Claude Code in CI — GitHub and GitLab

Claude Code in CI is where the agent stops being an interactive tool and becomes infrastructure. Three distinct documented paths:

Path	What it does
GitHub Actions	General-purpose: run Claude Code as a workflow step
GitHub Code Review	Automatic review on every pull request
GitLab CI/CD	The GitLab equivalent

GitHub Enterprise Server has its own setup page, and there is separate documentation for using Actions with cloud providers (Bedrock, Google Cloud, Foundry) as the model backend.

Permissions matter more here than anywhere

An agent in CI runs unattended, with repository credentials, on code that may have come from a fork. Three specifics:

• Untrusted input. Pull request titles, descriptions and diffs are attacker-controlled on public repositories. Prompt injection through a PR body is the obvious attack.
• Least privilege. Scope the token to what the job actually needs.
• No interactive approval. Nothing will prompt. Permissions are the whole control surface.

Routines overlap with this

Routines can also be triggered by GitHub events, and run in Anthropic’s cloud rather than your CI. If you want repository automation without a workflow file, that is the other option.

What this page does not claim

Workflow syntax, required secrets and action versions are documented upstream and change; they are deliberately not copied here.