Convert, compress and read files. Offline, free, no upload.
iLovePDF, but it runs on your computer and your files never leave it.
Builds are published on the releases page for Windows, macOS and Linux.
There is no installer wizard to sit through and nothing to configure. Download, open, drag a file in.
Twenty-seven tasks, each a card on the home screen rather than a coordinate in a format matrix.
PDF. Compress, merge, split, rotate, delete pages, reorder pages, add a watermark, add a password, remove a password you know, export pages as images, build a PDF from images.
Images. Convert between JPG, PNG, WEBP, TIFF, BMP and GIF; compress with a quality setting; resize by pixels or by a preset; crop to a shape.
Text. Read the words out of a photo or a scan, pull the text out of a PDF, turn a scanned PDF into one you can search and copy from.
Documents. PDF to Word, Word and Excel and PowerPoint to PDF, and the usual interchange between Office and OpenDocument formats.
Video and audio. Convert and compress video, pull the audio out of a clip, convert between audio formats, make a GIF.
Drag a file onto the home screen and the grid filters itself to what you can do with that file. Drop a photo and you see Convert, Compress, Resize, Extract text, Images to PDF. You never have to know which category anything is in.
Then the task screen: the file already attached, at most three options with sensible defaults, and one button carrying a verb.
Because the file you are converting is usually a document with your name, address or ID number on it, and an online converter means uploading it to a server belonging to someone you have never heard of.
Beyond that: no size limit, no watermark, no daily quota, no account, and it works with the network unplugged.
Lathe is a desktop interface over pdfcpu, Tesseract, FFmpeg and LibreOffice. It did not reinvent any of them, and it is not going to out-implement FFmpeg.
What Lathe contributes is the part those tools do not: an interface that gets out of your way, a job pipeline that will not damage your file, packaging that installs in seconds and grows only when you ask it to, and a guarantee that nothing you convert ever leaves the machine.
The engineering that actually took the time was the dependency tiering (BUNDLING.md), the cross-platform process and filesystem work (DISCOVERIES.md), and making a webview-based window stop feeling like a web page.
Compressing a PDF re-encodes the pictures inside it, which is what shrinks a scan. A lossless optimiser pass alone would report a scanned document as "already as small as it can get", which is true of its structure and useless to the person holding a 14 MB file and a 2 MB upload limit.
Measured on the sample scan from scripts/demofiles:
| Quality | Result |
|---|---|
| Smaller file | 1.5 MB to 197 kB, 87% smaller |
| Balanced | 1.5 MB to 288 kB, 80% smaller |
| Best quality | 1.5 MB to 493 kB, 66% smaller |
A PDF that is already tight comes back untouched with a note saying so, rather than being returned slightly larger.
Those figures are what a bare install produces. If Ghostscript happens to be installed, Lathe also runs the file through it and keeps whichever result is smaller, which helps most at the best-quality setting because Ghostscript can lower a scan's resolution and the built-in path cannot. Nothing prompts you to install it and no task requires it. The reasoning is in DECISIONS.md D7, and the limit it lifts in KNOWN_GAPS.md.
| Size | When | |
|---|---|---|
| Lathe itself | 18.1 MB | The download |
| Video and photo support (FFmpeg) | 111 MB | First video task, or the first HEIC photo |
| Text recognition (Tesseract) | 50 MB on Windows, detected elsewhere | First OCR task |
| Stronger PDF compression (Ghostscript) | 65 MB on Windows, detected elsewhere | Optional, never required |
| Office documents (LibreOffice) | detected, not downloaded | First Word or Excel task |
PDF and image work needs none of that: it is compiled into the binary and available the moment Lathe opens. Anything heavier says what it needs and how big it is before you commit, then works offline afterwards.
How a component arrives depends on the platform, not on the project. On Windows, Tesseract and Ghostscript publish a setup program and nothing portable, so Lathe downloads the exact file it has a checksum for and runs it into its own folder. Windows asks your permission first, because both installers request it, and Lathe says so beside the button rather than letting the prompt arrive unexplained. Removing the component afterwards is Lathe deleting its own directory.
On macOS and Linux those two come from a package manager, which needs a root password Lathe has no business asking for, so there they are detected and Lathe tells you the exact command to run. LibreOffice is detected everywhere. See BUNDLING.md.
Measured, not estimated: docs/evidence/binary-size.md.
| Measured | Target | |
|---|---|---|
| Cold start | 479 ms | under 2 s |
| Idle memory | 31.8 MB | n/a |
Median of five cold launches on Windows 11, Ryzen 7 6800HS. Method and raw numbers in docs/evidence/startup-windows.md. Reproduce them yourself; do not take the table's word for it.
Character-level accuracy against committed ground truth, from docs/evidence/ocr-accuracy.json:
| Input | Accuracy |
|---|---|
| Clean scan | 100% |
| Skewed, unevenly lit, noisy | 100% |
| Hard-edged shadow across the page | 99.6% |
| A third of the resolution | 99.8% |
Read that table with the caveat it deserves. The corpus is rendered from digital text and then degraded programmatically, which is an easier target than a photograph of physically printed paper, so real-world accuracy on a crumpled receipt will be lower. What the benchmark is genuinely good for is catching a regression in the preprocessing chain, and CI fails if any figure drops by more than two points.
The preprocessing is why the degraded sets score close to the clean one. Lathe flattens the lighting, deskews, trims the dark border from a page photographed on a desk, denoises, upscales below roughly 300 DPI, and only then binarises, using Sauvola's method: a threshold chosen for every pixel from its own surroundings rather than one chosen for the whole page.
Those last two steps are the ones that were measured rather than assumed. On a page with a hard-edged shadow across it, the kind cast by your own hand or by the gutter of a bound book, a single threshold for the page reads 37% of the text and a local one reads 99%. Flattening the lighting first, before the deskew and the border trim rather than after them, is worth a further fifteen points, because both of those steps judge a page by how dark its rows are and a shadow makes them answer about the lighting instead. The comparison is internal/engine/ocrengine/binarize_test.go and it fails if either choice is undone.
Lathe never uploads your files. There is no server, no account and no telemetry of any kind.
You can check this in ten seconds: disconnect from the internet and everything except a one-time component download works exactly the same.
It is also enforced in the code rather than promised in a document. No package
outside internal/deps and internal/update may so much as import network
code, and the whole conversion suite runs in CI with outbound access blocked.
Both are tests, not policies: test/isolation.
Full detail in PRIVACY.md.
Stated plainly, because you will find them anyway.
- PDF to Word is best-effort. LibreOffice does the conversion. Simple pages come through well; multi-column layouts with floating figures do not. Lathe says so on screen before the conversion, not after.
- HEIC needs the video component. No pure-Go HEIC decoder is worth depending on, so the first HEIC photo prompts for the FFmpeg download.
- Exporting PDF pages as images works on scans, not on vector pages. Lathe extracts the embedded page images rather than rasterising, so a page drawn as text and shapes has nothing to extract. It says that instead of failing.
- The Linux window looks different. Windows and macOS get custom chrome; Linux keeps the window manager's own title bar, because Linux window managers vary too much to own that surface safely in a first release.
- Windows builds are not signed yet. SmartScreen will show a blue warning: click More info, then Run anyway. A certificate costs a few hundred dollars a year and will be bought when the project has enough traction to justify it. Saying nothing about this would just produce confused users.
- No PDF text editing, no mobile app, no AI features. Deliberate, with reasons: KNOWN_GAPS.md.
The same task registry, headless:
lathe pdf-compress report.pdf --quality medium
lathe pdf-merge a.pdf b.pdf c.pdf -o merged.pdf
lathe image-convert photo.png --format webp
lathe text-from-image scan.jpg --lang eng+hin -o text.txt
lathe --list-tasks
Output paths go to stdout, one per line; progress and notes go to stderr, so piping stdout gives a clean list of results.
make deps # Go modules and npm packages
make check # format, vet, import boundaries, tests
make corpus # regenerate the golden test corpus
make build # the desktop app
make build-cli # the headless binary
Needs Go 1.25+, Node 22+ and, on Linux, libgtk-3-dev and
libwebkit2gtk-4.1-dev. scripts/linux-deps.sh installs those across
Debian, Fedora and Arch.
build/docker/test.Dockerfile carries Tesseract, FFmpeg and Ghostscript, so the
OCR, media and compression suites run on a machine that has none of them:
docker build -f build/docker/test.Dockerfile -t lathe-test .
docker run --rm -v "$PWD:/src" -w /src lathe-test go test ./...
Adding a task means adding one entry to the registry and teaching one engine to handle it; nothing else changes, and both the interface and the CLI pick it up automatically. ADDING_A_TASK.md walks through it.
Two rules are not negotiable, and both are enforced by tests:
- No task ever modifies an input file. Every output is written to a temp file beside its destination and atomically renamed.
- No engine error reaches a user untranslated. "exit status 137" is not a message.
Lathe is AGPL-3.0. The reasoning, including why not MIT, is recorded in DECISIONS.md as D4. Component licences are catalogued in LICENSES.md.