downloaded script into a shell, package lifecycle hooks, and prompt-injection phrasing aimed at the agent reading the skill. The final gate is deterministic, so the same artifact always grades the same way and a model is never the last word on what gets in.

04 同仓库更多 Skills