Review agent output like you review pull requests

What it does

Crit is a single-binary CLI that opens a local review interface for anything your AI agent produces. It renders markdown plans with proper formatting, syntax-highlights git diffs, and can even proxy a running dev server so you can click around and annotate the UI. Once you leave inline comments, the agent reads them and edits the code, creating an actual feedback loop instead of a one-way dump.

The interesting bit

The tool treats different outputs as different review surfaces—markdown files get a document UI, code gets a diff viewer, and web apps get an overlaid review layer via a proxied iframe. That adaptive approach, plus the ability to start real-time “live threads” with the agent, makes it less a viewer and more a shared workspace between human and machine.

Key highlights

• Adaptive UI for markdown plans, git diffs, static HTML, and proxied live web apps.
• Inline comments and round-to-round diff tracking as the agent edits.
• Programmatic CLI reviews and optional async sharing via crit.md.
• Bidirectional GitHub PR sync to pull existing comments or push new ones.
• Experimental real-time “live threads” that let an agent read, edit, and reply while you review.

Caveats

• Live mode proxies your app through an iframe on a different origin, so host-scoped session cookies do not transfer automatically; you must forward them manually or via a cookie file.
• The real-time “send to agent” feature is experimental and requires configuring an agent_cmd in your global config file.
• GitHub PR sync depends on having the GitHub CLI (gh) installed and authenticated separately.

Verdict

Worth a look if you are tired of copy-pasting AI output into GitHub PRs just to review it, or if you want to keep an agent in a tight local feedback loop. Less useful if you already have a review workflow you love, or if your agent never produces anything complex enough to warrant line-by-line commentary.