OpenCut's Ground-Up Rewrite: Building an Open-Source Video Editor with AI Agents as First-Class Citizens

A Counterintuitive Decision

In 2025, OpenCut was already the most popular open-source video editor on GitHub — a free CapCut alternative with hundreds of thousands of users and daily traffic on opencut.app. Then the team made a counterintuitive call: rewrite from scratch.

Not incremental refactoring — a clean break. The old version was renamed opencut-classic, and the new version is being written from zero. During the rewrite, opencut.app still runs the old version, while the new one grows gradually at new.opencut.app.

This looks risky — rewrites are famously "the thing you should never do" (per Joel Spolsky's classic 2000 essay). But the OpenCut team is betting that video editors need their architecture redefined for the AI era, and the old architecture can't get there.

Six Commitments of the New Architecture

The rewritten OpenCut's README lists six goals:

1. Editor API — all editor capabilities exposed via API 2. Plugin-first architecture — third-party plugins as first-class citizens 3. Unified cross-platform codebase — desktop, mobile, and browser from one codebase (Rust core) 4. MCP server — an interface for AI agents 5. Headless mode — automation and batch rendering 6. In-editor scripting — write scripts directly inside the editor

These six aren't random. Together they point in one direction: the video editor evolving from a "GUI tool for humans" into a "programmable rendering platform".

MCP Server: Letting AI Agents Edit Video

This is the most interesting of the six. MCP (Model Context Protocol), proposed by Anthropic in 2024, lets AI agents call external tools. A built-in MCP server means agents like Claude Code and Codex can directly operate the video editor.

Imagine these scenarios:

• "Cut this 10-minute video into five 2-minute shorts" — the agent calls OpenCut's cut API
• "Add subtitles using Whisper transcription" — the agent calls transcription + subtitle APIs
• "Batch-add intros and outros to 100 videos" — the agent uses headless mode for batch rendering
• "Swap this video's background music for something more upbeat" — the agent calls an audio replace API

In traditional editors, these tasks are either impossible or require hand-written scripts (After Effects ExtendScript, DaVinci Lua scripting). OpenCut exposes these capabilities over the MCP protocol so agents can call them directly.

This is isomorphic to ego-lite's approach: not bolting an agent tool onto a product, but designing the agent interface into the product's architecture. The difference: ego-lite does browsers, OpenCut does video editing.

Rust Core: The Key to Cross-Platform

Choosing Rust for the core isn't arbitrary. Video editors are extremely performance-sensitive — real-time preview, multi-track compositing, and effect rendering are all CPU-intensive. Traditional cross-platform approaches (Electron, Tauri) hit performance bottlenecks in video processing.

A Rust core plus native UI per platform means:

• Video decoding, encoding, and effects in Rust (high performance, memory safety)
• Native UI layers per platform (desktop, mobile, browser)
• One core logic, different UI shells

It resembles Blender's architecture — C/C++ core with native UIs — but Blender is a 30-year-old codebase while OpenCut starts from zero. Rust's memory safety makes "from scratch" much cheaper: you don't need 30 years of bug fixes to reach production-grade stability.

Plugin-First: The VS Code Model

OpenCut explicitly targets a "plugin-first architecture": a minimal core editor, with all advanced features (effects, transitions, subtitle styles, export formats) delivered as plugins.

That's the VS Code model. VS Code's core is just a text editor; all language support, themes, and debuggers are plugins. That architecture let VS Code overtake Sublime Text's decade-long position in five years.

Video editing has no true "VS Code" yet. CapCut is closed-source; DaVinci Resolve has a plugin SDK but isn't open-source; Kdenlive is open but has a weak plugin ecosystem. If OpenCut builds a plugin ecosystem, it could become video editing's VS Code.

fal.ai — a generative AI platform — is already an OpenCut sponsor, suggesting the plugin ecosystem may include AI generation: text-to-video, image-to-video, AI voiceover, AI subtitles.

Headless Mode: The Editor as a Service

Headless mode is the most overlooked but commercially valuable of the six. Traditional editors are GUI tools — you must open an interface. Headless mode means the video editor can run as a backend service, enabling:

• Automated video production pipelines (RSS → video → publish)
• Batch processing (watermarking, transcoding, cutting)
• Video generation APIs for web apps
• CI/CD integration (commit → auto-generated demo video)

Combined with the MCP server, headless mode makes OpenCut a "programmable video engine" — not just an editor, but video-processing infrastructure.

The Essential Difference from CapCut

CapCut is a ByteDance product. Its business model: free tool + cloud AI features + data collection. Your videos are uploaded to ByteDance servers for AI processing.

OpenCut is MIT-licensed open source. Its business model is still unclear (sponsor fal.ai supports development), but the architecture guarantees one thing: your video data stays local. The MCP server lets agents operate videos locally; headless mode runs batch processing locally.

This is the "local-first vs cloud-first" architecture philosophy divide. Cloud-first (CapCut, Adobe Premiere Rush) dominates video editing because AI features depend on cloud compute. OpenCut's bet: local compute (Rust core + GPU) plus AI agents (MCP) can beat cloud AI.

Whether that bet wins is unknown, but the direction is right. As local GPUs get stronger (Apple M4, NVIDIA RTX 5090), local 4K processing is no longer a problem. AI agents mean "local tools" no longer imply "manual operation" — agents can do what previously required cloud AI.

A Deeper Observation

OpenCut's rewrite illustrates one thing: AI-era tools aren't "old tools with AI features added" — they're "new tools designed for AI at the architecture level".

Compare two categories:

• Old architecture + AI features: Photoshop with generative fill, Premiere with auto-editing
• New architecture + AI-native: OpenCut (MCP server + headless + plugin-first)

Old-architecture products add AI as a feature — users still operate through a GUI, and AI is just a button. New-architecture products treat AI as a first-class citizen — agents can call every capability directly, and the GUI is merely a human-friendly interface to agent capabilities.

It's like the difference between a feature phone with a browser and a smartphone. Feature phones could browse the web, but the browser was an add-on. Smartphones were designed for the internet at the architectural level — browser, app store, push notifications all first-class.

OpenCut is the "smartphone attempt" in video editing. Success isn't guaranteed — rewrites are high-risk, and many projects die along the way. But if it succeeds, it won't define a better video editor; it will define a new architectural paradigm for tool software in the AI era.

---

Project: OpenCut-app/OpenCut

Website: opencut.app

Sponsor: fal.ai — generative image, video, and audio model platform