// the take
gh-aw is GitHub's own CLI extension for defining AI agent automation (issue triage, PR review, CI failure investigation, etc.) as Markdown files with YAML frontmatter, which it compiles into standard GitHub Actions workflows. It's for teams who already run CI/CD on Actions and want to add LLM-driven repo tasks without hosting a separate agent runtime.
Compiles to a plain `.lock.yml` Actions workflow, so there's no hosted service or opaque runtime to trust — you can read exactly what will execute before it runs. The safe-outputs design is genuinely thought through: the agent job itself is sandboxed and read-only by default, and any actual GitHub writes go through a separate job with its own scoped permissions rather than the agent having direct write access. It's engine-agnostic (Copilot, Claude Code, Codex, Gemini, Pi), so teams aren't locked into one vendor's CLI. Backed by GitHub directly with real adoption (614 forks, an active external contributor base).
It already shipped a disclosed vulnerability serious enough to retire a version range (0.83.3–0.85.4) — handing an LLM any path to repo writes is inherently high-risk, and this is evidence it's been exploited or found exploitable once already. The project surface is large and fast-moving: sandboxing, a firewall layer, an MCP gateway, a work-queue system, and five+ AI engine integrations, all versioned somewhat independently, which is a lot to keep secure and in sync. The README's own security section is configurable rather than enforced, and explicitly says 'use with caution, at your own risk' — the safe defaults help, but getting permissions/tools/network access right per workflow is still on you. Tying workflow behavior to the churn of multiple third-party AI CLIs (Copilot, Claude Code, Codex, Gemini) means compatibility breakage is an ongoing tax, not a one-time setup cost.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →