5 apps you pay for, rebuilt free

1BentoPowerPoint in one HTML file5.6K ★ GitHub stars

github.com/nyblnet/bento · MIT · try it: bento.page/slides

The whole app (editor, slideshow, morph transitions, charts, speaker view, live collaboration) lives inside a single ~720 KB .html file. Your deck sits inside that file too, so you send the file and the other person has the app.

• Get it: open bento.page/slides, or download Bento_Slides.bento.html from the GitHub releases page and open it in any browser. Saving rewrites the file in place.
• Morph: give two elements on neighbouring slides the same id and they animate between them, like PowerPoint's Morph.
• My test: its own showcase deck, presented in the browser, morphs and all.
• It's slides only. There's no Word or Excel.

2Indra OCRAcrobat-style "scan to text" on your CPU1.7K ★ GitHub stars

github.com/datalab-to/indra · code Apache-2.0, model weights under a modified OpenRAIL-M licence

pip install indra-ocr
indra-ocr scan.png            # markdown out
indra-ocr paper.pdf --out results/

• It's a 1-million-parameter model with no GPU needed. The first run downloads about 76 MB of models.
• My test: the 1969 "In Event of Moon Disaster" memo (public domain, US National Archives via Wikimedia Commons). It took 0.59 s of model time, about 0.8 s for the whole command, on a laptop CPU. It got every line, and its only slips were where the scan itself is smudged ("ordained", "their"). A copy tilted 28° and an upside-down copy read the same.
• It reads text and layout. It doesn't edit or sign PDFs the way Acrobat does.

3Woxithe Wolfram Language (Mathematica's language) in Rust960 ★ GitHub stars

github.com/ad-si/Woxi · AGPL-3.0 · try it: woxi.ad-si.com (runs in your browser)

cargo install woxi
woxi eval 'Integrate[x^2 Sin[x], x]'

• My test, in the browser playground: Plot3D[Sin[x] Cos[y], {x, -3, 3}, {y, -3, 3}], Integrate, Solve, DSolve and FactorInteger[2^64 + 1] all came back right.
• About 3,500 of the language's functions work so far, per its own functions.csv. That's not all of Mathematica, so check that the ones you need are there. It also has Jupyter support and a notebook app (Woxi Studio).

4bg0background remover in your browser492 ★ GitHub stars

github.com/opencoredev/bg0 · Apache-2.0 (the model weights keep their own licences) · try it: bg0.dev

• Drop in a photo and download a transparent PNG at full resolution. There's no account and no limit, and the photo isn't uploaded: it runs on WebGPU, or WebAssembly where WebGPU isn't available.
• My test: a CC0 photo of a Pomeranian on a lead, 2816×2112. The fur edges stayed and the lead was removed. I logged the page's network traffic while it ran: no uploads at all (zero POST requests). The only requests were downloads (the model and its runtime) and the site's analytics loading a survey.
• In a browser without WebGPU it took about 35 s for the full model. With WebGPU it's faster.
• remove.bg is moving into Canva (its own site says the standalone site closes on 1 Dec 2026), where background removal is a Pro feature.

5AutoClipOpusClip on your own computer234 ★ GitHub stars

github.com/artbyjazi/autoclip · MIT

git clone https://github.com/artbyjazi/autoclip.git
cd autoclip && uv venv --python 3.11 && uv pip install -e ".[dev]"
cd frontend && npm install && npm run build && cd ..
autoclip doctor
ollama pull llama3.1:8b          # fully local; or add an API key for better picks
autoclip serve                   # http://localhost:8000

• Give it a long video. It transcribes it (Whisper), an LLM picks the moments, then it reframes to 9:16 following the speaker's face, burns in animated captions and exports MP4s. There's no watermark and no account.
• Needs Python 3.11 or 3.12 (not 3.13) and an ffmpeg with libass (brew install ffmpeg-full on a Mac).
• My test: JFK's 18-minute 1962 moon speech at Rice University (public domain, NASA), fully offline (Whisper "small" plus a local llama3.1 8B). It took about 20 minutes on a laptop CPU and gave 4 captioned 9:16 clips of 20-29 s: "Well, space is there", the George Mallory story, the heat of re-entry, and the space budget. It skipped "We choose to go to the Moon", so a bigger model would likely pick better.
• Two snags I hit, so you don't have to:
  • If transcription fails with open() got an unexpected keyword argument 'metadata_errors', run pip install "av<16".
  • With Ollama, also pick the model in Settings, or clip-finding fails with "No Ollama model is selected".
• It's pre-1.0, so expect rough edges. The README says clip quality tracks the model: a local 8B model works, and a big hosted model picks better.

Safety: these run on your machine or in your browser. Read the README before installing anything.

These apps belong to their authors (links above). I'm not affiliated with any of them or with the paid apps named.