Crit.md Deep Dive — A Local-First Review Tool That Gives AI-Coding-Agent Output the "Pull-Request Review" Experience, With First-Class Support for Claude Code / Cursor / Copilot / Codex / Gemini / Qwen Four Review Modes (Plans & Docs, Code Diffs, Live Apps, Static HTML), Per-Line Comments Persistent Across Iteration Rounds

Crit.md is a local-first review tool that provides a pull-request-style review experience for AI-coding-agent output.

Four takeaways:

1. Four review modes in one UI — Plans & docs / Code diffs / Live apps / Static HTML preview
2. Per-line comments that persist across iteration rounds — "comment on line 47 → agent fixes it → the comment is still tied to line 47 next round"
3. Local-first — no account, binds to 127.0.0.1, zero telemetry
4. Multi-agent support — first-class plugins for Claude Code / Cursor / Copilot / Codex / OpenCode / Gemini / Qwen; file-based protocol means no vendor API and no lock-in

Modern AI coding agents propose large amounts of change at once. The usual responses:

- Scroll a 100-file diff in the terminal — impractical
- "Yolo accept" — merge without close review; regressions surface later
- "Have the agent review it" — reviewer is also an agent; humans drop out

Crit.md's answer: check agent output line-by-line like a GitHub PR review in the browser, with comments stored in a format the agent can read and fed back into the next iteration.

(1) Plans & docs: cleanly rendered Markdown for the agent's plans and doc drafts, with per-line comments.

(2) Code diffs: syntax-highlighted diffs, essentially the GitHub PR Files tab. Navigate across files, comment on changed regions.

(3) Live apps: comment directly on the DOM — Crit.md talks to the dev server the agent started, and you click on UI elements to leave comments ("make this button bigger," "change this color").

(4) Static HTML preview: browser-preview an agent-generated static HTML file, comment per element. Great for landing-page and email-template review.

The core mechanic: comments live in structured files the agent can read, so on the next iteration:

- The agent detects unresolved comments and works them
- After fixes, comments stay attached to the same line
- The set of "still-unresolved comments" stays visible

Just like a GitHub PR review — resolved / unresolved tracking and cross-round discussion continuity.

First-class plugins:
- Claude Code
- Cursor
- GitHub Copilot
- Codex
- OpenCode
- Google Gemini
- Alibaba Qwen (Qwen 3.6-35B)
- Others

File-based protocol — instead of a proprietary API, Crit.md communicates with agents via shared file formats (structured Markdown / JSON). Agents don't ship a Crit.md SDK; they just read and write review files inside the project directory.

Benefits: no agent-side changes are required, and future agents (including those built on Gemma 4) are supported automatically.

Design principles:

- No account — no login screen, launches immediately
- 127.0.0.1 bind — access is local-only, never reachable from outside
- Zero telemetry — usage isn't reported anywhere
- Source code doesn't leave the local machine — easy to adopt in enterprise environments

The agent itself may still talk to Anthropic / OpenAI / etc. — that's on the agent's terms — but Crit.md itself is fully local. Sails through corporate security review.

The two currents in 2026 AI-agent tooling:

1. Fully autonomous: Claude Cowork, Devin, OpenAI Operator — the agent finishes the task; the human receives the artifact
2. HITL (human-in-the-loop) assistants: Crit.md, Cursor, Claude Code — the agent proposes; the human reviews and adopts

Crit.md maxes out the HITL current. It tightens the loop of "agent proposes → Crit.md line-review → comments → agent fixes → Crit.md re-review."

The strategic read: as fully autonomous agents proliferate, HITL review quality determines aggregate quality. Crit.md doesn't aim to boost AI productivity — it aims to boost AI quality.

Individual developers: keep Crit.md in a browser tab while running Claude Code / Cursor and review each iteration there. Swap terminal-diff eyeballing for the Crit.md UI.

Team development: put Crit.md between agent PRs and merge — comment → fix → merge. Standardize a quality bar for agent output across the team.

Enterprise: local-first design smooths security review, and combined with existing Git-PR workflows Crit.md becomes a gate for agent output.

Crit.md is a genuinely useful review substrate for the AI-coding-agent era. As fully autonomous agents accelerate through late 2026, HITL review quality shapes product quality. The combination of a four-mode unified UI, persistent per-line comments, multi-agent support, local-first operation, and a file-based protocol is a clean way to extend the GitHub PR experience to AI agents.

Related services from us — software development, AI consulting, and OpenClaw setup. For help designing enterprise adoption of AI coding agents or a review workflow, get in touch.

- Crit.md
- Related columns: Claude Cowork Web/Mobile expansion, Claude Fable 5 return and Devin/Cursor, Cursor iOS support, Qwen 3.6-35B, Gemma 4