UltimateSlice

The AI connectable video editor.

UltimateSlice is a Final Cut Pro–inspired non-linear video editor built with Rust and GTK4, powered by GStreamer for media playback and export.

Built in MCP server allows for AI collaboration.

Features

✅ Implemented

• GTK4 application scaffold with dark theme styling
• Media library import with duration probing (video/audio/image)
• Source monitor with playback, scrubber, in/out marks, and timecode
• Timeline with multi-track rows, zoom/pan, clip selection, trim, move, razor
• Undo/Redo command history
• Inspector panel for selected clip properties
• MP4/H.264 export via GStreamer pipeline
• FCPXML 1.10–1.14 import + 1.14 export
• Optional MCP server (--mcp) for JSON-RPC control

Third-Party Crates and Libraries

UltimateSlice uses open-source crates and runtime libraries, including:

• gtk4-rs / gdk4 / gio / glib / pango — GTK/Pango runtime libraries LGPL-2.1-or-later
• gstreamer-rs + GStreamer — LGPL-2.1-or-later
• quick-xml — MIT
• serde / serde_json — MIT OR Apache-2.0
• uuid — MIT OR Apache-2.0
• anyhow / thiserror / log / env_logger — MIT OR Apache-2.0
• rustfft — MIT OR Apache-2.0
• fuzzy-matcher — MIT
• whisper-rs — Unlicense
• ort (ONNX Runtime) / ndarray — MIT OR Apache-2.0
• tokenizers (Hugging Face) — Apache-2.0
• hound (WAV I/O) — Apache-2.0
• resvg / usvg / tiny-skia — MIT OR Apache-2.0
• tempfile — MIT OR Apache-2.0
• FFmpeg (tooling/runtime) — LGPL-2.1-or-later (Flatpak build enables GPL options)
• x264 (Flatpak build dependency) — GPL-2.0-or-later

UltimateSlice can also use the following AI models, installed by the user at runtime (not bundled with the binary):

• MODNet — photographic portrait matting (background removal). Apache-2.0.
• SAM 3 — Segment Anything Model v3 (Meta AI) for mask generation from box prompts. Apache-2.0 (model weights from wkentaro/sam3-onnx ONNX exports).
• Whisper (GGML) — speech-to-text for subtitle generation. MIT.
• MusicGen-small — text-to-music generation (Meta AI). CC-BY-NC-4.0 (research/non-commercial use; check the model card before commercial use).
• RIFE — Real-time Intermediate Flow Estimation for AI slow-motion frame interpolation. MIT (model weights from third-party ONNX exports of the upstream RIFE project).

GPU acceleration

AI inference (SAM, MODNet, RIFE, MusicGen, and optionally Whisper STT) can run on GPU via ONNX Runtime execution providers. The default build is CPU-only; opt in to GPU with one of the ai-webgpu / ai-cuda / ai-rocm / ai-openvino Cargo features. Cross-vendor WebGPU is the recommended starting point (one feature flag, prebuilt binaries, works on Intel Arc / AMD / NVIDIA). Native EPs (CUDA / ROCm / OpenVINO) deliver higher peak throughput but CUDA needs the NVIDIA toolkit and ROCm/OpenVINO require source-building ONNX Runtime 1.24.2 via scripts/build_onnxruntime.sh. See docs/gpu/README.md for the decision tree and per-vendor setup guides.

For exact versions and full dependency tree, see Cargo.toml, Cargo.lock, io.github.kmwallio.ultimateslice.yml, cargo-sources.json, and onnxruntime-sources.json.

🔜 Planned

See ROADMAP.md for upcoming features like thumbnails, audio waveforms, multi-track editing, transitions, and a program monitor.

Project Structure

See docs/ARCHITECTURE.md for the full layout and design notes. Highlights:

• src/app.rs – GTK application setup and CSS loading
• src/model/ – core data model (Project, Track, Clip, MediaItem)
• src/media/ – playback, thumbnails, and export
• src/ui/ – GTK widgets (timeline, inspector, media browser, preview)
• src/fcpxml/ – FCPXML parser/writer

Requirements

• Rust (edition 2021, via rustup)
• GTK4 development libraries
• GStreamer + plugins for playback and export
• ffmpeg on $PATH (for export)

Linux (Ubuntu/Debian):

The default feature set builds whisper-rs from source, so install the compiler toolchain, pkg-config, and cmake alongside the media libraries:

sudo apt install \
  build-essential \
  cmake \
  pkg-config \
  libgtk-4-dev \
  libgstreamer1.0-dev \
  libgstreamer-plugins-base1.0-dev \
  gstreamer1.0-plugins-good \
  gstreamer1.0-plugins-bad \
  gstreamer1.0-libav \
  ffmpeg

macOS (Homebrew):

Install the Xcode Command Line Tools first:

xcode-select --install

Then install the Homebrew dependencies (the default speech-to-text feature needs cmake during cargo build):

brew install cmake gtk4 gstreamer gst-plugins-base gst-plugins-good gst-plugins-bad gst-libav ffmpeg

Then add to your shell profile so cargo can locate the libraries:

export PKG_CONFIG_PATH="$(brew --prefix)/lib/pkgconfig:$(brew --prefix)/share/pkgconfig"

Recommended System Specs

Three tiers targeting up to 4K source media, each with suggested UltimateSlice preference settings.

Minimum (1080p editing, basic 4K with proxies)

Component	Spec
CPU	Dual-core, 2.0 GHz+
RAM	4 GB
GPU	Integrated graphics (Intel HD / AMD APU), 512 MB shared VRAM
Storage	HDD (SSD recommended for proxy cache)

Settings: Proxy Mode → Half or Quarter, Preview Quality → Quarter, Hardware Acceleration → Off, GSK Renderer → Cairo, Playback Priority → Smooth

Recommended (4K editing)

Component	Spec
CPU	Quad-core, 3.0 GHz+
RAM	8 GB
GPU	Integrated or discrete with VA-API support, 2 GB VRAM
Storage	SSD

Settings: Proxy Mode → Off, Preview Quality → Half or Auto, Hardware Acceleration → On, GSK Renderer → Auto (OpenGL), Playback Priority → Balanced

Ideal (4K real-time, multi-track compositing)

Component	Spec
CPU	6+ cores, 3.5 GHz+
RAM	16 GB+
GPU	Discrete GPU with VA-API, 4 GB+ VRAM
Storage	NVMe SSD

Settings: Proxy Mode → Off, Preview Quality → Full, Hardware Acceleration → On, GSK Renderer → Vulkan, Playback Priority → Accurate, Real-time Preview → On

Notes:

• VA-API hardware decoding supports H.264, H.265/HEVC, VP9, and AV1.
• Export uses FFmpeg (CPU-based) — more cores = faster exports.
• Flatpak includes --device=dri for GPU access; native installs need VA-API drivers.

Build & Run

# from the project root
cargo build
cargo run

To run with MCP server enabled:

cargo run -- --mcp

To open a project file at startup from program arguments:

cargo run -- /path/to/project.uspxml

Python MCP Socket Client

When using the MCP socket transport (running instance), you can use the Python bridge client:

python3 tools/mcp_socket_client.py

Optional socket override:

python3 tools/mcp_socket_client.py --socket /tmp/ultimateslice-mcp.sock

The client reads JSON-RPC lines from stdin and writes responses to stdout. See docs/user/python-mcp.md for complete command examples.

Native Install

After building, run install.sh to install the binary, desktop entry, icons, MIME type, and AppStream metainfo to standard XDG locations:

# Install to /usr/local (default)
sudo ./install.sh

# Install to /usr (distro-style)
sudo ./install.sh --system

# User-level install (no sudo needed)
./install.sh --prefix=$HOME/.local

# Remove all installed files
sudo ./install.sh --uninstall

The script will automatically run cargo build --release if the binary is not yet built. Run ./install.sh --help for full usage.

Flatpak

A Flatpak manifest is provided at io.github.kmwallio.ultimateslice.yml.

Routine builds can use the checked-in cargo-sources.json and onnxruntime-sources.json source mirrors as-is. The first Flatpak build takes longer because ONNX Runtime is compiled from source inside the sandbox.

flatpak-builder --user --install --force-clean flatpak-build io.github.kmwallio.ultimateslice.yml
flatpak run io.github.kmwallio.ultimateslice

If Cargo.lock changes, regenerate the Rust vendored-source manifest before rebuilding:

python3 flatpak-cargo-generator.py Cargo.lock -o cargo-sources.json

onnxruntime-sources.json only needs to be refreshed when the pinned ONNX Runtime version or its mirrored CPU/shared-lib cmake/deps.txt inputs change.

Notes

• GTK4 callbacks cannot unwind panics. Avoid RefCell double-borrows in UI callbacks.
• The project shares a single GStreamer playbin for source and timeline playback.

License

This project is licensed under the GNU General Public License v3.0.