Offline OCR

Handwriting in.
LaTeX out.

makaar turns photos of your notes, handwritten or printed, into Markdown, LaTeX and Word with real equations. It runs on your laptop. Nothing is uploaded.

Download Other platforms

Free and open source · Windows, macOS, Linux · v0.1.0

Three handwritten lines: a sentence and two equations about the total energy of molecules

makaar writes

(b) Total energy of molecule is const. \[ E = n_{1}E_{1} + n_{2}E_{2} + n_{3}E_{3} + \cdots + n_{i} E_{i} = \text{constant}. \]

Unedited output

Your notebook, typed.

Take a photo or scan a PDF. makaar finds every block on the page, reads it, and writes the equations as real LaTeX, so they compile, and as native Word equations you can edit.

Anything it is unsure of is marked for a quick check, so nothing changes silently.

Three steps

Works in three steps.

01

Choose your pages

Images, PDFs or a whole folder. Tick Handwritten notes for handwriting: more accurate, no slower.

02

Convert

Each page appears as it is read. You get notes.md, .tex and .docx.

03

Review

See the scan beside what was read. Fix the red boxes, save, and every file updates.

A handwritten page: blocks are found, uncertain ones marked red, then shown beside the Word document with real equations

Measured, not claimed

Beats Gemini on print. On your laptop.

We test against Gemini on public benchmarks and publish every number, including where we lose.

37%

fewer errors on printed pages than Gemini 3.5 Flash-Lite: error rate 0.130 vs 0.207 on 126 OmniDocBench pages.

Level

with Gemini 3.5 Flash on handwritten maths: 0.203 vs 0.225 on 30 CROHME expressions.

Behind

on handwritten prose: 0.079 vs Gemini's 0.029 on 30 IAM lines. We're working on it.

Error rate is normalised edit distance, the share of characters you would have to fix; lower is better. Speed: about 7 seconds a page with an NVIDIA GPU from the command line, 1–2 minutes on a laptop CPU. The desktop app uses the CPU for now. Full method and tables: RESULTS.md.

The review screen: the scan with red and blue boxes beside the recognised text and equations

Private by design

Works with Wi‑Fi off.

Exam papers, lab notebooks and medical notes never leave your computer. After a one-time model download, makaar needs no internet, no account and no API key.

Download

Free. Open source.

Windows

Installer (.exe) Windows 10/11

Windows will say the publisher could not be verified; click Run. Or, on a blue screen, More info → Run anyway. We're working on signing.

macOS

Disk image (.dmg) Apple silicon

Drag makaar to Applications. The first time, right-click it and choose Open.

Linux

AppImage any distribution .deb Debian, Ubuntu

Run chmod +x on the AppImage and start it, or sudo apt install ./makaar_*.deb.

The first start downloads the reading models: about 1.9 GB, once, checked with SHA-256. The command-line version and all files are on GitHub Releases.