Source: ai-research/opencut-github-readme-2026-05-27.md (GitHub README fetched 2026-05-27). Repo: github.com/OpenCut-app/OpenCut. Homepage: opencut.app (classic) / new.opencut.app (rewrite, pre-release). Stars: 51.6k. Forks: 5.6k. License: MIT. Languages: TypeScript 92.6%, Rust 5.9%, WGSL 0.5%. Latest release: v0.3.0 (Apr 15, 2026). Last commit: May 18, 2026. Maintainer: mazeincoding (96 contributors).
Open-source CapCut alternative — a free video editor for web, desktop, and mobile under MIT. Currently being rewritten from the ground up with a Rust core and an explicitly agent-native feature set: an Editor API, plugin-first architecture, MCP server for AI agents, headless rendering for batch automation, and a scripting tab inside the editor. The classic version at opencut.app (repo: opencut-app/opencut-classic) is the version to use today; the rewrite lands at new.opencut.app when ready.
Key Takeaways
- 51.6k stars, 5.6k forks, MIT, 96 contributors. This is the largest OSS video-editor repo the wiki has tracked — adoption signal is unambiguous. Lead maintainer
mazeincoding; primary sponsor fal.ai (generative image/video/audio API). No paid tier; no upsell on the README. - The rewrite is the story, not the classic version. README was refocused on the rewrite on 2026-05-18 (
docs: refocus readme on rewrite). The classic codebase atopencut-classicstill powers opencut.app today; new.opencut.app is the rewrite’s home until it ships and takes over. - Six rewrite features make it agent-native, not just OSS-CapCut: (1) Editor API, (2) first-class third-party plugins, (3) one Rust-core codebase across web/desktop/mobile, (4) MCP server for AI agents, (5) headless mode for automation and batch rendering, (6) scripting tab directly in the editor. The combination positions OpenCut as the agent-driven NLE that the AI-video stack currently lacks.
- Rust + WASM is the architectural commitment. Recent commits show the GPU renderer migrated from WebGL to wgpu/WASM (2026-04-01), time utilities ported from TypeScript to Rust WASM (2026-03-31), and a WASM compositor added (2026-04-13). The 5.9% Rust + 0.5% WGSL share is hot — performance-critical paths move to Rust as the rewrite progresses.
- Plugin model parallels Claude Code skills. A “plugin-first architecture” + headless mode + in-editor scripting tab is the same pattern Claude Code uses for skills — agents can ship editor capabilities like prompts ship to other agents. Combined with the MCP server, OpenCut becomes a callable surface for Claude Code / Codex / OpenClaw / Hermes the same way HeyGen and Higgsfield already are.
- No outside contributions yet. The README explicitly says they’re not taking outside PRs while the architecture is being designed. Discord (
zmR9N35cjK) or GitHub issues for questions. ^[inferred] Likely opens once the Editor API stabilizes. - MIT license + browser deployment = no platform lock-in. opencut.app is hosted on Cloudflare Pages/Workers (
wrangler.jsoncin repo). Self-host the same code on your own Workers, R2, or static infra. There is no SaaS-only path — the rewrite ships web/desktop/mobile from one codebase.
Where this fits in the AI-video stack
| Tool | Layer | Trust model | Agent integration |
|---|---|---|---|
| OpenCut (rewrite) | NLE / timeline editor | OSS, self-hostable | First-class — MCP server + headless mode + Editor API + plugin system |
| Remotion | Programmatic motion graphics | OSS, React-based | Code-as-source (no UI affordance) |
| Hyperframes | HTML composition | OSS, Apache 2.0 | Skill-driven (Claude Code / Cursor / Codex / Gemini CLI) |
| Claude Code Video Toolkit | Multi-skill production workspace | OSS, MIT | Native (10 skills + 13 slash commands) |
| video-use | Conversational editor | OSS | Skill (Python + ffmpeg) |
| HeyGen | Avatar + translate + video generation | Managed SaaS | Skill bundle + MCP |
| Higgsfield | Generative image + video API | Managed SaaS | Skill bundle + MCP + Supercomputer |
| CapCut | NLE | Closed SaaS (ByteDance) | None |
OpenCut occupies a slot the OSS stack didn’t have: a full NLE that’s agent-callable. Today, agents either drive code-as-video tools (Remotion, Hyperframes) or generative endpoints (HeyGen, Higgsfield). OpenCut’s MCP + headless mode makes a timeline-based editor scriptable from an agent without requiring the user to be in the UI.
Architecture (extracted from repo)
- Rust core — animation helpers split by domain (
refactor: split animation helpers by domain, 2026-04-26); WASM compositor; GPU renderer on wgpu/WASM - TypeScript shell — 92.6% of the codebase; the web app surface, with apps/ subfolder housing platform entry points
- Turborepo + Bun —
turbo.jsonorchestration,bun.lockfor dependencies - Cloudflare Pages/Workers deployment —
wrangler.jsoncin repo - ESLint + Prettier (migrated from Biome on 2026-05-02) —
chore: switch from biome to eslint + prettier - Cursor-assisted development — several commits carry
Made-with: Cursormetadata in their messages
Try It
- Use the classic version today. Visit opencut.app and start editing — no account, browser-native. The classic codebase still ships at
opencut-app/opencut-classicif you want to self-host or fork. - Watch the rewrite branch. new.opencut.app is the pre-release landing page. The MCP server, headless mode, and Editor API will surface there first.
- Track v0.3.0 → v0.4 release notes. Latest tag is v0.3.0 (Apr 15, 2026). Whether the next minor release ships in classic-mode or rewrite-mode signals the rewrite’s maturity. Subscribe to GitHub releases.
- Join the Discord (zmR9N35cjK) when the Editor API draft drops — this is where third-party plugin schema discussion will happen.
- Compare against Claude Code Video Toolkit for your use case. CC Video Toolkit is already shipping with OSS-model stack + skills + slash commands; OpenCut’s rewrite isn’t there yet. Today CCVT is the right answer for agent-driven OSS video work. OpenCut becomes the right answer when the MCP server + headless mode land.
Open Questions
- MCP server schema. When the README says “MCP server (for AI agents)” — does the editor expose timeline operations (insert clip, trim, render) as MCP tools? Resources? Prompts? The exact surface determines whether Claude Code / Codex can drive non-trivial edits or only fire-and-forget commands.
- Editor API surface. Documented? Read-only or read-write? Auth? Available cross-platform (web/desktop/mobile) or per-shell?
- Headless mode CLI. What’s the invocation?
opencut render --project foo.opencut --output out.mp4? Or is it API-driven? - Plugin model. Are plugins JS/TS like browser extensions, or Rust crates compiled into the binary, or Markdown skills like Claude Code? The “first-class third-party plugins” claim is structurally vague in the README.
- Format support. Codec coverage (H.264 / H.265 / AV1 / ProRes)? Audio (AAC, Opus)? Subtitle formats (SRT, VTT, ASS)? Not stated.
- Render performance. Pure WASM means CPU-bound or wgpu-on-CPU-fallback. Does it use WebGPU when available? GPU acceleration story on desktop (native CUDA / Metal) vs browser?
- Funding sustainability. Single named sponsor (fal.ai). The README invites more. Is there a parent company or a Foundation behind this, or is it maintainer-led?
- Comparison to Olive Editor, Kdenlive, Shotcut, DaVinci Resolve Studio (free tier). OpenCut is the OSS-CapCut-alternative narrative — but those four are the existing OSS NLE field. Where does OpenCut win, lose, or differentiate?
Related
- AI Video Tools topic landing
- Claude Code Video Toolkit — the closest existing OSS / agent-native production workspace; nearest comparison point
- Remotion Motion Graphics — programmatic React-based video framework (different paradigm; OpenCut is timeline-based)
- HeyGen Hyperframes — open-source HTML composition framework (different paradigm; HTML-based not timeline)
- video-use — conversational editing via Claude Code (different paradigm; transcript-driven edits vs OpenCut’s timeline)
- Claude Code skills overview — pattern OpenCut’s “plugin-first architecture” likely echoes
- Claude Code plugins and marketplaces — adjacent design space; how OpenCut’s “first-class third-party plugins” might surface in Claude Code
- OpenMediaTools — sibling browser-native OSS media-processing project; smaller scope (converter/extractor), no NLE timeline, no agent surface