by Shalimov04
MCP Airlock is a proxy placed between an AI agent and an MCP server. It adds per-caller tool permissions, dry runs, human confirmation for risky calls, and an audit trail.
PluginContact for pricing
Step 1: Install with uvx mcp-airlock --help or pip install mcp-airlock, adding extras such as postgres, otlp or regex if needed.
Step 2: Write a policy.yaml that lists the tools and their risk tiers per environment.
Step 3: Start the proxy with --policy, --upstream and --env, setting AIRLOCK_TRUST_PRINCIPAL_HEADER=1 on a laptop or using a JWT elsewhere.
Step 4: Point your agent at the proxy instead of the server; docs/clients.md covers Claude Code and Cursor.
Step 5: Check the audit file and run airlock-policy lint on your policy.
An L2 call runs first as a dry run and comes back as input_required with a signed requestState. After a person approves, the agent repeats it and it executes once.
From a JWT bearer token or an X-Airlock-Principal header from a trusted gateway, never from the request body. With no principal, the call is refused.
Yes. The image ghcr.io/shalimov04/mcp-airlock includes all three extras and runs as a non-root user.
It marks suspected injection phrases but does not change the content. It does cut responses to the output cap and hides tools the policy does not list.
Requirements
IronClaw MCPRun untrusted agent shell commands inside a gVisor sandbox.Proxy AIOps MCPProxy AIops is an MCP server and CLI for audited operations on Traefik, Caddy and HAProxy reverse proxies. The same 28 tools work across all three.