SkillAgentSearch skills...

airlock

An MCP server that lets any agent run commands in a sandbox, asking you first about the ones that matter

Install / Use

claude mcp add alpine-ronin -- npx -y github:alpine-ronin/airlock

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

80/100

Supported Platforms

Claude Code
Claude Desktop

Tags

Our assessment of airlock

airlock scores 80/100 on our quality scale, 3564th of 4,705 Development & Engineering skills we index.

Its MCP Server is 30 KB long, well organised into 17 sections with 23 code examples: a thorough specification that gives an agent plenty to work with.

It has 3 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
20/20
Description
12/15
Adoption
3/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated today, so airlock is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 92/100, with 1 caution from licensing, adoption, age or documentation. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

Review

Our scan of the whole file found 1 pattern worth reviewing before you install airlock. An AI review of the same text found nothing harmful.

  • mediumSends the output of a local command to a remote serverline 375
    curl (`-d @.env`, `-F 'key=@id_rsa;filename=public.txt'`) was asked about, but not *every time*.
  • noteInstalls by piping a downloaded script into a shellline 102
    curl -fsSL https://raw.githubusercontent.com/alpine-ronin/airlock/main/install.sh | sh
  • noteInstalls by piping a downloaded script into a shellline 108
    curl -fsSL https://raw.githubusercontent.com/alpine-ronin/airlock/main/install.sh | sh -s -- 0.1.0

AI review by kimi-k2.7-code on 2026-11-16. Automated pattern scan on 2026-11-16. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

airlock compared with similar skills

All 4 of these similar skills score higher than airlock; compare them before choosing.

SkillScoreStarsUpdatedFormat
airlock (this skill)by alpine-ronin803todayMCP Server
taste-skillby Leonxlnx100117.0k8d agoCLAUDE.md
Agent-Reachby Panniantong100112.4k9d agoCLAUDE.md
rufloby ruvnet10081.9ktodayCLAUDE.md
headroomby headroomlabs-ai10079.9ktodayCLAUDE.md

Frequently asked questions

How do I install airlock?
Run claude mcp add alpine-ronin -- npx -y github:alpine-ronin/airlock. The install tabs above show the steps for each supported agent.
Which AI agents does airlock work with?
It is written for Claude Code and Claude Desktop, as a MCP Server file. Other agents that read the same format can often use it too.
Is airlock safe to use?
Our scan of the whole file found 1 pattern worth reviewing before you install airlock. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 92/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is airlock still maintained?
The repository was last updated today, so airlock is actively maintained.
<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="docs/brand/logo-dark.svg"> <img src="docs/brand/logo.svg" alt="Airlock" width="380"> </picture> </p> <p align="center"> <strong>Every agent command passes through it.</strong><br> An MCP server that lets any agent run shell commands on your machine, without trusting the agent with your machine. </p> <p align="center"> <a href="https://github.com/alpine-ronin/airlock/actions/workflows/check.yaml"><img src="https://github.com/alpine-ronin/airlock/actions/workflows/check.yaml/badge.svg" alt="check"></a> <img src="https://img.shields.io/badge/PHP-8.5-12161C?style=flat-square" alt="PHP 8.5"> <img src="https://img.shields.io/badge/Symfony-8.1-12161C?style=flat-square" alt="Symfony 8.1"> <img src="https://img.shields.io/badge/MCP-2026--07--28-12161C?style=flat-square" alt="MCP 2026-07-28"> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-E8930C?style=flat-square" alt="MIT license"></a> <a href="#counting-installs"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fapi.github.com%2Frepos%2Falpine-ronin%2Fairlock%2Freleases%2Ftags%2Finstalls&query=%24.assets%5B0%5D.download_count&label=installs&color=12161C&style=flat-square" alt="first installs"></a> </p> <p align="center"> <a href="#quick-start-five-minutes">Quick start</a> · <a href="#try-the-demo">Demo</a> · <a href="https://youtu.be/ve3vKHaWW2k">Video</a> · <a href="#configuration">Configuration</a> · <a href="docs/design.md">Design</a> · <a href="docs/judges.md">For judges</a> </p>

How it works

Any MCP client connects to Airlock like to any MCP server: Claude Code, Cursor, VS Code, Zed… Every command the agent sends goes through five steps:

| | Step | What happens | |---|---|---| | 1 | Read | Parsed the way a shell reads it: a grant for make test never covers make test; rm -rf ~. | | 2 | Weighed | Plain; asked every time (sudo, a recursive delete, git push, a download piped into a shell, a secret file, anything that cannot be read whole); or refused by your policy. | | 3 | Asked | In your client, with the reason first: once, this session, always for this project or refuse. | | 4 | Run | In a sandbox, bubblewrap on Linux and sandbox-exec on macOS: it writes only to the project, /tmp and package caches, cannot read your keys or Airlock's own files, and on the network reaches the package registries and the hosts you allow (a command reaching a git or SSH remote, any host). | | 5 | Checked | The output goes back marked as data, with secrets masked, and a model flags it when it looks like a prompt injection. |

Every decision is kept in an audit log, readable as the MCP resource airlock://audit. How and why: docs/design.md.

A command never gets a token it could copy, so gh cannot reach GitHub from the sandbox. Airlock reads a pull request for the agent instead, with your own gh login: github_pr_checks gives its CI checks, github_pr_comments its reviews and comments, line comments included, the current branch's pull request or the one numbered. Neither asks you, and both come back like a command's output: checked, marked as data, secrets masked, and in the audit log. github_open_pr opens a pull request from the current branch, once the branch is pushed: Airlock never pushes. It asks you every time, showing the repository, the branch and the one it merges into, the title, whether it is a draft, and the body; you allow it once or refuse, and no grant covers it. On GitHub Enterprise, log in with gh auth login --hostname <host>: Airlock does not pass on GH_ENTERPRISE_TOKEN, which gh would send to whatever host the repository names.

Built with Symfony AI and the PHP MCP SDK

  • The MCP server is the MCP bundle (symfony/mcp-bundle, on the PHP SDK mcp/sdk): tools and the audit resource are classes in src/Mcp, served over stdio only.
  • Elicitation is the permission system. A command that needs your answer returns an MCP 2026-07-28 input_required result carrying a form; the SDK turns it into a classic elicitation/create for clients that only speak 2025-11-25. The request state that rides along with the retry comes back through the client, so the bundle signs it, and Airlock binds it to the exact command and to a one-use token: an answer for make test cannot be replayed to run anything else.
  • A model reads the output as data. symfony/ai-platform and its generic bridge call any API that speaks OpenAI's chat completions, configured through the AI bundle, with structured output: the model answers a field, clean or injection, not prose to parse. A second call says in one line what a risky command would do. Neither decides anything: both are optional, and nothing about running a command depends on them.

Documentation

| Document | What it covers | |---|---| | judges.md | For hackathon judges: running it in five minutes, and where to look in the code | | design.md | The problem, what happens to a command, the protocol, what Symfony AI does, what is not in scope | | architecture.md | The modules of src/, the path of one run_command call through them, where the tests live | | decisions.md | Each technical and functional decision: why, and what was rejected | | network.md | The registries-only network: the proxy, the hosts, what it does not stop | | formats.md | The policy, network, grants, audit and models files | | mcp-spec.md | Each MUST and SHOULD of the MCP spec on tools, elicitation and request state, and what Airlock does |

Quick start: five minutes

Two steps: install Airlock, one of three ways, then connect your MCP client to it. Then try the demo.

1. Install Airlock

Each way ends with the command that starts the server. Your client runs it; you never run it yourself.

With a binary (no PHP, no Docker)

Each release holds one executable per platform, PHP 8.5 included: airlock-macos-aarch64 (Apple Silicon), airlock-linux-x86_64 and airlock-linux-aarch64, with their SHA256SUMS. The install script picks the one for this machine, checks it against SHA256SUMS and, when the GitHub CLI is logged in, checks its attestation (see below; releases before 0.3.4 have none, so theirs is not checked), then puts it in ~/.local/bin (AIRLOCK_INSTALL_DIR to choose another directory). Read it first if you like: install.sh.

curl -fsSL https://raw.githubusercontent.com/alpine-ronin/airlock/main/install.sh | sh

A given version rather than the latest:

curl -fsSL https://raw.githubusercontent.com/alpine-ronin/airlock/main/install.sh | sh -s -- 0.1.0

To update a binary installed this way, run airlock update: it downloads the latest release's binary and checks it as the install script does, then replaces itself (airlock update 0.6.0 installs a given version). An update is not counted as an install. A server already running keeps working on the version it started with: restart your MCP client to use the new one. From 0.6.1 or older, with the GitHub CLI logged in, airlock update refuses 0.7.0 and later: those Airlock versions only accept the attestations made before the repository moved. Run the install command above once instead; airlock update works from 0.7.0 on. From a clone, update with git pull and composer install; with Docker, git pull in the clone, then build the image again.

Counting installs

On a first install, when no airlock is in the install directory yet, the script also downloads first-install.txt from the installs release, and says so. Its download count, the installs badge above, is how many people installed Airlock; an update or a reinstall over it is not counted. Nothing else is sent: GitHub counts the download, as it counts the binary's. AIRLOCK_NO_COUNT=1 skips it. Each day the installs workflow writes that number and each release's binary downloads (updates and CI included) to its run summary.

Or by hand: download airlock-<platform> and SHA256SUMS from the release, check that this prints the line SHA256SUMS has for it, then make it executable. On Linux, install bubblewrap as below; macOS has sandbox-exec built in. The macOS binary is not signed by Apple yet: downloaded with a browser, Gatekeeper refuses its first run. Clear the download mark with xattr -d com.apple.quarantine airlock-macos-aarch64, or allow it in System Settings › Privacy & Security › Open Anyway. The install script downloads with curl, which sets no such mark.

shasum -a 256 airlock-<platform> && chmod +x airlock-<platform>

Each binary is attested from 0.3.4 on: GitHub signs where it was built (this repository, its tag and commit, the release workflow) and what it holds, the PHP packages listed in airlock.cdx.json, the release's CycloneDX SBOM. The static PHP runtime and its extensions are not in the SBOM. With the GitHub CLI:

gh attestation verify airlock-<platform> --repo alpine-ronin/airlock --signer-workflow alpine-ronin/airlock/.github/workflows/binaries.yaml --source-ref refs/tags/v<version>

Releases up to 0.6.1 were built before the repository moved to Alpine Ronin: check them with --owner hbenabdallah --signer-workflow hbenabdallah/airlock/.github/workflows/binaries.yaml instead. The same command with --predicate-type https://cyclonedx.org/bom checks the SBOM attestation.

The server command:

~/.local/bin/airlock mcp:server

Natively, with PHP 8.5

git clone https://github.com/alpine-ronin/airlock.git ~/airlock
composer install --working-dir ~/airlock

On Linux, install bubblewrap (macOS has sandbox-exec built in):

sudo apt-get install -y bubblewrap

On Ubuntu 23.10 and later, AppArmor forbids the unprivileged user namespaces bubblewrap needs. Allow them (this is system-wide; put it in /etc/sysctl.d/ to keep it after a reboot):

sudo sysctl -w kernel.apparmor_restrict_unprivileged_userns=0

The server command:

php ~/airlock/bin/console mcp:server

With Docker (no PHP needed)

git clone https://github.com/alpine-ronin/airlock.git ~/airlock && docker build -t airlock ~/airlock
mkdir -p ~/.config/airlock

On an Ubuntu 23.10 or later host, also allow the user namespaces bubblewrap needs, as above: the restriction reaches into containers.

The server command, with your project's absolute path in place of /path/to/project (twice) and your home directory in place of /home/you:

docker run --rm -i -u "$(id -u):$(id -g)" \
  --security-opt seccomp=unconfined --security-opt apparmor=unconfined --security-opt systempaths=unconfined \
  -e AIRLOCK_AI_BASE_URL -e AIRLOCK_AI_API_KEY -e AIRLOCK_OUTPUT_CHECK_MODEL -e AIRLOCK_RISK_MODEL \
  -e XDG_CONFIG_HOME=/airlock-config -v /home/you/.config/airlock:/airlock-config/airlock \
  -v /path/to/project:/path/to/project -w /path/to/project airlock
  • Commands then run inside the container, with its tools (bash, git, PHP 8.5, coreutils), on your project mounted at the same path. That is enough to try Airlock; for your own work, where commands need your tools, install it natively or as a binary.
  • The three --security-opt let bubblewrap create its namespaces inside the container: Docker's defaults forbid it. They loosen Docker's confinement of the container, while

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars3
CategoryDevelopment
Updated15h ago
Forks0

Languages

PHP

Trust signals

92/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 low