Cutline

A video editor that runs in the browser, with recording built in.

Capture your screen, your camera and your microphone in one take. Each one is written to its own file on a shared clock — so afterwards you can re-time, mute, crop or replace any of them without disturbing the others. Nothing leaves your machine: recordings, imports and projects are ordinary files under ~/Cutline, and the export runs locally.

What works today

Record, edit, export — the whole loop, on your own machine, in one full-page editor.

Projects

Save, autosave, crash recovery, duplicate, delete. Media that has gone missing between sessions is detected and named rather than rendering black.

Media

Drag-and-drop or pick files — video, audio, images. Thumbnails, waveforms, duration, resolution, codec, bins, search, sort, favourites.

Timeline

Everything from one take stays linked: cut the picture and the voice is cut with it, at the same instant, so lip sync cannot quietly drift. When you want them apart, Detach audio — or unlink the whole take — and edit each freely.

Unlimited video and audio tracks. Move, trim, ripple trim, slip, split, ripple delete, duplicate. Snapping to clip edges, markers and the playhead. Per-track lock, hide, mute and solo. Drag clips between tracks. Waveforms drawn on audio clips. Markers, in/out points, JKL shuttle.

Layers

Click a layer in the preview to select it, drag it to move, pull a corner to resize, an edge to stretch, or the handle above it to rotate — with snapping to the centre, the thirds and the edges. Or set any of it numerically: position, scale, stretch, rotation, anchor point, opacity, flip, crop, rounded or circular framing, drop shadow, and all sixteen blend modes. Rectangle and ellipse masks with feather and invert. Chroma key with similarity, smoothness and spill suppression.

Colour and effects

Exposure, brightness, contrast, highlights, shadows, whites, blacks, saturation, vibrance, temperature, tint, hue, sharpen, fade. A stacked, reorderable effects rack: blur, gaussian blur, sharpen, glow, film grain, vignette, pixelate, posterize, black & white, sepia, invert, tint, chromatic aberration, scanlines.

Motion

Keyframe any slider from the diamond beside it — linear, ease, ease-in, ease-out or hold. Fourteen transitions in and out of every clip.

Text, shapes and captions

Text layers with font, weight, size, tracking, leading, colour, stroke, shadow, backing and six animation presets. Shape layers — rectangle, ellipse, line, triangle, star, arrow. Captions written by hand or imported from SRT or WebVTT, styled and burned in, exported back out.

Monitoring

Histogram, luma waveform, RGB parade and vectorscope, measured off the program monitor so they reflect every grade exactly as it will export. Grid, rule of thirds, centre guides, action safe and title safe.

Export

MP4/H.264 or WebM/VP9 with delivery presets for YouTube, Shorts, Reels and TikTok, plus resolution, frame rate, quality and explicit bitrate — and an in/out range. Encoded on your machine at around 4× realtime; nothing is uploaded.

The preview and the exporter render through the same function, so the file you get is the frame you approved.

Recording

Start from what you want to capture — screen, camera, both, or just your voice — and the right sources arm themselves. The camera rides over the screen in the preview exactly as it will in the edit, so the framing decision is made once.

• Screen, camera, microphone and system audio, any combination, each to its own file on a shared clock
• Device pickers, frame rate and noise-suppression controls
• Live level meters that warn when a microphone has gone silent — while the take can still be saved, rather than after
• Pause and resume, with the paused spans recorded
• Streamed to disk as it records, so a long screen capture neither fills memory nor disappears if the tab dies — then moved into ~/Cutline when the take ends, and deleted from the browser only once every byte has arrived
• A library to play back, download and delete past takes

What does not work yet

No audio effects or mixer, stabilisation, motion tracking, multicam, LUTs, curves or colour wheels, batch export, proxies, collaboration, cloud or plugins. ProRes, DNxHR, Dolby Vision and native plugins are not reachable from a browser at all. See CLAUDE.md.

Running it

Needs Bun.

bun install
bun run dev

Then open http://localhost:5310. That starts two things together: the web app, and a small local server that keeps your projects, recordings and imported media as ordinary files in ~/Cutline — visible, backed up by whatever backs up your home folder, and not one "clear browsing data" away from gone. The server listens on 127.0.0.1 only, and every request needs a per-run token.

bun run start builds the app and serves everything from the one process.

With Docker

Needs Docker, and works the same on macOS, Windows and Linux:

git clone https://github.com/Forkscout/cutline && cd cutline
docker compose up -d

Then open http://localhost:5311 in Chrome or Edge. Projects and recordings still live in ~/Cutline, as ordinary files. Two things need the native install instead: the cursor track, which reads your system's cursor, and GPU speed for a local whisper.cpp server. Run whisper.cpp on your machine as usual and connect it in the Captions panel — the container reaches it for you. Hosted transcription services work the same either way.

On Linux, Docker cannot reach a service bound to 127.0.0.1 on the host: start whisper-server with --host 172.17.0.1 (Docker's bridge address). Docker Desktop on macOS and Windows needs nothing extra.

As a single binary

One file: the server with the web app inside it, nothing else to install.

bun scripts/package.ts          # for this machine, into out/
bun scripts/package.ts --all    # macOS, Linux and Windows

Run it and open the address it prints. Projects, recordings and media go to ~/Cutline, the same as every other way of running it.

From a published image

Nothing is published yet. When you want an image, run the image workflow (Actions → image → Run workflow) or push a v* tag: it builds for amd64 and arm64 and pushes to ghcr.io/<owner>/cutline. It never runs on a merge — publishing is a decision, not a side effect. Then:

docker run -d -p 127.0.0.1:5311:5311 -v ~/Cutline:/data ghcr.io/<owner>/cutline:latest

Auto-captions

Right-click anything with sound — a clip on the timeline or an item in the media pool — and choose Generate captions. The captions land where that clip is heard, and captions under other clips are left alone. They show in a Captions lane under the timeline's ruler — click one to go there, double-click to edit it — and on the picture while the playhead is inside one. The Captions panel does the same with a choice of language.

Captions and the transcript are two things. The transcript is every word with its own time, kept with the media file; captions are lines made from it for viewers. Editing or deleting captions leaves the transcript alone, and it is the transcript an agent reads (the transcript tool) to time a cut, a title or an animation to a word.

Choose where transcription runs in the Captions panel. On this Mac, a whisper.cpp server is free and nothing leaves the machine — but a long video wants a fast one. Otherwise, a hosted service with your key: OpenAI, Groq, OpenRouter or ElevenLabs (Scribe is strong on Hindi and mixed-language speech). Any OpenAI-compatible endpoint works. For the local route:

whisper-server -m ~/.cache/whisper-cpp/ggml-large-v3-turbo.bin --inference-path /v1/audio/transcriptions --convert -l auto --port 8178

Then press Connect in the Captions panel (the address is filled in). Asking for captions before a service is connected opens that panel. LM Studio has no transcription endpoint yet, so it cannot serve this.

Letting an agent edit

Cutline is also an MCP server. With it running, register it with Claude Code once:

claude mcp add --transport http cutline http://127.0.0.1:5311/mcp --header "Authorization: Bearer $(cat ~/Cutline/mcp-token)"

Open a project and ask. The agent edits the project open in your editor through the same actions the interface uses, so you watch each change land, and its whole turn is one undo step. It can render any frame, check the audio for gaps and clipping, and export the finished video to ~/Cutline — and it is told to look before it reports.

It works like a director rather than a button-presser. It reads the project's brief (the Brief tab: goal, platform, layout, brand, references, rules), asks you what is missing in a form that opens in the editor, looks at the footage first — graphics already burned into it, where you sit in the frame, what is said — and shows you the video in a few design themes to choose from. Then it builds one scene for your approval, and the rest from the theme you chose: titles, lists, flows, bars, lower thirds, the speaker moving into a side panel. Changing the look afterwards is one step. In Claude Code, /mcp__cutline__direct starts that workflow; /mcp__cutline__review checks an edit.

Chrome or Edge. The recorder streams to disk through createSyncAccessHandle, which only exists in a worker, and the exporter encodes with WebCodecs. Neither has a fallback here, and pretending otherwise would just fail later and less clearly.

Built with React 19, TypeScript, Vite, Tailwind v4, shadcn/ui and mediabunny for demuxing, decoding and muxing.

Checking your work

bun run typecheck

Two harnesses run in the browser and need no permissions — a canvas stream and an oscillator are indistinguishable from a camera and a microphone as far as MediaRecorder is concerned:

• /dev-check.html covers the recorder: chunk ordering, the OPFS worker, track offsets, pause accounting, reading files back.
• /dev-editor-check.html covers the edit model, linked clips, subtitles, moving a take to the server, import and export — then decodes the file it just exported and reads its pixels, checking that the background, the screen layer, the camera overlay, the text, the captions, an effect and a colour grade all actually arrived.

That last part is not belt-and-braces. The first export this project produced was the right duration, the right codec and the right resolution, and entirely blank. Nothing short of looking at the pixels would have caught it.

Contributing

Issues and pull requests are welcome.

Two things worth knowing before you change rendering code, both explained at length in CLAUDE.md:

1. The preview and the exporter render through one function, drawFrame in src/editor/compositor.ts. Effects go there and nowhere else — adding one to the preview alone is the bug that design exists to prevent.
2. Clips recorded together are linked, and every timeline edit propagates through the group. Split the picture without the sound and lip sync drifts silently.

Run bun run typecheck and both harnesses before opening a PR.

Licence

MIT.