🧲 Magnetic

A free, open-source, Final Cut Pro-style non-linear video editor for Windows.

Magnetic timeline · WebCodecs/WebGL2 playback · edit-by-transcript · auto silence removal · keyframe animation · smart-render export. Built for long-form creator workflows — multi-hour recordings play with a flat memory footprint and export at disk speed.

Portfolio project. The design is an homage to Apple's Final Cut Pro — the magnetic timeline, three-panel layout, JKL transport, and edit grammar are modeled on it as the design reference. Not affiliated with Apple; no Apple trademarks are used and "Magnetic" is an original product name.

New to Magnetic? Start with the User Guide — a five-minute first edit, the magnetic-timeline workflow, edit-by-transcript, and troubleshooting.

✨ Features

Organize & import

• Library model — Libraries → Events → Projects; import copies media into the .mglib bundle (hardlinks instead of copies when source and library share a volume)
• Filmstrip browser — hover-skim filmstrips, favorite/reject ratings, text search, background thumbnail/waveform generation
• Missing-media relink — assets whose file vanished get an alert badge; relink accepts a replacement whose duration matches ±1 frame

Edit

• Magnetic timeline — spine + connected clips, lanes, gap clips, snapping, zoom, skimming; clips never overlap, edits ripple; minimap strip with drag-to-navigate when the sequence is wider than one screen, and the view pages along with the playhead during playback
• Edit grammar — append E, insert W, connect Q, overwrite D, ripple delete, lift; select A / blade B / trim T tools; trim ripples, edit points roll, clip bodies slip; drag-rearrange with magnetic collision; full undo/redo
• Viewer — JKL transport (tap again for faster), frame-accurate stepping (←/→, ±10 with Shift), in/out marks with play-the-marked-range (/), source/sequence timecode you can click and type to seek, loop playback (Ctrl+L), fullscreen review (Shift+F), live audio meter, transport buttons in both source and sequence modes, and Space plays the browser selection when nothing else is up
• Review grid — watch up to 9 selected clips side by side (browser "Watch N" button); click a cell for its audio, double-click to open it in the viewer
• Resizable layout — drag the panel splitters (double-click one to reset it); Reset Layout button restores the defaults
• Clipboard — copy/paste/duplicate clips (Ctrl+C/V/D), paste-connect (Ctrl+Shift+V), paste attributes incl. keyframes (Ctrl+Alt+V); one undo step per paste
• Keyframe animation — animate transform, opacity, and color params from the Inspector (linear/smooth easing); diamonds on timeline clips; WYSIWYG in export
• Markers — color-coded notes (M at the playhead) that ride the media through blades, trims, and moves; ruler diamonds seek on click, the Inspector edits note/color, Ctrl+;/Ctrl+' navigate; agents drop and read them too
• Transitions — cross dissolve, wipe left/right, fade-to-black
• Titles — free text + 3 presets, rendered as live canvas layers
• Color board — exposure / contrast / saturation / temperature per clip

Playback & long-form

• Sequence playback — WebCodecs decode → WebGL2 compositor; per-clip transforms (position/scale/rotation/opacity); AudioContext-clocked A/V sync
• Multi-GB clips — streaming demux fetches samples on demand with bounded memory; a 10.85 GB, 4.2-hour clip plays video + audio with a flat ~87 MB renderer heap
• Windowed audio — long clips stream 10-second audio windows double-buffered ahead of the clock instead of decoding whole files

Audio & transcript

• Audio — mixing, per-clip volume/pan, fade in/out
• Roles — tag clips Dialogue / Music / SFX (looped beds auto-tag as music); role stripes on the timeline, Dia/Mus/SFX mute buttons in the toolbar silence a whole role across playback and export
• Loudness normalization — one click measures a clip's source loudness (ffmpeg EBU R128, cached) and sets its volume to the −14 LUFS streaming standard; agents can propose it for every dialogue clip at once
• Auto-ducking — "Duck under dialogue" dips a music bed −12 dB wherever speech is detected (ramped, merged across short pauses so it never pumps); dips draw as darker bands on the bed and apply identically in live playback, offline render, and smart-render export
• Detach audio & split edits — pull audio onto its own lane and trim its edges independently for J-cuts and L-cuts
• Loop to fill — loop a connected music bed to the end of the spine in one action; seam ticks on the timeline, whole-bed fade envelope, one undo step
• Edit-by-transcript — local whisper.cpp transcription, word-click seek, select-words-and-cut (button or Delete), filler-word removal, transcript search; agents get the same power as a cut_words MCP tool that matches a spoken quote and proposes the cut
• Auto silence removal — background audio analysis, tunable dead-air detection with timeline preview, one-click jump-cuts, one-step undo
• Rough Cut — one button merges silence + filler detection into a single pass (one aggressiveness slider) and ghost-renders it before anything commits: red hatched strikethrough over the cuts, a green preview strip showing the result, Accept lands it as one undo step, Discard leaves zero history; then review each AI-badged cut on the ruler and reject individual cuts spell-checker style. Imported footage gets hands-off draft cards ("2 cuts · 4.1s tighter") — one click appends and proposes
• Copilot — chat with Claude that answers and edits: it sees the clips, detected dead air, and transcript, and performs edits (blade, ripple delete, trim, roll, slip, move, transitions) against a working copy only — every batch previews as a ghost diff with a per-change checklist (dependent changes link into one decision), a scrubbable before/after with real rendered frames, applies as one undo step on Accept (partial accepts replay transactionally), and vanishes without a trace on Discard; accepted AI edits leave session-only attribution (clip dot + "Edited by Copilot" in the Inspector). There is no export tool, and every proposed sequence passes the timeline's invariant guard. Two providers, picked in the Copilot tab: your Claude subscription via a locally installed, signed-in Claude Code CLI (auto-selected when detected; each turn spawns the CLI as a fresh process — total turn time depends mostly on the model, not the spawn), or your own Anthropic API key (app settings, never logged) — with Rough Cut, the human+agent co-editing roadmap in action (docs/agent-copilot-vision.md)
• Agent Access (MCP) — Magnetic is an MCP server: Claude Code or any MCP client connects over loopback (claude mcp add magnetic -- node scripts/magnetic-mcp.mjs), reads the open timeline (clips, roles, markers, transcript), and proposes edits that arrive as the same ghost-diff preview with a timeline banner — including the high-level cut_words (quote → cut), duck_music, and normalize_loudness tools. Nothing applies until you Accept, there is no export tool over MCP, the sidebar toggle severs all connections instantly (token-protected, 127.0.0.1 only), and requests arriving mid-drag queue until your gesture finishes
• Voice cleanup — one-click background denoise (ffmpeg FFT) per asset or as part of Rough Cut; playback and export automatically prefer the cleaned track, originals untouched
• Flow score — every accepted AI pass grades itself (residual dead air, untransitioned jump cuts, slivers): a colored score chip plus clickable orange flags on the ruler; the Copilot runs the same check on its working copy before proposing
• Captions — burned-in captions derived live from the transcript (pop-in / karaoke / block presets); SRT/VTT sidecar export

Finish & export

• Export — H.264/AAC MP4 at 1080p/720p/source via bundled ffmpeg, with progress and cancel; WYSIWYG (exported pixels match the live compositor)
• Smart-render export — when the video track is a single, possibly trimmed, visually untouched asset, export stream-copies the H.264 bitstream and renders only the mixed audio: a 6-hour VOD exports at roughly disk speed, bit-identical video
• Keyboard-shortcut overlay — Shift+? lists every live binding
• NSIS installer — npm run package, binaries bundled

Explicitly out of scope: multicam, color wheels/curves/scopes, proxy workflow (preview proxies for non-decodable codecs only), compound clips, 360°/HDR, plugins, Motion templates, ProRes encode, object tracking, collaboration.

🖥️ Requirements

• Windows 10/11 x64
• Node.js 24+, npm 10+

🚀 Setup

# 1. Install dependencies
npm install

# 2. Fetch native binaries (ffmpeg, ffprobe, whisper.cpp, base.en model)
#    into resources/bin/ — pinned URLs, sha256-verified, idempotent
npm run fetch-binaries

# 3. Generate test media into fixtures/ (uses the fetched ffmpeg + Windows SAPI TTS)
npm run fixtures

Development

npm run dev          # start electron-vite dev server + app (HMR)

Build & test

npm run typecheck    # tsc across main, preload, renderer, shared, e2e
npm run lint         # eslint
npm test             # vitest unit tests
npm run build        # electron-vite production build -> out/
npm run test:e2e     # Playwright Electron E2E (requires `npm run build` first)
npm run package      # NSIS installer + win-unpacked -> dist/ (bundles resources/bin)

The installer bundles ffmpeg/ffprobe/whisper and the base.en model (~250 MB of binaries), so the packaged app has no first-run download. e2e/packaged.spec.ts boots dist/win-unpacked/Magnetic.exe and verifies the bundled binaries resolve; it is skipped until npm run package has produced that output.

Scripts

⌨️ Keyboard shortcuts

Generated from the live shortcut registry (scripts/dump-shortcuts.mjs); Shift+? in the app shows the same list. Shortcuts are suppressed while typing in text fields. J/K/L and Space are contextual — they drive the source viewer when it has a clip open, otherwise the timeline.

🏗️ Architecture

flowchart TB
    subgraph rendererBox["Renderer (src/renderer)"]
        ui["React UI<br/>browser · viewer · timeline · inspector"]
        store["zustand timeline store<br/>undo / redo"]
        engine["Playback engine<br/>streaming WebCodecs demux + decode → WebGL2 compositor<br/>AudioContext-clocked sync"]
    end
    subgraph sharedBox["Shared (src/shared)"]
        kernel["Magnetic timeline kernel<br/>pure TS · immutable Sequence<br/>every op returns its inverse"]
    end
    subgraph mainBox["Main process (src/main)"]
        ipc["IPC<br/>zod-validated handlers"]
        lib["LibraryStore<br/>.mglib JSON + media"]
        jobs["JobQueue<br/>filmstrip · waveform · proxy<br/>transcribe · envelope"]
        exp["Export sink<br/>rawvideo pipe → ffmpeg<br/>or smart-render stream-copy"]
    end
    ui --> store
    store --> kernel
    store --> engine
    ui -->|window.api preload bridge| ipc
    ipc --> lib
    ipc --> jobs
    ipc --> exp

• src/shared/timeline/ — the magnetic-timeline kernel: pure functions over an immutable Sequence; clip positions are derived by summation so overlaps are unrepresentable; every op returns its inverse for undo
• src/main/ — Electron main process: window, menu, library persistence, background jobs, export (including the smart-render planner); every IPC handler zod-validates its payload
• src/preload/ — contextBridge bridge exposing the typed window.api
• src/renderer/ — React UI (dark FCP-style 3-panel shell), zustand state, and the playback engine (streaming WebCodecs demux + decode, WebGL2 compositing, AudioContext-clocked sync)
• scripts/ — binary fetcher, fixture generator, shortcut-table dump
• e2e/ — Playwright Electron tests; resources/bin/ and fixtures/ are gitignored artifacts

The visual language is documented in DESIGN.md (tokens, type scale, spacing, motion — Apple dark-mode chrome where the footage is the hero) and PRODUCT.md (register, users, brand personality).

Renderer security: contextIsolation: true, nodeIntegration: false, sandbox: true, CSP set in index.html, no remote URLs; library files are served over a custom mfile:// protocol scoped to the library, except video/audio playback, which goes through a token-guarded loopback HTTP server (src/main/media-server.ts) because Chromium's media pipeline cannot seek large files through Electron custom-protocol responses. Ctrl+Shift+D toggles a diagnostics overlay that spawn-verifies the bundled binaries.

📸 Screenshots

Known limitations

• The "Explicitly out of scope" list above (multicam, scopes, compound clips, …) is deferred by design
• Reverse playback: J in the timeline pauses; true reverse playback is viewer-only
• 4K media plays at degraded fidelity (frame drops accepted; no proxy editing pipeline)
• Codecs WebCodecs cannot decode natively (e.g. ProRes) play through an auto-generated H.264 preview proxy; export always re-renders from source
• Export encodes H.264/AAC MP4 only; no ProRes/HEVC encode
• Auto-transcribe skips assets over 30 minutes (manual transcribe still works at any length)
• Windows-only: packaging, binary fetching, and E2E are wired for win32

📄 License

MIT — free for everyone, for any use.

The fetch-binaries script downloads third-party tools that are not part of this repository and carry their own licenses: FFmpeg/FFprobe (GPL builds from gyan.dev), whisper.cpp (MIT), and the Whisper base.en model (MIT).

Support

If my tools save you time, you can support my work here: