# MONARC

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Salman Shaikh](https://cerebralvalley.ai/u/imselmon)
- **GitHub:** https://github.com/monarcpro/matte
- **Demo video:** https://youtu.be/HAmUJOC89D8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5352

Matte is an open-source agentic video editor that runs in the browser, built on Meta's models. Live at https://matte.monarcpro.app (no sign-in).

The problem: editing a short video still means masking, keyframing, captioning and designing graphics by hand, and most AI editing tools only see people. We make MONARC PRO, a desktop video editor, and our users want to say what they want and get a finished cut.

How it works: drop in a clip or paste a link and say what you want: "edit this", "put SUMMER behind the dog", "blur every face", "make a 15-second promo from this site".

1. Muse Spark 1.3 watches the clip (summary, findable objects, timed moments, spoken hooks, mood) and Muse Voice transcribes it with times.
2. Matte plans before it cuts: a storyboard of 3 to 6 scenes with times and the exact edits, shown as a card you approve or change.
3. Build: SAM 3.1 finds any object you can name and keeps one id on it through the clip. That drives text behind the subject, blur, pixelate, colour pop, spotlight, outline, a Muse Image background under the cutout, a 9:16 or 1:1 crop that follows the subject, punch-ins and trims. Captions come from the transcript.
4. Motion graphics have no templates: Muse writes each scene as a Volito program (our open-source motion library) in the brand's colours and fonts, read from the page. Matte runs it, checks frames and sends errors back to fix.
5. Muse then reviews four frames of the finished cut as an art director and applies up to three fixes.
6. Export renders every frame to H.264 in the browser with ffmpeg.wasm, the clip's audio muxed back.

Matte grew out of a benchmark we ran during the hackathon: SAM 3.1 vs MediaPipe on nine stock clips, scored against chroma-key truth. On people, SAM 3.1 cut edge error by 40% (0.070 vs 0.118); on anything that isn't a person, only SAM works. Meta's Roman Rädle featured it in his thread on hackathon projects.
Demo : https://matte.monarcpro.app/

Benchmark: https://matte.monarcpro.app/bench/report.html
Feature tests: https://matte.monarcpro.app/bench/feature-report.html
Write-up: https://monarcpro.app/blog/sam-3-1-vs-mediapipe

## More from Meta Global AI Developer Hackathon

- [Impact Gate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5349)
- [MergeConflict](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5350)
- [LegalBuddy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5351)
- [BitPOS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5353)
- [hireMarco.ai](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5354)
- [RoverEdge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5355)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5352. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt