actions: Run Outfitter Agents in GitHub Actions

Overview

ai-outfitter/actions connects GitHub Actions workflows to Outfitter, which assembles an agent profile and launches its CLI in headless mode. Each workflow run performs a bounded unit of work and exits, making it possible to attach an agent to repository events without hosting a persistent service.

It is intended for teams that want to build their own event-driven or scheduled reviewer and task agent. The repository provides an action, example workflows, and an Outfitter skill for designing agentic workflows. It does not supply an agent profile or provider credentials; those are configured by the workflow and catalog.

Key Features

• Runs agents from a workflow prompt and an Outfitter catalog, with support for pi, claude, and codex harnesses.
• Resolves agent selection from an input, an issue or pull request label, or the catalog default.
• Supports scheduled, pull request, path-based, assignment, and manually dispatched workflows through reusable examples.
• Can make Chromium available to browser MCP servers for preview-environment reviews.
• Saves a self-contained HTML session transcript as a workflow artifact when using pi.
• Exposes controls for the GitHub token, Git identity, working directory, catalog version, and strict preflight behavior.
• Includes guidance on least-privilege tokens and handling prompt injection and other untrusted repository content.

Use Cases

• Repository maintainers can trigger reviews when a pull request is ready for review or when changes touch sensitive paths.
• Teams can schedule an agent to review commits that landed since its previous run.
• A project can assign issues or pull requests to a bot account and have an agent work on the task.
• Developers can dispatch an implementation workflow on demand, including from an issue-triage workflow.
• Teams with a deployed preview environment can ask an agent to inspect it using a browser MCP server.

Getting Started

Add ai-outfitter/actions@v1 to a GitHub Actions workflow, provide a required prompt, select an agent and catalog as needed, and pass provider credentials through step environment variables. For example:

- uses: ai-outfitter/actions@v1
  with:
    agent: reviewer
    prompt: Review the changes in this pull request and post your findings.
  env:
    ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

See the repository README and examples/ for workflow setup and input details.

Alternatives

• run-gemini-cli: Runs Gemini CLI in GitHub Actions for coding tasks and issue or PR help, while actions runs configurable Outfitter profiles as headless jobs.

Considerations

• You need an Outfitter profile or catalog, an agent harness, and credentials for the selected model provider. These are not provided by the action.
• An agent can access a shell and the token supplied to its job. Use explicit least-privilege workflow permissions, and avoid human personal access tokens.
• Pull request diffs and issue text are untrusted inputs. The README advises against running the action on pull_request_target with a write token against untrusted fork code.
• Pin source-ref for catalogs you do not own. Catalog profiles can affect agent configuration and launch behavior.
• Browser access requires a compatible browser MCP server and careful handling of sandbox settings, especially on root-runner containers.
• The repository reports 2 stars and 5 open issues, so users should review its current implementation and issue tracker before relying on it in sensitive workflows.