open-ribbi/velocut

What it solves

Velocut is a local-first, browser-based video editor designed to integrate AI agents as first-class users. It eliminates the need to upload footage to a server, ensuring privacy and performance, while allowing users to edit video using natural language commands instead of manually manipulating a timeline.

How it works

The system uses a "protocol-first" architecture where both the human user (via the UI) and the AI agent (via an LLM) issue the same JSON command protocol to a central document model.

• Engines: It features a canonical Rust engine (compiled to WASM) and a TypeScript reference engine for flexibility.
• Rendering: It leverages WebGPU for compositing and WebCodecs for decoding and exporting video.
• AI Integration: An LLM agent (supporting Anthropic-compatible providers) can "see" the video through frame grabs and shot-boundary detection, "hear" via loudness and silence analysis, and perform edits by sending JSON commands directly to the engine.
• Storage: Media is stored in the Origin Private File System (OPFS) and project data in IndexedDB, enabling a local-first experience with multi-tab synchronization via CRDTs.

Who it’s for

Video editors and content creators who want an AI-assisted workflow that keeps their data local and provides a transparent, correctable way to automate tedious editing tasks.

Highlights

• AI-Native Editing: Natural language commands (e.g., "cut out silent parts") are converted into precise engine commands.
• Local-First Privacy: No installation or uploads required; footage stays on the user's machine.
• Branching History: Supports a non-linear edit history where human and AI actions are color-coded and can be forked.
• Advanced Browser Tech: Uses WebGPU and WebCodecs for high-performance rendering and mp4 export without memory bloat.
• Agent Perception: The AI agent can analyze loudness, silence, and shot boundaries to make informed editing decisions.