cupcat

🧁 CupCat

The free, local, AI-native video editor for Windows.

Claude edits your timeline. Everything runs on your machine. No credits, no watermark, no account.

⬇  Download for Windows  ·  🌐  Website & docs

Free · one installer · nothing else to install

---

CupCat is a real timeline video editor where an AI agent is a first-class editor. You describe the edit in plain language — "cut this 30-minute interview to a highlight reel, remove the filler, add Italian captions, make it 9:16" — and Claude does it directly on your timeline, using ~100 real editing tools. Almost everything runs offline on your PC: transcription, captions, dubbing, stem separation, reframing, slow-mo, motion tracking, exports.

The whole point: an AI editor that is free, local, and yours. No credit meters, no subscription gates, no cloud upload of your footage, no watermark, no account. Just a fast editor with an agent that actually operates the timeline.

Highlights

• 🎬 The agent edits the timeline — ~100 real tools (cut, trim, ripple, keyframe, caption, mask, color, multicam, compound, export prep) driven by plain language.
• 🔒 Local-first & private — your media never leaves your machine; the AI toolbox runs on-device.
• 💸 Zero credits — every local feature is free and unlimited; optional generation uses your own account.
• 🪟 Windows-native — a single self-contained installer, no preinstalls, no Python.
• 🧩 Open source (GPLv3).

Two ways to drive the AI

1. In-app chat — talk to Claude inside the editor (Opus 4.8 / Fable 5 / Sonnet / Haiku), with @-mentions of your media, an interruptible run, and a live build-log of every tool call.

2. Claude Code / Claude Desktop over MCP — CupCat exposes your open project as a local MCP server. Point any MCP client at it and edit your timeline from your own Claude:

   claude mcp add --transport http cupcat http://127.0.0.1:19789/mcp
   

Both paths share the exact same tool surface — editing by chatting and editing over MCP are the same engine.

What it can do

Agentic editing — describe an edit, get it done. First-cut assembly from a folder of footage, transcript-based text editing (delete words → the video ripples), one-click filler removal ("um / uh / cioè …"), long→short clipping into vertical shorts with karaoke captions, hooks and titles, a Ctrl+K command palette that falls back to the assistant.

Local AI toolbox — free, offline, no credits:

Tool	What it does	Engine (local)
auto_rough_cut	Folder of footage → assembled first cut + music bed	ffmpeg scene analysis
Transcript panel	Edit the video by editing its text; one-click filler removal	Whisper (on-device)
auto_clips	Long video → N vertical shorts w/ captions, hooks, titles	Whisper + Claude
auto_reframe	Reframe to 9:16/1:1 by tracking the subject per shot	ffmpeg saliency
separate_stems	Split audio into voice + music	sherpa-onnx spleeter
smooth_slowmo	Fluid, motion-interpolated slow motion	ffmpeg minterpolate
track_motion	Pin text/stickers to a moving subject	template matching
dub_timeline	Transcribe → translate → speak → time-fit a dub track	Whisper + Piper TTS
add_captions / translate_captions	Karaoke captions + translations	Whisper + Claude
caption_sounds	Name what is heard — (applause), (laughter), (siren)	sherpa-onnx + CED (AudioSet)
auto_multicam	Cut a multi-camera conversation on who is speaking	diarization + face detection
auto_pan	Place each voice where its owner stands on screen	face detection + ffmpeg
suggest_thumbnails	Pick a cover by measuring focus, exposure, faces	Laplacian variance + YuNet
add_motion_graphic / make_transition	Claude writes an animated HTML/CSS overlay → alpha video	Edge headless + VP9
identify_speakers	Speaker diarization	sherpa-onnx
Semantic search_media	Find clips by name, generation prompt, or what's spoken	token index
blur_faces	Find every face, follow it, render an anonymised copy	YuNet (ONNX) + ffmpeg
stabilize_video	Two-pass smoothing for handheld footage	ffmpeg vidstab
auto_color	Measured exposure / white-balance fix, or match another shot	ffmpeg signalstats
apply_lut	Grade with any .cube / .3dl LUT pack	ffmpeg lut3d
match_loudness	Normalise to YouTube / TikTok / podcast / broadcast targets	ffmpeg loudnorm (2-pass)
repair_audio / enhance_audio	Rebuild clipping & clicks; remove hiss and hum	ffmpeg adeclip · afftdn
duck_music	Music drops under speech, returns in the gaps	ffmpeg sidechaincompress
auto_chapters	Topic chapters + YouTube timestamps from the transcript	Whisper + Claude
quality_report	Pre-publish check: loudness, clipping, dead frames, flash risk	ffmpeg (read-only)

Generative media — when you want to generate (never required, never metered by us): the agent calls Higgsfield through your own account (video, images, audio, upscaling, AI reframe, background removal).

Pro editing — multicam, compound clips (live nesting), adjustment layers, bezier keyframe curves, pen masks with feather, chroma key, HDR (Dolby Vision decode-once → HLG HEVC export), frame-exact lossless smart-cut, beat sync, markers, rich text, custom keyboard shortcuts, screen/webcam recorder, URL import, project versioning.

Export — H.264 · H.265 · AV1 (10-bit) · HDR HEVC · ProRes · FCPXML / FCP7 XML for pro NLEs · lossless stream-copy. Exports are always user-initiated — the agent prepares the edit and hands you the Export button; it never renders on your behalf.

Install

Just want to use it: download the installer from cupcat.meetaly.agency or the latest release — CupCat_<version>_x64-setup.exe — and run it. It's a self-contained Windows app; every local engine (ffmpeg, Whisper, Piper, sherpa-onnx, YuNet) is bundled. No preinstalls, no Python, no credits. CupCat checks for new versions on launch and offers a one-click download.

To use the in-app assistant or MCP, sign in to Claude Code on the same machine. To generate media, sign in to Higgsfield (higgsfield auth login).

Windows will warn you when you run the installer: it is not code-signed yet. You can check the file's SHA-256 against the value on the release page. How the installer is built, and where signing stands, is written down in the code signing policy.

How it works

Claude Code / Desktop ──MCP (http 127.0.0.1:19789/mcp)──▶  CupCat bridge (bun, compiled)
In-app chat ───────────────────────────────────────────┤   ├─ ~100 timeline + media tools
                                                        │   ├─ local AI: Whisper · Piper · sherpa-onnx · YuNet · ffmpeg
                                                        │   └─ Higgsfield CLI (generate / upscale, your account)
                                                        │ WebSocket (live state)
                                                        ▼
                              CupCat editor (React + Vite) ── @cupcat/editor-core
                              timeline · preview · media library · inspector · transcript

• packages/editor-core — pure-TypeScript timeline model: commands, undo/redo, keyframes, selectors. Framework-free and unit-tested.
• apps/web — the editor UI (React 19 + Vite + Tailwind).
• apps/bridge — the local process: hosts the MCP server + WebSocket, runs the AI toolbox, drives Higgsfield + ffmpeg. Compiles to a single sidecar binary.
• apps/desktop — the Tauri shell that bundles the bridge, the web UI, and every local engine into one NSIS installer.

Build from source

Requires Bun, Rust (for the Tauri shell), and the Windows build tools.

bun install
bun run typecheck          # editor-core + bridge + web
bun run build:core         # the shared model
bun --filter @cupcat/web build

# dev: run the web editor + bridge (uses port 19789)
bun run web
bun run bridge

The bundled AI engines (ffmpeg, Whisper/GGML models, Piper voices, sherpa-onnx models) live under apps/desktop/src-tauri/sidecars/ and are git-ignored because of their size — 1.3 GB. One command fetches all 403 of them from pinned upstream sources, verifies each download against a recorded SHA-256, and checks the assembled folder against a lock file:

bun run sidecars                # → apps/desktop/src-tauri/sidecars/
bun run sidecars:check          # verify an existing folder without downloading

cd apps/desktop/src-tauri
bunx @tauri-apps/cli build      # → target/release/bundle/nsis/CupCat_<version>_x64-setup.exe

Releases are built by .github/workflows/release.yml on GitHub-hosted runners, which is what makes them signable — see docs/SIGNING.md.

Roadmap

See ROADMAP.md for where CupCat is heading, and the releases for what already shipped.

License

GPL-3.0-or-later — see LICENSE and NOTICE.md. Generative AI is provided by Higgsfield via its CLI, under your own account — not bundled.