Skip to content
Fibi66Public

About

No description, website, or topics provided.

Resources

Stars

19 stars

Watchers

0 watching

Forks

Latest commit

 

History

113 Commits

Folders and files

Repository files navigation

cutti

AI-powered video editing app for macOS. Import your footage, click Start — cutti's AI handles transcription, scene analysis, and edit suggestions.

Status

  • macOS — usable. Pre-built DMGs are published to Releases.
  • iOS (iPhone + iPad) — work in progress, not usable yet. The ios/CuttiMobile/ target builds against the simulator so the shared kit (shared/CuttiKit/) keeps cross-platform parity, but the iOS UI is incomplete and most features are stubbed out. Don't expect a working app from it. There's no TestFlight build and no ETA — the macOS app is the focus right now.

Minimum requirements (macOS)

To run the local AI features (Qwen3-ASR transcription, speaker diarization, on-device scene analysis):

  • Apple Silicon Mac (M1 or later) — Intel / Rosetta is not supported
  • macOS 14 Sonoma or later
  • 16 GB RAM recommended (Qwen3-ASR runs on the Apple Silicon GPU via MPS; 8 GB works but leaves little headroom alongside the editor)
  • ~8 GB free disk for the local ASR sidecar: a private Python runtime + venv (~2 GB, torch / transformers / etc.), the Qwen/Qwen3-ASR-1.7B weights (~3.5 GB), the Qwen/Qwen3-ForcedAligner-0.6B weights (~1.2 GB), plus the sherpa-onnx diarization model (~47 MB)
  • Internet connection on first launch to download the sidecar and model weights; offline afterwards. If the sidecar isn't installed (e.g. you skipped the prompt), cutti falls back to Apple's built-in SFSpeech recognizer for transcription.

Setup

Prerequisites

  • macOS 14+ with Xcode 16 (Swift 6 toolchain) installed
  • For iOS builds: brew install xcodegen

1. Build & run macOS

cd macos/CuttiMac
swift build
swift run

On first build, SwiftPM auto-downloads the vendored sherpa-onnx and onnxruntime xcframeworks (~45 MB total) from this repo's GitHub release into the SwiftPM cache. The Qwen3-ASR sidecar (private Python runtime + model weights, ~7 GB) is not fetched at build time — it's installed on demand from the in-app prompt the first time you transcribe, into ~/Library/Application Support/cutti/qwen-asr/ (runtime + venv) and ~/Library/Caches/cutti/qwen-asr/huggingface/ (model weights). Both directories are removed when you uninstall the sidecar from Settings.

2. Build iOS (contributors only — UI is incomplete)

The iOS target exists so the shared kit (shared/CuttiKit/) doesn't drift out of cross-platform parity. It does not currently produce a usable app — most surfaces are stubs. Build it only if you're working on CuttiKit and need to verify your changes still link on iOS.

cd ios/CuttiMobile
xcodegen generate           # MUST re-run after editing project.yml
open CuttiMobile.xcodeproj

For Simulator builds: works out of the box.

For device builds: open project.yml, set DEVELOPMENT_TEAM to your own Apple Developer Team ID and change bundleIdPrefix to your own reverse-DNS prefix, then re-run xcodegen generate. There is intentionally no TestFlight or App Store build pipeline yet.

Testing

# macOS app + cross-platform package (one combined run)
cd macos/CuttiMac
DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer swift test

# Shared package alone
cd shared/CuttiKit
DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer swift test

Known issues / future improvements

These are known limitations that we intend to fix in future releases. None of them block normal use — they just mean some workflows are rougher than they should be:

  • One-click first cut can't be paused or resumed. Once the pipeline starts (transcribe → scene analysis → audio quality → LLM edit), the only way to stop it is to quit the app or wait it out. Re-opening the project re-runs the analysis from scratch instead of picking up where it left off.
  • No mid-flight cancel for transcription. A long clip (e.g. a 24-minute lecture) can take 10–15 minutes to transcribe locally with Qwen3-ASR. Closing the analysis chat panel doesn't actually cancel the in-flight request; the sidecar keeps working until it finishes.
  • No background / "headless" analysis. Closing the app stops the local sidecar, which means transcribing a long clip requires keeping the app in the foreground for the full duration.
  • No per-segment re-analysis. If a single clip's auto-cut comes out wrong, you have to re-run the analysis on the whole project, not just that clip.
  • Limited language coverage for the local model. Qwen3-ASR currently ships with Chinese, Cantonese, and English aligners. Other languages fall back to Apple Speech and won't get per-word timing.

If you hit something else that feels broken or missing, please open an issue at https://github.com/Fibi66/cutti/issues.

License

AGPL-3.0.

About

No description, website, or topics provided.

Resources

Stars

19 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages