MarkAI

Reviewing a document is easy. Getting the review applied is the part that wastes your afternoon. You read a thesis, a spec or a set of docs, and you leave forty comments — a wrong figure reference in §3.2, a paragraph that contradicts the previous chapter, a term that should be spelled the same way everywhere. Then the real work starts: every one of those comments has to be carried back into the source by hand. And that is exactly the work an AI coding agent could do for you — except no annotation tool will hand the notes over. Word comments live inside the .docx. PDF annotations live inside the PDF. Google Docs suggestions live inside Google. You can read them one at a time and retype each into a chat window, which is slower than just fixing the document yourself.

MarkAI is the missing export step. You annotate the document the way you always have — click a word, select a phrase, type a note — and MarkAI writes the whole review out as a compact JSON file built for an agent to consume: what to change, where it is (Chapter 3 / 3.3 Chunking and pre-processing) and the exact quoted text to find it by. Point your agent at the file, let it apply the fixes, and it flips each note to done in a companion status file that MarkAI reads back — so the notes panel in front of you empties itself as the work lands. Review in a UI made for reading; let the agent do the typing.

How it works

   you                          MarkAI                        your AI agent
    │                              │                                │
    │  click / select → note       │                                │
    ├─────────────────────────────►│                                │
    │                              │  <doc>.notes.json              │
    │                              ├───────────────────────────────►│
    │                              │       (source folder)          │  applies each
    │                              │                                │  note to the
    │                              │  <doc>.notes_status.json       │  real source
    │   notes flip to "done"       │◄───────────────────────────────┤
    │◄─────────────────────────────┤        status: done            │

Link a document to a source folder — the directory where the actual source of that document lives — and MarkAI keeps two files there in sync on every change:

• <doc>.notes.json — the review itself.
• <doc>.notes_status.json — id/status pairs the agent writes back as it applies each note.

No source folder? Use Export notes… in the toolbar and pick either the notes file alone or a zip containing the same pair, and hand it to the agent yourself.

The export format

Deliberately terse — a document full of notes should not eat an agent's context window:

{
  "document": "Master Thesis",
  "notes": [
    {
      "id": "9f3c1e7a4b6d40f2a1c8e5d2b7093f14",
      "status": "pending",
      "note": "This contradicts the definition given in 2.1 — pick one and use it everywhere.",
      "location": "Chapter 3 / 3.3 Chunking and pre-processing",
      "quote": "chunks are split on token boundaries"
    }
  ]
}

location is the full heading path (root → leaf), not just the nearest heading, so an agent can find the right section even when a phrase repeats across a long document. quote is the selected text verbatim, or a few words either side of the anchor for a single-point note.

The status file the agent writes back is just as small:

{
  "instructions": "For each note you have applied to the source, set its status to 'done'. MarkAI reads this file back and reflects the status in its UI.",
  "notes": [{ "id": "9f3c1e7a4b6d40f2a1c8e5d2b7093f14", "status": "done" }]
}

Features

• Markdown, PDF and DOCX in one reader, with the same annotation model across all three.
• Precise anchoring. Click a word or select a phrase; notes are anchored to the exact character offsets, not just "somewhere in this paragraph" — markers and highlights land on the glyphs.
• PDF that behaves like a PDF reader. Real text selection (pdf.js's own text layer), a chapter outline read from the PDF's bookmarks, and zoom that responds instantly at any page count: the page resizes on the spot and canvases re-render in the background, nearest to the viewport first. − / +, a fine-grained slider popover (40–400 %), Fit-width presets, and Ctrl/Cmd + wheel.
• Chapter outline on the left as a collapsible tree, starting closed, with a Collapse-all button.
• Notes panel on the right grouped into Pending / Done, with a selection mode for bulk mark-done / mark-pending / delete.
• Two-way sync. Notes are written to the source folder on every change; statuses your agent writes back are pulled in on a 5 s poll, on demand via Refresh, and before every note edit.
• Drag & drop upload anywhere on the library page.
• Local-first: a single Flask process and one SQLite file. Nothing leaves your machine.

Quick start

Requires Python 3.10+. With uv installed, there is nothing to clone and nothing to set up:

uvx markai

That downloads MarkAI into a throwaway environment, starts it on http://localhost:8765 and opens your browser. To keep it installed instead:

uv tool install markai     # or: pipx install markai
markai

Register an account on first run — it is stored in your own SQLite file, on your own machine — and upload a document.

Options

markai --port 9000        # a specific port (fails loudly if it's taken)
markai --data-dir ~/notes # where the database and uploads live
markai --no-browser       # don't open a browser window

MarkAI binds to 127.0.0.1 only, so nothing outside your machine can reach it. Everything it stores lives in one directory — %LOCALAPPDATA%\MarkAI on Windows, ~/Library/Application Support/MarkAI on macOS, ~/.local/share/MarkAI on Linux — holding app.db, uploads/ and a session key generated on first run. Delete that directory to reset the app; back it up to keep your notes. MARKAI_DATA_DIR overrides the location.

From a checkout

git clone https://github.com/follen99/MarkAI && cd MarkAI
python -m venv .venv
.venv/bin/python -m pip install -e .        # Windows: .venv/Scripts/python
.venv/bin/python run.py                     # dev server on :5000, data in ./data

Using it with an AI agent

1. In the library, click Source folder on a document and enter the path where its source lives (e.g. the repo folder containing the .md file, or the LaTeX project behind the PDF).
2. Annotate. Every note rewrites <doc>.notes.json and <doc>.notes_status.json in that folder.
3. Tell your agent something like:

   Read thesis-4.notes.json. For each note with status: "pending", find the spot using location and quote, apply the change described in note, then set that note's status to "done" in thesis-4.notes_status.json.

4. Watch the notes move from Pending to Done in the panel as the agent works.

Tech

Flask, SQLite (plain sqlite3, no ORM), server-rendered Jinja templates, and vanilla JS/CSS — no build step, no frontend framework. PDF rendering and interaction are entirely client-side via pdf.js (3.11.174, bundled with the package rather than pulled from a CDN, so MarkAI works with no network at all); the server just streams the raw bytes. Markdown is parsed with markdown, DOCX with python-docx. Serving is handled by waitress.

CLAUDE.md in the repository root is the deep architecture document — how notes are anchored, why the PDF text layer works the way it does, and which fallbacks are load-bearing. Read it before changing anything non-trivial.

Status and limitations

This is a working single-user local tool, built to be deployable later but not deployed yet:

• Built for localhost. There is no CSRF protection and AI-provider API keys are stored in plaintext, so don't put this on a public host — --host 0.0.0.0 exists but exposes your documents to anyone who can reach the machine.
• No password recovery and no email verification, by design.
• The "Resolve with AI" button is an intentional stub (app/ai/): provider settings can be saved, but the in-app resolve endpoint returns 501. The export/agent loop above is the supported path.
• Paragraph grouping in PDFs is a heuristic; unusual layouts (multi-column, tables) can group oddly.
• Cross-page PDF selections are rejected rather than silently truncated.

Contributing

python tests/smoke_test.py runs the end-to-end check (it installs nothing, but expects MarkAI to be installed — pip install -e . first). CI runs it on Linux, macOS and Windows against a freshly built wheel.

License

MIT — see LICENSE.