FableCut: A Zero-Dependency Browser Video Editor for AI Agents

FableCut: A Zero-Dependency Browser Video Editor for AI Agents

FableCut enables AI-driven video editing by treating the project file as the primary interface

FableCut is a Premiere-style non-linear video editor that runs entirely in the browser with zero npm dependencies. Unlike traditional AI video tools that hide editing logic behind an API, FableCut exposes its entire timeline—including media, clips, tracks, effects, keyframes, and transitions—as a single project.json document. This architecture allows any process capable of writing JSON to edit video, while a browser UI hot-reloads changes within approximately 150ms via server-sent events (SSE).

AI Agent Integration and Control Surfaces

AI agents can operate FableCut end-to-end by following the schema and recipes provided in the project's CLAUDE.md file. The system provides three primary ways for agents to interact with the editor:

Model Context Protocol (MCP)

For users of Claude Code or Claude Desktop, FableCut includes a zero-dependency MCP server. This interface is designed for token efficiency, allowing agents to patch the timeline with small operations (fablecut_patch_project) rather than rewriting the entire document. Agents can also fetch compact summaries of the project or specific documentation sections on demand.

Direct File Manipulation

Agents can read project.json, modify the content, and increment the revision counter. Because the UI live-reloads, these changes appear instantly in the editor.

REST API

FableCut provides a REST API for programmatic control, featuring endpoints for project management (GET/PUT /api/project), media uploads (POST /api/upload), and a library of assets (GET /api/library).

Core Editing and Motion Features

FableCut provides a professional feature set for manual and agentic editing:

  • Timeline Management: Supports 4 video tracks and 3 audio tracks with drag, trim, split, and snap functionality. It includes a rubber-band marquee for multi-selection and beat/cue markers for rhythmic editing.
  • Visual Effects: Includes 12 filter presets, adjustment layers for global grading, and full controls for brightness, contrast, saturation, hue, temperature, tint, blur, and film grain.
  • Advanced Compositing: Features chroma key (green screen) with spill suppression and AI-powered background removal via MediaPipe.
  • Motion and Transitions: Supports keyframe animation on 25 properties with easing, speed ramps that time-remap both video and audio, and 17 different transitions including glitch and whip-pan effects.
  • Typography: Offers kinetic captions (typewriter, karaoke, etc.) with neon glow effects and automatic loading of any Google Font by name.
  • SVG Overlays: Supports frame-accurate rendering of CSS-animated SVGs, allowing agents to author vector overlays like lower-thirds or confetti as plain .svg files.

Reference Video Analysis and Remaking

FableCut includes a specialized tool (analyze.js) that can extract an "edit blueprint" from a reference video. This blueprint includes shot boundaries, music beats, BPM, a loudness curve, and per-shot energy levels. An AI agent can use the fablecut_analyze_reference MCP tool to analyze a reference reel and then rebuild the same structure using the user's own footage.

Technical Architecture and Concurrent Editing

FableCut is built as a single-page application with a Node.js backend (server.js). It uses a Chromium-based browser for rendering and optionally uses ffmpeg for fast, frame-accurate MP4 exports.

To prevent data loss during concurrent editing, FableCut implements a revision counter. If a human editor makes a change in the UI while an AI agent is modifying the project, the system detects the conflict. The REST API returns a 409 error, and the UI notifies the user via a toast notification rather than silently overwriting changes.

Community Insights

Users on Hacker News highlighted the efficiency of using JSON and SSE for agent interaction over traditional APIs. Some contributors suggested further improvements for future versions, such as:

"I was hoping to see some very clever use of vision and maybe waveform analysis to give the model 'eyes and ears' to edit audio and video."

Other users noted that the JSON-based approach opens opportunities for modular templating and agentic demo video narration.

Sources