YoloCut

A new kind of video editor

YoloCut is a local-first desktop editor where a human and an Agent work on the same real project. Trim and arrange clips on the timeline, or ask the built-in Agent, Codex, Claude Code, Gemini CLI, Cursor, or another MCP client to use the same editing commands.

This is not a one-shot video generator that returns an opaque result. Agent work becomes normal media, tracks, clips, captions, transitions, effects, keyframes, and export jobs. You can preview it, review the proposed changes, apply or reject them, undo them, and continue editing by hand.

Media + script + editing goal
        ↓
Agent creates a reviewable edit session
        ↓
Real timeline changes → preview → refine → export

Related MCP server: Cutible MCP Server

Why YoloCut

Product tour

All screenshots above were captured from the current YoloCut v0.0.2 desktop application.

What is available in v0.0.2

Agent-native by design

The detachable Agent workspace is a control surface, not a decorative chat sidebar:

• It shares project state and editing commands with the manual UI.

• It can dock left or right, open as a separate wide desktop window, and return to the editor through corner docking.

• Mutating work is staged in an edit session for review before it is considered applied.

• The built-in Agent and external MCP clients use the same canonical 119-tool catalog.

• When no project is open, YoloCut exposes only the smaller server-direct surface and reports the limitation honestly.

Connect an external Agent

Install the YoloCut Agent Skill:

npx skills add Hhz0823/YoloCut --skill yolocut

Start YoloCut, open the target project, then open Agent Connection Center (MCP) and copy its live URL and Bearer token. The default endpoint is:

http://localhost:5199/api/external-mcp/mcp

The desktop application can choose another loopback port when 5199 is occupied, so clients should use the address displayed by the connection center rather than permanently hard-coding the default.

Required completion flow:

yolocut_status → get_connection_manifest → list_projects → target_project
→ load_skill / ToolSearch → begin_edit_session → editing tools
→ review_edit_session → get_edit_session (status=applied)

An Agent should report completion only when full editing is ready, capability coverage is complete, and the final edit session is applied. Configuration examples for Codex, Claude Code, Gemini CLI, Cursor, and generic Streamable HTTP clients are in YOLOCUT_AGENT_CONNECTION.md.

Batch auto-editing

Give YoloCut a media directory, edit script, narration script, and optional reference video. The durable queue accepts up to 10,000 jobs, with a separate project and lifecycle for every output instead of placing thousands of videos on one timeline.

1. Scan an authorized local media directory.

2. Attach the edit script, narration script, and optional reference cut.

3. Analyze rhythm, shot structure, captions, transitions, and color intent.

4. Let the Agent create reviewable edits with hardware-aware concurrency.

5. Render, quality-check, and read back success, failure, or cancellation per job.

Optional local reference analysis uses the Apache-2.0 SmolVLM2-500M-Video-Instruct-GGUF package with llama.cpp. Missing models or runtimes fail visibly; YoloCut does not fabricate an analysis result.

Long 4K projects on modest machines

YoloCut separates the interactive proxy path from final-quality media. Analysis and preview can use smaller derivatives while final export reads the original files.

• NVIDIA systems prefer NVDEC → scale_cuda → NVENC after live capability probes.

• H.264 routing probes NVENC, Intel QSV, AMD AMF, and then software libx264.

• AV1, VP9, and unsupported hardware paths can fall back to FFmpeg software decoders such as libaom-av1 and libvpx.

• CPU, memory, VRAM, codec support, and active workload decide proxy size and queue concurrency.

• Runtime results show the backend that was actually used and any fallback reason; model or GPU names alone are never treated as proof.

Fish S2 model weights use the Fish Audio Research License. Commercial use requires separate written permission. Hardware tiers are conservative policy contracts, not performance guarantees for every driver and machine.

Faster desktop startup

YoloCut v0.0.2 now stages desktop startup instead of blocking the first window on the complete editing stack:

• Electron displays a lightweight native startup shell before loading the embedded application server.

• The server starts in parallel, while GPU discovery continues asynchronously after the editor becomes usable.

• Models, thumbnails, updater checks, and the packaged render runtime are deferred until the relevant feature or post-startup window needs them.

• Renderer surfaces and dashboard dialogs stay behind lazy boundaries so the full editor graph is not part of the initial dashboard load.

• Desktop smoke tests use a unique temporary profile and data directory, preventing tests from touching a user's projects, credentials, media, or live YoloCut process.

On the Windows release-verification machine, the unpacked build displayed its native shell in about 85 ms, brought up the embedded server in about 0.72 s, and loaded the first renderer in about 2.79 s on a cold run; a warmed renderer loaded in about 0.51 s. These are diagnostic measurements from one machine, not cross-device guarantees. Set YOLOCUT_STARTUP_TRACE=1 when launching the desktop build to print stage timings for another system.

Download

GitHub Release assets intentionally contain only the directly installable Windows .exe and macOS .dmg files. Installer SHA-256 values are listed in the release notes rather than attached as another asset. Windows packages are currently unsigned; macOS packages are ad-hoc signed but not notarized.

Linux remains built and smoke-tested in CI but is not attached to GitHub Releases. Linux users can run YoloCut from source using the instructions below. Packaged builds use the in-app version check and open the GitHub Release page for manual installation when an update is available.

Run from source

Requires Node.js >=24 <25 and npm.

git clone https://github.com/Hhz0823/YoloCut.git
cd YoloCut
npm install
cp .env.example .env.local
npm run dev

Desktop development and Windows packaging:

npm run desktop:dev
npm run desktop:dist:win

Development profiles isolate projects, media, credentials, tasks, and settings per Git checkout/worktree. Configure only the AI or media providers you actually intend to use.

Data and security

• Projects, chat, versions, indexes, and settings are local by default under ~/.yolocut.

• Existing legacy data roots can be mounted by the migration layer so users do not lose earlier projects.

• Provider keys stay on the service side and are not exposed through VITE_ browser variables.

• MCP binds to loopback by default and requires a generated Bearer token.

• Directory access uses opaque grants; absolute local paths do not cross the desktop IPC boundary into Agent tools.

• AI traffic leaves the computer only for providers the user explicitly configures.

• YoloCut targets a single-user desktop workflow, not an unisolated multi-tenant server deployment.

Architecture and verification

npm test                         # full regression suite
npm run build                    # type checking and production build
npm run lint                     # static analysis
npm run verify:architecture      # enforced dependency boundaries
npm run verify:mcp               # Agent/MCP contract checks
npm run verify:media-performance # codecs, proxies, acceleration, fallbacks
npm run verify:auto-edit         # batch auto-edit contracts
npm run desktop:smoke             # embedded server and desktop bridge
npm run desktop:smoke:render      # real packaged rendering path
npm run desktop:smoke:post-startup # deferred startup work
npm run desktop:smoke:agent-window

Current release boundaries

YoloCut 0.0.2 is an early public release. Core editing, Agent, MCP, batch, desktop packaging, and local model management are actively verified, but several surfaces remain experimental:

• Fish S2, SmolVLM2, and some GPU paths depend on local runtimes and drivers.

• The project format and Agent catalog will continue to evolve; back up important projects before upgrades.

• Signing and notarization are not yet production-grade.

• A configured provider or installed model is required for its corresponding AI feature; manual editing remains available without one.

See CHANGELOG.md for version history and GitHub Releases for published artifacts.

License, provenance, and contributing

YoloCut is independently maintained by hhz0823 and licensed under the GNU Affero General Public License v3.0 or later.

The codebase continues from the AGPL-licensed 0xsline/OpenChatCut project. This attribution preserves source and license provenance; upstream branding, contributors, community channels, and commercial relationships are not part of the YoloCut product or its public release history.

Third-party libraries, models, fonts, skills, and bundled binaries retain their own licenses. See THIRD_PARTY_NOTICES.md, src/agent/skills/NOTICE.md, and assets/fonts/LICENSES.md.

• Issues: github.com/Hhz0823/YoloCut/issues

• Releases: github.com/Hhz0823/YoloCut/releases

• Agent guide: YOLOCUT_AGENT_CONNECTION.md