FableCut: An AI Video Editor MCP That Skips the Timeline

Most video editors assume a person is holding the mouse. FableCut is an AI video editor MCP built on the opposite assumption: that an agent might be the one making the cuts, and a human might just be watching it happen. Instead of learning a timeline UI, an agent (or a person) edits a single JSON file that describes every clip, track, effect, and keyframe in the project. Change the file, and the open browser tab updates in under 200 milliseconds.

That is the whole pitch, and it is a genuinely different approach from the usual "AI video generator" category. FableCut does not generate footage. It edits footage you already have, the way a real non-linear editor does, except the interface it exposes to automation is a document rather than a set of buttons.

What this AI video editor MCP actually does

FableCut is a free, open-source, zero-dependency non-linear video editor that runs entirely in a browser. Clone the repo, run node server.js, and it opens at localhost:7777 with a Premiere-style layout: three video tracks, four audio tracks, drag-trim-split editing, keyframe animation, transitions, color grading, and a text engine with kinetic caption styles. All of that is standard editor territory.

The distinctive part is project.json. Every element in the timeline, from clip positions to crop values to caption animation curves, lives in that one file. The editor watches it and live-reloads over server-sent events. That means three separate things can all point at the same timeline: a human dragging clips in the browser, a script writing to the file directly, or an AI agent talking to it over MCP or REST. A revision counter keeps them from clobbering each other; if an agent's write conflicts with an unsaved change in the UI, it gets rejected instead of silently overwritten.

For an agent, the actual editing tools are fablecut_status, fablecut_get_project, fablecut_patch_project, fablecut_import_media, and fablecut_analyze_reference, plus a docs tool that serves the project's CLAUDE.md manual on demand. The patch tool is worth noting: an agent sends small edit operations rather than rewriting the whole document each time, which keeps token use down on anything past a trivial cut list.

One feature stands out beyond basic cutting: point it at a reference video you like, and fablecut_analyze_reference (or the CLI equivalent) returns an edit blueprint, shot boundaries, music beats and BPM, a loudness curve, per-shot energy, and the reference's music track pulled out and dropped into your media bin. An agent can then rebuild that structure with different footage. This is the closest FableCut gets to "smart" automation; everything else is precise, literal timeline manipulation.

Setup and requirements

Node 18+ and a Chromium-based browser cover the base install. ffmpeg on the system PATH is optional but enables the fast, frame-accurate export path; without it, FableCut falls back to a realtime MediaRecorder capture. The server now binds to 127.0.0.1 only by default (as of v1.3.1), so using it from another device on a LAN requires explicitly setting HOST and an allowed-hosts variable.

Registering it as an MCP server is one command:

claude mcp add -s user fablecut -- node "<path-to>/fablecut/mcp-server.js"

There is also a Claude Code plugin install (/plugin marketplace add ronak-create/FableCut) that bundles two skills, edit-video and remake-reel, and keeps project data outside the checkout so an update never touches your footage.

Why "describe the edit" beats clicking a timeline

The practical case for this approach shows up once you try to hand off real editing work to an agent. A conventional editor has no machine-readable state; an agent can only drive it by simulating mouse clicks against a UI, which is brittle and slow. FableCut sidesteps that by making the project state the API. An agent reads a compact, one-line-per-clip summary of the timeline, writes a patch, and the change appears on screen immediately, so a person supervising the edit can catch a bad cut before it goes further. That live-reload loop is what makes this a describe-the-edit tool rather than a black-box generator: you can watch the agent work in something close to real time.

It also means the skill transfers. The same JSON schema an agent edits is the same schema a script can edit, or a person can edit by hand in a text editor, without switching tools.

Where this fits and where it doesn't

FableCut is worth adding to an AI stack when the workflow is raw footage in, and an agent assembling a rough cut, grading it, captioning it, and cutting it to a beat, especially for short-form vertical content, where the built-in 9:16 preset and beat-marker tooling are aimed. It is not a generator: it will not create shots you do not already have, and it leans on ffmpeg for a fast export path rather than shipping its own encoder.

It is also worth being clear about scope: this is a single-project, mostly local tool built by an independent developer, not a hosted product with support guarantees. The open issues list sits at 13 as of this writing, which is normal for an actively developed repo but worth checking before depending on it for anything time-sensitive.

The traction, verified

As of today, the FableCut repo shows 589 stars, 60 forks, and 13 open issues on GitHub, released under the MIT license and actively pushed to within the last day. It has also picked up a Hacker News front-page mention and a listing in the official MCP registry as io.github.ronak-create/fablecut, on top of the growth the scan flagged. For a project built by one person as a zero-dependency tool, that is a fast climb, and the star count keeps moving.

Bottom line

FableCut is a narrow, well-built answer to a specific problem: giving an AI agent a real, inspectable, live-updating video timeline instead of a UI it has to fake its way through. If the editing task fits inside "cut, grade, caption, export" and the footage already exists, it is worth trying before reaching for a heavier hosted editor. If the need is generating footage from scratch, this is the wrong category of tool entirely.