CI integration

Add agentgate ci to your pipeline so a rug-pulled tool description or a severe finding fails the build before your agent ever sees it.

Prerequisites in the repo being gated:

1. A committed agentgate.lock (run agentgate lock locally and review it in a PR).
2. An MCP config the runner can read — pass it explicitly with --config for reproducible CI runs.

AgentGate is published on npm as mcp-agentgate (the installed command is still agentgate), so every recipe below is a one-liner with npx mcp-agentgate.

GitHub Actions

Using the bundled composite action:

name: mcp-gate
on: [push, pull_request]

jobs:
  gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: wookat/agentgate/packages/action@main
        with:
          command: ci
          args: --config .mcp.json --fail-on high
          lockfile: agentgate.lock

To surface scan findings in GitHub code scanning, set sarif-file: agentgate.sarif on the action and add:

      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with:
          sarif_file: agentgate.sarif

GitLab CI

mcp-gate:
  image: node:22
  script:
    - npx mcp-agentgate ci --config .mcp.json --fail-on high

CircleCI

version: 2.1
jobs:
  mcp-gate:
    docker:
      - image: cimg/node:22.17
    steps:
      - checkout
      - run:
          name: MCP gate
          command: npx mcp-agentgate ci --config .mcp.json --fail-on high
workflows:
  gate:
    jobs: [mcp-gate]

Jenkins

pipeline {
  agent { docker { image 'node:22' } }
  stages {
    stage('MCP gate') {
      steps {
        sh 'npx mcp-agentgate ci --config .mcp.json --fail-on high'
      }
    }
  }
}

Azure Pipelines

pool:
  vmImage: ubuntu-latest
steps:
  - task: UseNode@1
    inputs:
      version: 22.x
  - script: npx mcp-agentgate ci --config .mcp.json --fail-on high
    displayName: MCP gate

Tips

• Pin the config: auto-discovery looks at the runner’s home directory — in CI, always pass --config pointing at a config committed to the repo.
• Choose the threshold deliberately: --fail-on high (default) blocks on high/critical findings; use --fail-on medium once your baseline is clean.
• Stdio servers run in CI: ci/diff/lock connect to stdio servers to read their live tool surface, so the runner needs any runtimes those servers require (e.g. uv for Python servers). Restrict with --server if only some servers matter.
• Exit code 2 means the gate itself broke (missing lockfile, unreachable server) — treat it as a failure, don’t mask it.