tomasz-tomczyk/crit

Your feedback loop with the agent

What Crit does

Crit is a local, single‑binary tool that lets you review any kind of text output that an AI agent produces – plans, code diffs, or even a running web UI – and comment on it directly. Those comments are stored locally and can be sent back to the agent so it can fix the issue, creating a fast “review‑and‑revise” loop.

Core workflow

  1. Run the tool – crit automatically detects changed files in a Git repo, or you can point it at a specific markdown file, an HTML page, or a live web server (crit http://localhost:3000).
  2. Review UI – The UI shows the file with syntax highlighting, a split or unified diff, and lets you click line numbers to add inline comments (single lines or ranges).
  3. Send feedback to the agent – When a comment is marked “Send now”, Crit pipes the comment, its context, and the file location to a configured AI agent (Claude, Aider, Cursor, etc.). The agent’s response (or edited files) is captured and displayed inline.
  4. Iterate – The agent edits the file, Crit watches for changes, updates the diff, and you repeat until the review is approved.

Highlights

  • Adaptive UI for markdown, code diffs, static HTML, or live web apps (via a proxy).
  • Round‑to‑round diff – see exactly what changed after each agent edit.
  • Programmatic comments – agents can add comments via crit comment … without opening a browser.
  • Share reviews – generate a public URL (or QR code) so others can view or add comments; self‑hostable via crit-web.
  • GitHub PR sync – pull existing PR comments into a Crit review and push new comments back to the PR (requires the GitHub CLI).
  • Per‑branch isolation – each Git branch gets its own review file stored under ~/.crit/reviews/.
  • Live mode – proxy a running dev server, forward cookies, and review the UI as you develop.
  • Configurable via global (~/.crit.config.json) and project (.crit.config.json) JSON files; many flags can also be set with environment variables.

Who would use it?

  • Developers who let LLMs (Claude, Cursor, Aider, etc.) write code or plans and need a human‑in‑the‑loop review step.
  • Teams that want a lightweight, local alternative to PR‑style code review for AI‑generated changes.
  • Anyone building AI‑assistant‑driven workflows who wants a visual, comment‑driven feedback channel without leaving the terminal.

Installation & quick start

# Homebrew (macOS/Linux)
brew install crit
# Or via Go, Nix, Windows binary, or direct download

Then integrate with your agent (e.g., claude plugin marketplace add tomasz-tomczyk/crit) and invoke the /crit slash command from the agent, or run crit manually.

In a nutshell: Crit is a locally‑run review UI that bridges human feedback and AI agents, turning plain‑text outputs into an interactive, comment‑driven workflow.