Doc Review

Comment on your documents the way you do in Google Docs, then let an AI handle your comments. Open a folder, read a .md / .txt / .tex file in a clean page, select text, leave a margin comment, press Send to AI. The AI edits the files and answers in each thread; you review its changes as a diff and accept or revert each one.

Free, open source (MIT). No account, no telemetry, nothing is uploaded except what you send to the AI provider you choose. Your files stay on your computer.

Two ways to use it

Related MCP server: overleaf-mcp

Desktop app (no technical knowledge needed)

1. Go to the Releases page and download Doc Review for your computer: the .dmg for Mac, the .exe for Windows.

2. Install it. The app is not signed yet, so your computer will warn you once:

   • Mac: open the .dmg, drag Doc Review to Applications, then right-click it, choose Open, and click Open again.

   • Windows: if "Windows protected your PC" appears, click More info, then Run anyway.

3. Open it, choose the folder with your documents.

4. Click AI not set up (top right), pick a provider, paste your key, press Test connection.

Where to get an AI key: OpenAI, Anthropic or OpenRouter. A ChatGPT or Claude chat subscription is not an API key; the API is billed separately, by use. Free option: install Ollama, pick "Ollama" in the settings. Quality depends on the model.

The AI can read and edit only files in the folder you opened, and only with exact find-and-replace edits. It cannot run programs or browse the web. Every change shows up in the Changes view, where you can revert it. Settings include a per-round spending limit, and the pill has a Stop button while it works.

MCP server (Claude Code, Cursor, others)

Needs Node.js 18+ and git. Download once, then register it (two lines):

git clone https://github.com/MichaelEight/doc-review.git ~/doc-review
claude mcp add --scope user doc-review -- node ~/doc-review/bin/doc-review.js --mcp

No git? Download the ZIP from the green Code button, unzip it, and use that folder's path instead of ~/doc-review. Update later with git -C ~/doc-review pull.

Other MCP clients: add this to their config (use the full path, not ~).

{ "mcpServers": { "doc-review": { "command": "node", "args": ["/full/path/to/doc-review/bin/doc-review.js", "--mcp"] } } }

Optional skill (teaches Claude Code the review loop; the server's own instructions already cover the basics):

ln -s ~/doc-review/.claude/skills/doc-review ~/.claude/skills/doc-review

Then ask your AI: "open this folder in doc-review". The AI gets a link; you comment; Send to AI delivers the comments to it. When an MCP agent has a folder open, it takes priority over the built-in AI.

Run from a terminal (no desktop app)

node ~/doc-review/bin/doc-review.js ~/path/to/folder          # standalone, with the built-in AI
node ~/doc-review/bin/doc-review.js --serve ~/path/to/folder  # UI only; comments wait for an MCP agent

DOC_REVIEW_PORT sets the first port to try (default 4848).

Known limits

• Not signed yet. Hence the one-time warning on install. Signing needs a paid Apple developer account and a Windows certificate; planned.

• API key is stored in plain text in ~/.config/doc-review/ai.json (readable only by you), not in the system keychain. It is never sent to the browser page or written into your documents' folder.

• Works offline except for the AI call itself: KaTeX and PDF.js are bundled in web/vendor/ (see THIRD-PARTY-NOTICES.md).

• Word files (.docx) are not supported yet. See PLAN-standalone.md.

• LaTeX preview needs latexmk or pdflatex installed (MacTeX / BasicTeX / MiKTeX). Without it the PDF tab is hidden; commenting on the source works.

• Tested on macOS. Windows and Linux are covered by automated tests in CI but have had no manual testing.

How it works

• Server (src/): MCP over stdio, plus a local HTTP UI on 127.0.0.1. Zero runtime dependencies, Node 18+. src/ai.js is the built-in AI: it waits for Send like an MCP agent would, then runs a tool-calling loop against any OpenAI-compatible /chat/completions endpoint.

• UI (web/): vanilla JS. Markdown is rendered with source positions so comments anchor to exact text.

• Desktop app (app/): a thin Electron window around the same server, plus a native folder picker. Build locally with npm install && npm run dist; releases are built by GitHub Actions on v* tags.

• Skill (.claude/skills/doc-review/SKILL.md): the review loop for MCP agents.

• Data: <folder>/.doc-review/comments.json (threads, rounds, tags, display rules) and baselines.json (what the diff compares against). Add .doc-review/ to the folder's .gitignore unless you want to keep the review history.

• Token: the UI link carries a secret kept in ~/.config/doc-review/token (mode 600; override with DOC_REVIEW_TOKEN_FILE). Delete the file to revoke old links.

Using it

Commands

Start a comment with / to pick a command. The menu shows each command's parameters and examples. The command applies to the selected text; for a click comment, to the paragraph around the spot; add all (for /fix) to cover the whole file.

Editing is locked while the AI is working (comments still work) and while there are AI changes waiting for review. This keeps your edits from mixing into the AI's diff.

Display rules

Colouring and notes that change how a file looks, never what it holds. You ask for it in a comment, in plain words: "colour this nick blue", "grey out everything in <>", "dim the timestamps in the whole logs folder", "show me how long passed since the previous line". The AI turns that into rules with review_style; the document on disk is untouched.

A rule is a scope, a match, and then a style, a note, or both:

{delta} needs annotate.delta, which reads each match as a time: 2013-11-15 16:02:59, 2013-11-15T16:02, or a bare 16:02:59 that wraps at midnight. It is written the way it reads out loud, +0h 0m 16s, with a day count only when there is one. The first match has nothing to compare against, so it gets no note.

Notes are decoration. They are never written to the file, cannot be selected or copied, and appear in the reading view only: the source view's overlay has to line up with the textarea underneath it character for character, so nothing that is not in the file can be drawn there.

Rules show up in a Display rules panel at the bottom of the sidebar, listing the ones that apply to the open file with a preview of each. The checkbox switches a rule off without deleting it, so you can ask for a colouring, try it, and turn it off again. They apply in both Read and Source, and overlapping rules stack with the later one winning.

Recipes

The ☆ button above the file list opens Display recipes: the half-dozen rules people ask for most, already written out. Pick one, fill in the blanks, tick the files or whole folders it should cover, edit the wording if you want, and it goes in as an ordinary /style comment that the next round picks up. Add and send delivers it straight away.

Folders are ticked as a whole, so a folder whose files are all picked becomes one glob: forty paths turn into logs/*, and the rule keeps working when the folder grows.

Adding a recipe is data, not code. An entry in RECIPES in web/recipes.js:

{
  id: "time-gap",
  name: "Time since the previous line",
  blurb: "Writes +0h 0m 16s after every timestamp",
  fields: [{ name: "what", label: "What to fade", placeholder: "the timestamps" }],
  prompt: "Add a display rule that dims {{what}}.",
}

{{field}} is replaced by what the user typed, or by that field's default, or by its placeholder. Single braces are left alone, so {delta} reaches the AI untouched. The scope line is appended for you, so a prompt should never say which files to change. A test walks every recipe and fails if a field is unused, a placeholder survives, or a prompt writes its own scope.

They live in .doc-review/comments.json alongside the comment threads, so they persist and are shared between the UI and the agent. They are per workspace, not per user.

Tools

How anchoring works

A comment stores the quoted text and 32 characters of context on each side, not an offset (web/anchor.js). When a file changes, whether typed in the UI or rewritten by the AI on disk, the server maps each thread through a line and word diff of the old and new text (web/remap.js), so comments follow their text. A comment whose text is gone shows under "Text no longer found".

Tests

npm test