GuideIntermediate

Claude Code in CI/CD Pipelines

Take Claude Code from your terminal into your pipeline. In this guide Claude Code runs inside GitHub Actions and GitLab CI/CD as a real agent: it reads the repository, runs the tests, follows your `CLAUDE.md`, comments on pull requests and opens branches with changes, all with nobody at the screen and with scoped permissions and budget. You start with `anthropics/claude-code-action@v1`: you install it with `/install-github-app` or by hand, choose how to authenticate (API key, subscription token, identity federation, or a cloud provider through OIDC) and use it in its two modes, the interactive one that answers `@claude` and the automation one that runs with a `prompt` on any event. You build an automated PR review with inline comments and your team's conventions, and compare it with Code Review, Anthropic's managed service. Then you leave GitHub: you run `claude -p` in GitLab and in any other CI with validated JSON output, permissions for unattended runs and caps on turns and dollars, and you write your own pipeline agent with the Claude Agent SDK in Python and TypeScript. The Messages API shows up once, as the no-agent alternative for when a single call is enough. A full module is dedicated to the agent's security in CI: least-privilege tokens, secrets and short-lived credentials, prompt injection from fork PRs, budget guards and auditing. With that in place you apply Claude to real delivery: changelog and release notes, readiness validation, human approval gates, fixing broken CI, security scanning with triage, and diagnosis after a rollback. The capstone builds the full pipeline for Colibrí, an online store with a FastAPI service, from an opened PR to production, hardened and documented.