Human Review: A Local Visual Editor That Batches File Feedback for Your AI Agent
A visual tool to edit HTML and Markdown files, leave comments like a Google Doc, and send feedback to your AI agent. Works inside your favorite AI harness.
At a glance
- What is it?
- Human Review is a local browser-based editor for HTML and Markdown files that collects your direct edits and anchored comments into a single batch for an AI coding agent. The design is sound for reviewing AI-generated documents; the interesting trade-offs are in how feedback persists and what happens when a review session ends.
- Who is it for?
- Human Review is worth adopting if you regularly review AI-generated Markdown plans, landing pages, or localhost apps and find chat-based feedback slow or imprecise. It is not the right tool if you need multi-user review, cloud sync, or a hosted collaboration layer; everything runs locally with no account or database.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 11 days ago.
- What is it written in?
- Mainly JavaScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The chat-feedback problem Human Review targets
When you ask an AI agent to revise a document, the feedback loop runs through chat. You describe what to change in prose, the agent edits the file, and you re-read the output to check whether each instruction landed. The README calls this out directly: the alternative to the tool is typing something like "change X to Y" in a paragraph reference, then verifying the agent understood every part. That works for one or two edits. It breaks down on long plans, multi-page sites, or anything where you want to change a sentence yourself rather than describe the change. Human Review's premise is that direct manipulation plus anchored comments, batched into one send, removes the translation step. The tool is built for people already using Claude Code, Codex, or a similar harness and reviewing generated files, not for teams doing collaborative document editing.
How the local review session is wired together
The repository splits responsibilities across five files. cli.js implements the human-review, poll, status, and setup commands. server.js runs the local review session. sdk.js handles editing, comments, highlights, and feedback. chrome-client.js is the visual interface you interact with in the browser. markdown.js renders Markdown files for review. SKILL.md is what teaches the agent how to participate. The data flow, based on the README, is: the agent invokes the skill against a file or localhost URL, the browser opens the rendered page, you edit and comment, and on Send the collected feedback goes to the waiting agent, which updates the source and refreshes the page. The README states that in Claude Code the agent waits in the background and picks up feedback the moment you click Send, while in Codex and other agents it waits during its turn; if the turn already ended, you send a message and it picks the feedback up. Closing the tab or clicking End review releases the agent either way. That distinction between harness behaviors is the most operationally important detail in the README and the one most likely to surprise new users.
Installing the skill and opening a review target
Two installation paths are documented. The first is pasting a prompt into your agent: "Install the /human-review skill globally from https://github.com/petergyang/human-review". The second is running npx -y human-review setup --global from a shell. Once installed, you invoke it against a file or a running page: /human-review (your file) for an HTML or Markdown document, or /human-review (localhost URL) for a page served locally. The README does not document additional configuration keys, environment variables, or a config file, so there is nothing to tune beyond the global install flag. The tool requires no account, cloud service, database, or API key according to the README, which keeps the setup surface small but also means there is no server-side component to configure or secure.
What persists after you close the tab
Persistence differs by file type, and the README is explicit about this. For HTML files, direct edits and resizes save automatically, so closing the tab does not undo them; you can use Discard on your next open or Revert all during the review to put the file back. For Markdown and localhost pages, edits are not applied to source until you click Send, at which point the agent applies them. Unsent feedback is retained: the README states that feedback you never sent is kept, and the next time you open that page you can restore or discard it. Comments are editable after writing, including ones already sent; a reworded comment reaches the agent with the next Send. This asymmetry between HTML and Markdown is a deliberate design choice, not an oversight, but it means your mental model of "did my edit stick" has to change depending on what you are reviewing.
The editing surface and its interaction model
The README lists the supported operations: direct text editing with basic formatting such as bold and italic; bulleted and numbered lists via typing "- " or "1. " at the start of a line or pressing Command-Shift-8 and Command-Shift-7; Tab and Shift+Tab to indent and outdent; links via Command-K on selected text, with Command-K inside an existing link to edit or remove it; image resizing by dragging a corner and repositioning by dragging; block rearrangement by hovering and dragging the handle on the left edge; pasting images from the clipboard, which for file reviews are saved beside the document and for localhost reviews are staged for the agent to place in the app source; comments anchored to selected text or attached to an element by clicking it; element deletion with Undo; and Command-click on links to review multiple pages without losing feedback. The scope is intentionally shallow on formatting. This is a review surface, not a document editor, and the feature list reflects that.
Where the design creates friction
The harness-dependent wait behavior is the main failure mode. If you are in Codex or a similar agent and its turn has already ended, clicking Send does not deliver feedback until you send a follow-up message. The README describes this plainly, but it means the round trip is not uniform across agents, and a user who expects Claude Code's background pickup in every harness will be confused. The second friction point is the Markdown and localhost path: because edits only reach source through the agent after Send, a closed tab with unsent edits leaves the file unchanged, and you are relying on the restore prompt on next open. The third is that there is no multi-user layer. Everything runs on your computer, which is the point, but it also means no shared review state, no comment threads between people, and no audit trail beyond what your agent harness keeps. For a solo reviewer working with an agent, these constraints are acceptable. For a team, they are disqualifying.
How it compares to editing in an IDE or a hosted doc tool
The closest alternative is editing the file directly in your editor and then describing the changes to the agent in chat. That approach gives you full editor power (multi-cursor, search and replace, extensions) but loses the anchored-comment model and the single-batch send. Human Review trades editor depth for a review-specific interaction: comments tied to exact text or elements, and one Send that delivers everything at once. A hosted document tool like Google Docs gives you the comment and suggestion model across multiple people, but it cannot write back to your source file or hand a structured batch to an agent. Human Review sits between those two: more structured than chat, more connected to source than a hosted doc, and deliberately narrower than a full editor. The trade-off is that anything outside its supported operations, such as complex tables or custom HTML components, has to be handled another way.
Maintenance, licensing, and what to check before adopting
The project is MIT licensed, which permits commercial use, modification, and redistribution with the license and copyright notice preserved; that is a permissive baseline, and nothing in the README suggests additional terms. This is not legal advice, and if you are embedding it in a product you should read the LICENSE file yourself. On maintenance: the release history shows v0.8.1, v0.8.0, and v0.7.2 all dated 2026-09-04, with release notes describing incremental work such as diagrams, a Back control, drop behavior, undo, and counting only what the agent does not already have. That cadence suggests active iteration, but the README does not describe a deprecation policy, a versioning contract, or a test suite, so upgrade cost is hard to estimate from the material. The practical check before adopting is to read src/SKILL.md, since it defines what your agent is expected to do with the feedback, and to confirm on your own harness whether feedback is picked up after a turn ends. If you rely on Markdown review, test the close-without-Send path once so the restore behavior is not a surprise.
Editorial conclusion
Human Review is worth adopting if you regularly review AI-generated Markdown plans, landing pages, or localhost apps and find chat-based feedback slow or imprecise. It is not the right tool if you need multi-user review, cloud sync, or a hosted collaboration layer; everything runs locally with no account or database. Before committing, verify two behaviors on your own setup: whether your agent harness actually picks up feedback after its turn ends, and whether Markdown edits survive if you close the tab without clicking Send. The SKILL.md file is the contract between the tool and your agent, so read it before trusting the round trip.
Community notes