WeftCut

WeftCut

The video editor your AI agent can actually drive. Point Claude, Cursor, or any MCP client at a real desktop NLE — and watch every edit land on your timeline while it plays.

weftcut.com

An agent at work over MCP while playback runs: restyling the lower third, trimming the B-roll, then undoing both — every edit lands in the UI in real time, and the Agent panel records what it did.

Most editors bolt AI on as a feature: a button that generates something, a sidebar that suggests a cut. WeftCut exposes the editor itself as a tool surface. A localhost MCP server publishes 88 editing tools — place, trim, split, restyle, keyframe, group, caption, mix, checkpoint — and whatever agent you already have open drives them. The intelligence lives outside the app: the installer bundles no model weights, and the local engines for speech and vision are an opt-in download when you want them.

None of that rides on a stripped-down editor. Everything an agent can reach, you can do by hand, in a full NLE: A/B-roll timeline, keyframes with a curve editor, effects, captions, a role-based audio mixer, and hardware-accelerated export.

Install

| Platform | Download | |---|---| | Windows | WeftCut-win-x64.exe | | macOS (Apple Silicon) | WeftCut-mac-arm64.dmg | | Linux | AppImage · .deb |

Windows updates itself in the background. The macOS build is ad-hoc signed and not notarized, so it neither self-updates nor opens on the first try — allow it under System Settings → Privacy & Security → Open Anyway. All releases: github.com/WeftCut/WeftCut/releases.

Hand your timeline to an agent

Copy the MCP URL and token the app prints on startup (also available in-app) into your client's config, and your agent is holding the editor.

• It edits the project you are watching. No import/export round trip, no

separate headless copy. The agent calls trim_layer, the block shortens under the playhead while audio keeps playing, and your next click continues from there.

• You can always see what it did. Every tool call becomes a row in the

Agent panel — Trimmed clip · Ember.mp4, Added marker · Needs a look — and the status bar echoes it. Nothing happens off the record.

• Every batch is reversible. Agents checkpoint at logical boundaries

(checkpoint), and each checkpoint is one click from being restored. The agent can also rehearse a whole multi-step edit against a throwaway clone first (dry_run) and find the collision before it touches your project.

• Multi-agent, with rules. Sessions are per-connection. An agent that is

mid-batch can lock_history so a stray Ctrl-Z doesn't land in the middle of its work — and the lock, plus the reason it gave, is shown to you.

For a longer run, an agent can call begin_agent_session and fold the UI down to preview, scrub, and a record of what it is doing:

Code as video: Motifs

A Motif is an on-screen element written as code — a real web page (HTML, CSS, SVG, canvas, WebGL, whatever you reach for) dropped on the timeline as a layer. It isn't a fixed preset: the author declares the knobs, and the inspector builds the form from that declaration, so seconds, label and accent below are editable in the app because the Countdown's manifest says they are.

The one rule that makes it an editor feature instead of a screen recording: a Motif renders as a pure function of time. It never advances itself. The harness owns the clock and drives the page to each composition frame, so the same t always yields the same pixels — when you scrub, when you re-export, and in preview and export alike.

// The built-in Countdown's whole script, near enough verbatim — the rest of the
// file is the SVG ring it animates. `props` is whatever the manifest declares;
// `frame` runs once per composition frame and keeps no state of its own.
motif.define({
  setup: async function (props, ctx) {
    _label = props.label != null ? String(props.label) : "GO";
    num.style.color = props.accent;
    ring.style.stroke = props.accent;
    ring.setAttribute("stroke-dasharray", C);
    ring.animate([{ strokeDashoffset: 0 }, { strokeDashoffset: C }],
      { duration: ctx.duration * 1000, easing: "linear", fill: "both" });
  },
  frame: function (t, ctx) {
    var n = Math.max(0, Math.ceil(ctx.duration - t));
    num.textContent = n > 0 ? String(n) : _label;
  },
});

Which is where agents and video stop being two separate ideas. Writing a web page is the thing coding agents are already best at — so an agent can author a brand-new overlay for your edit instead of picking one from a catalog:

write_motif_draft writes it · preview_motif_draft renders a frame back as a PNG so the agent can look at its own output and fix it · add_motif puts the draft on the timeline · install_motif publishes it to your catalog for good. The authoring spec ships inside the app as an agent skill, so the model gets the contract without you pasting documentation.

Lower thirds, countdowns, karaoke text and animated title cards are the obvious uses. Anything you can build on a page is the actual limit.

A complete editor underneath

Timeline — A/B-roll rows with filmstrips and waveforms, frame-accurate SMPTE editing, ripple delete that closes the gap behind it, linked A/V that trims as one clip, cross-track groups, and nested compositions. Agent: move_layer · trim_layer · split_layer · ripple_delete_gap · links_create · groups_create · move_layers_to_composition

Keyframes — animate any parameter, with bézier easing, a curve editor, tangent control, motion paths and extrapolation. Agent: set_keyframe · set_keyframe_easing · set_keyframe_tangents · smooth_keyframes · set_extrapolation

Speech and captions — transcribe a clip and get editable caption layers packed onto your caption tracks; import SRT/VTT/ASS the same way. Transcription runs against a cloud provider, or entirely on your machine once you let the app fetch a local engine (whisper.cpp, FunASR). Text-to-speech for scratch voiceover. Agent: transcribe_clip · apply_subtitles · synthesize_speech

Audio — role-based mixing (dialogue / music / SFX / voiceover) with live per-role metering, gain, pan, fades and denoise. Pauses finds the dead air in a take and cuts it as one undoable edit, keeping a pad so speech still breathes. Agent: detect_pauses · remove_pauses · set_role_gain · set_role_flags

Understanding the footage — shot-boundary detection with per-shot brightness, motion and sharpness, frame comparison, and vision-model descriptions of what a clip actually contains. An agent can cut on content, not just on timecode. Agent: analyze_clip · auto_split_by_shot · describe_clip · compare_frames

Titles, effects and transitions — styled text layers with outlines, per-layer effect chains including chroma key with an eyedropper that picks from the live frame, and transitions between clips. Agent: add_effect · update_effect · add_transition · update_transition

Export — H.264 / HEVC / AV1 up to 10-bit, hardware or software encoders, resolution / fps / quality controls, and streamed muxing that never buffers the whole render in memory.

Find anything — one Ctrl+K palette over commands, media, clips, captions and markers, with pinyin support.

Under the hood

Electron shell with a React 19 UI; the preview composites through PixiJS v8 + WebCodecs on a `, and export runs the same compositor in a Worker on an OffscreenCanvas. A **Rust** core (napi-rs) does decoding, audio mixing, media analysis and background jobs; **ffmpeg** handles the codecs it doesn't; mediabunny` demuxes and muxes. The agent surface is MCP over streamable HTTP.

The design choices behind that, and why several of them were reversed, are written down as [ADRs](docs/adr/).

Build from source

Prerequisites: Node 24+, Rust (stable via rustup), and your platform's C++ build tools — per-OS commands in [docs/setup.md](docs/setup.md).

npm install       # JS dependencies
npm run bootstrap # one-time: fetch ffmpeg + build the Rust addons
npm run dev       # start the editor

Also useful: npm run typecheck, npm test, npm run e2e, npm run build, npm run package (installers).

Documentation

Start with [architecture](docs/architecture.md) for the system map, or [MCP server & agent UX](docs/mcp.md) for the tool surface, resources and multi-agent behavior. [Motifs](docs/motifs.md) and [motif authoring](docs/motif-authoring.md) cover the overlay engine and its contract.

Everything else

• [Data model](docs/data-model.md) — project state schema, history, persistence, validation.
• [Render](docs/render.md) — PixiJS + WebCodecs renderer architecture.
• [Preview](docs/preview.md) — interactive preview surface and transport.
• [Export](docs/export.md) — export settings and range, audio export, final mux, proxies, background jobs.
• [Captions](docs/captions.md) · [Audio](docs/audio.md) — caption ingestion and the audio engine.
• [Conformance](docs/conformance.md) — media fixtures and E2E gates for frame alignment, audio sync, colour.
• [Features](docs/features.md) — small-feature contracts: undo-stack scope, groups, search palette, colour picker.
• [Status / log system](docs/status-log.md) — the bottom-of-editor log bus.
• [Setup](docs/setup.md) — per-OS toolchain prerequisites and first-run flow.
• [Licensing](docs/licensing.md) — MIT app plus the two FFmpeg lanes and their build-time compliance gates.
• v1 target — release scope and open work, tracked as an issue: docs/ describes what exists today.

Maintainer

WeftCut is built and maintained by UncleChair.

It exists to speed up my own video work. I wanted an editor an agent could actually drive, and a timeline I could keep watching while it did — so the MCP surface is the part I use daily, not a demo bolted onto the side. Motifs came from the other half of that: the overlays I wanted were easier to write than to find, and an agent that can write a web page can write one for the shot in front of it.

The project sits under the WeftCut organization so the name, domain and releases have a stable home, but it is a one-person project — issues and pull requests all reach me.

License

WeftCut is licensed under the [MIT License](LICENSE). Packaged installers bundle FFmpeg binaries under their own licenses (LGPL shared libraries for in-process decode, GPL command-line tools run as a separate process) — see [THIRD-PARTY-NOTICES.md](THIRD-PARTY-NOTICES.md) and [docs/licensing.md](docs/licensing.md) for the full model.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: WeftCut
• Source: WeftCut/WeftCut
• License: MIT
• Homepage: https://weftcut.com

Install and usage instructions live in the source repository linked above.