Vursor

A cursor-like AI video editor that lets you trim, cut, add effects, smart-zoom, subtitles, and sound - all through natural language commands. No timelines, no hassle, just talk to your video.

Comment

Inspiration

Editing videos is powerful but painful. Traditional editors require scrubbing through timelines, hunting clips, and stacking effects. We asked: what if editing was as easy as telling your computer what to do?

That’s how this project was born — inspired by Cursor, but reimagined for video editing. Instead of cut() and drag-and-drop, you just say:

“Trim the part where the guy speaking is silent.”

…and it happens.

How We Built It

  • Backend (FastAPI + TwelveLabs + FFmpeg): Handles video/audio processing, trims clips, adds effects, and generates previews.
  • NLP (Cohere): Parses natural language into structured editing commands.
  • Video Search (Twelve Labs): Finds key moments (like “LeBron silent” or “when the person dies”).
  • Frontend: A minimal Cursor-like UI with chat-driven commands, video preview, and instant feedback.
  • Dev Environment (Windsurf): Used Windsurf’s AI-first IDE to rapidly prototype, refactor, and debug the entire stack under hackathon time pressure. It cut down iteration time massively.

The workflow looks like this:

User Command → Cohere (parse intent) → Twelve Labs (find moment) 
→ Executor (FFmpeg) → Preview video

What We Learned

  • How much AI-powered development environments like Windsurf accelerate building — almost like pair-programming with a senior engineer on demand.
  • How multimodal AI (language + video) can completely reshape creative tools.
  • The trade-offs between real-time performance vs. hackathon prototyping .
  • How important it is to scope ruthlessly: better to demo 3 magical features than 10 broken ones.
  • That even for creative tools, structured pipelines (NLP → search → execution) make everything easier.

Challenges We Faced

  • Latency: Running Cohere + Our Server + Twelve Labs can get heavy. We solved it with mocks + pre-indexed demos.
  • Parsing ambiguity: Natural language is messy. We had to carefully prompt Cohere and fallback to heuristics.
  • Media handling: Combining audio overlays with video timelines isn’t trivial — ffmpeg saved us here.
  • Time pressure: 30 hours forced us to keep scope razor-thin.

What’s Next

  • More robust effect libraries (glitch, transitions, auto-cuts).
  • Real-time collaboration — multiple users editing the same video via chat.
  • Scaling up beyond demos: distributed video rendering for fast results.

Final Thought

We set out to answer one question:

What if video editing was as simple as talking to your video?

This project is our first step toward that future.

Built With

Share this project:

Updates