tomasz-tomczyk/crit

Review AI coding agents' plans, diffs and running apps in the browser. Local-first, works with any agent.

Crit – Interactive Review UI for AI‑assisted coding

What it is – A single‑binary, locally‑run application that lets you (and any LLM‑powered coding agent) review plans, diffs, or live web apps in a browser. You can click any line, add inline comments, and the agent can read those comments, edit code, and reply in real time.

Why it matters – While many AI‑code tools generate code, Crit focuses on the feedback loop: a UI that mirrors GitHub‑style reviews for any text output (markdown plans, git diffs, static HTML, or a running dev server). It bridges the gap between an LLM’s textual output and a human’s line‑by‑line review, making the edit‑review‑repeat cycle faster and more visual.

Quick start

Install – brew install crit (or go install, Nix, or download the binary from the releases page).
Hook an agent – Add the Crit plugin to Claude Code (claude plugin marketplace add tomasz-tomczyk/crit) or use any of the supported agents (Cursor, GitHub Copilot, Aider, Gemini, etc.).
Run – In a terminal inside a repo, simply run crit. It auto‑detects changed files and opens a local web UI. You can also target a specific file (crit plan.md), a running server (crit http://localhost:3000), or a static HTML page.

Core features

Adaptive UI – Different renderers for markdown plans, git diffs, live web apps, and static HTML.
Round‑to‑round diff – After the agent edits a file, Crit shows a split or unified diff of what changed.
Inline comments – Click a line or drag a range to add a comment, just like a GitHub PR review.
Programmatic comments – Agents can add comments via crit comment … without opening the browser.
Story mode – Generates a high‑level, chaptered overview of large PRs/branches to give context before line‑by‑line review (LLM‑driven, cost‑aware).
Live mode – Proxies a running dev server, injects a review overlay, and can forward session cookies so authenticated apps work.
Sharing – One‑click upload of a review to a public URL (crit.md or self‑hosted targets) for async collaboration.
GitHub / GitLab sync – Pull and push comments to PRs/MRs, with full support for approvals and request‑changes.
Send to agent – Click “Send now” on a comment to have a configured LLM (e.g., Claude Code) respond and optionally edit code automatically.
Per‑branch isolation – Review data lives in ~/.crit/reviews/ and is scoped to the current git branch.
Zero telemetry – No hidden analytics; the only network call is an optional version‑check.

Typical workflow

Agent writes a plan or code change.
You invoke /crit (or run crit locally). The UI opens showing the plan, diff, or live app.
You click lines, add comments, or use the “Send to agent” button.
The agent reads the comment, edits files if permitted, and replies.
Repeat until the review is approved, then optionally crit push to sync comments back to the PR/MR.

Integration points

Claude Code, Cursor, GitHub Copilot, OpenCode, Aider, etc. – All can invoke Crit via a slash command or by running the binary.
CLI‑only – crit comment, crit share, crit pull, crit push, crit live, crit status, crit resume, crit stats, crit cleanup.
Config – Global (~/.crit.config.json) and project (.crit/config.json) JSON files control live‑mode cookies, share targets, default forge, and the agent_cmd used for “Send to agent”.

Who should use it

Developers who already use LLM coding assistants and want a richer, visual review loop.
Teams that need to review AI‑generated plans or code before merging.
Solo engineers who like the GitHub‑style comment experience without leaving their local machine.

Limitations / notes

All code and review data stay local unless you explicitly share.
“Send to agent” requires you to trust the configured agent_cmd; only a global config can set it to avoid malicious repo‑level execution.
Live‑mode may need extra cookie handling for apps with authentication.

Resources

Demo video – https://www.youtube.com/watch?v=LHwfdvePf5A
Documentation – docs/story-mode.md, docs/live-mode.md, docs/renderers.md
Source – https://github.com/tomasz-tomczyk/crit

Bottom line – Crit is a practical, locally‑run UI that turns any LLM‑generated text (plans, diffs, or running apps) into an interactive review surface, letting humans and agents iterate quickly while keeping everything under the developer’s control.