Monet

<h1 align="center">Monet</h1> <p align="center">AI video editor and design canvas for coding agents</p> <p align="center"> <img src="./resources/monet-mark.png" alt="Monet logo" width="120" /> </p> <p align="center"> Local-first video editing, a live design canvas, terminal-native agent workflows, real timeline operations, local transcription, and export in one macOS app. </p> <p align="center"> <strong>Monet is no longer just an AI video editor.</strong> Use Claude Code or Codex directly inside Monet Canvas to design vector graphics, animate, simulate physics, and drop the result straight into your video timeline. </p> <p align="center"> Built by Het Patel </p> <p align="center"> <code>curl -fsSL https://raw.githubusercontent.com/Monet-AI-Editor/Monet/main/install.sh | bash</code> </p> ## What Monet Is Monet is a desktop video editor designed for people who want coding agents to work on video directly, not through a fragile bridge into another editor. The app combines: - a real timeline and preview - local project files with autosave and recovery - an embedded PTY terminal for Claude Code, Codex, and other terminal agents - deterministic editor tools through `editorctl`, a local API bridge, and MCP - local transcription with `faster-whisper` - semantic search powered by embeddings when an OpenAI key is configured - the same OpenAI key can also be reused for GPT Image 2 generation Monet is not a chat demo wrapped around a timeline. The terminal, project graph, and editor runtime are the product. ## Features <table> <tr> <td width="42%" valign="top"> <h3>Agent-native terminal</h3> Claude Code and Codex can run inside Monet's built-in terminal with project-aware context, live editor access, and deterministic commands through <code>editorctl</code>. </td> <td width="58%"> <img src="./resources/ai_demo.png" alt="Monet terminal and preview" width="100%" /> </td> </tr> <tr> <td width="42%" valign="top"> <h3>Real timeline editing</h3> Import media, scrub, trim, split, move, duplicate, ripple edit, generate captions, add markers, and export from a real multi-track timeline. </td> <td width="58%"> <img src="./resources/timeline.png" alt="Monet timeline" width="100%" /> </td> </tr> <tr> <td width="42%" valign="top"> <h3>Monet Canvas — design alongside the editor</h3> A new artboard-based design surface inside Monet. Ask Claude Code or Codex to draw with <strong>Paper.js</strong>, simulate physics with <strong>Matter.js</strong>, or build any HTML/CSS/JS frame. <td width="58%"> <img src="./resources/canvas.png" alt="Monet Canvas" width="100%" /> </td> </tr> </table> - **Terminal-first AI workflow** — Use Claude Code or Codex directly in the app instead of relying on a built-in chatbot abstraction - **Monet Canvas** — design surface with Paper.js + Matter.js + raw HTML modes, live agent loading per frame, auto-export to the video media library, JSON export/import for canvas state, and PNG render for individual frames - **Deterministic editing tools** — `editorctl`, local API bridge, and MCP entrypoint for inspect/edit/export operations across both timeline and canvas - **Local transcription** — Uses `faster-whisper` on-device by default, with OpenAI fallback when configured - **Semantic search** — Search spoken content, metadata, and embedded project context - **Autosave and recovery** — Reopen current work, recover sessions, and manage multiple saved projects - **Export pipeline** — 720p, 1080p, and 4K outputs, export progress UI, and one-click `Show in Finder` - **React video composition** — [Remotion](https://www.remotion.dev) is built in with `PaperCanvas` and `PhysicsScene` compositions plus title cards, lower thirds, kinetic text, audio visualizers — all auto-import as timeline assets - **Basic motion and compositing** — transforms, opacity, text overlays, chroma key, and layered export baking - **Privacy-aware telemetry** — Anonymous usage analytics are optional; Sentry handles crash reporting ## Why Monet Most AI video workflows today still depend on a human driving Premiere, Resolve, or another editor by hand. Even when tools exist, they usually stop at prompts, scripts, or rough cuts. Monet is built around a different model: - the project is a structured graph, not an opaque binary - the terminal is a first-class editing surface - agents can inspect the live editor state, import media, search transcripts, cut sequences, and export results The goal is simple: make video editing something coding agents can actually operate. ## Install ### macOS app Monet is currently macOS-first. Install with one command in your terminal — no Gatekeeper prompts: ```bash curl -fsSL https://raw.githubusercontent.com/Monet-AI-Editor/Monet/main/install.sh | bash ``` This fetches the latest release ZIP directly (bypassing browser quarantine), installs `Monet.app` to `/Applications`, and launches it. For a local packaged build from source: ```bash npm install npm run dist:mac ``` That produces stable release artifact names like: - `release/Monet-macOS-arm64.dmg` - `release/Monet-macOS-arm64.zip` - `release/mac-arm64/Monet.app` ### Run from source Requirements: - Node.js 18+ - `ffmpeg` on your system path - macOS - optional telemetry env vars via a local `.env` Optional local transcription runtime: ```bash npm run setup:local-transcription ``` Start the app: ```bash npm install npm run dev ``` If you want Sentry and Aptabase enabled locally or in release builds, set env vars outside git-tracked files: ```bash cp .env.example .env ``` Available vars: - `MONET_SENTRY_DSN` - `MONET_APTABASE_APP_KEY` - `VITE_MONET_SENTRY_DSN` For local use, `VITE_MONET_SENTRY_DSN` can usually match `MONET_SENTRY_DSN`. ## First Run On first launch, Monet asks for: - an OpenAI API key for embeddings and semantic search, with the same key reusable for GPT Image 2 generation - an explicit anonymous usage analytics choice - nothing else up front Claude Code and Codex installation help is shown next to the terminal when needed, instead of blocking onboarding. ### Codex sandbox Monet's embedded terminal automatically launches Codex with `-s danger-full-access` so that `editorctl` can reach the live app API on `localhost:51847`. Without this flag, Codex's default `read-only` sandbox blocks outbound localhost connections and every `editorctl` command falls back to file mode. This flag is injected by Monet's wrapper script (written to `userData/bin/codex` on each terminal start) and **only applies when Codex is launched from inside Monet**. Codex sessions started from any other terminal are unaffected. ## What Agents Can Do Through the embedded terminal, `editorctl`, API bridge, and MCP surfaces, agents can operate **both** the video editor and the design canvas. ### Video Editor - inspect the current project and active sequence - list assets, tracks, clips, markers, and segments - import media from disk - split, trim, move, duplicate, rename, and remove clips - add tracks and transitions - generate captions from transcript segments - create rough cuts and selects from search results - extract frames and create contact sheets - run transcription - generate images with GPT and import to media - export the active sequence ### Monet Canvas - create artboard frames in `paperjs`, `matterjs`, or `html` mode - run any Paper.js script (vector graphics, animation, generative art) - run any Matter.js scene (2D physics simulations) - inject inline HTML/CSS/JS for arbitrary web-native designs - update, rename, resize, or delete individual frames - export entire canvas state as JSON (full backup) or render any single frame as PNG - import JSON canvas state, adding to existing frames without overwriting - auto-import every rendered frame into the video editor's media library - show live grid loading overlays per frame as the agent works ### Tool Surface | Surface | Used by | Purpose | | --- | --- | --- | | `editorctl` CLI | Claude Code, Codex, terminal scripts | Deterministic timeline + canvas commands inside the embedded terminal | | MCP server | MCP-aware agent hosts | Same operations exposed as Model Context Protocol tools | | HTTP API bridge | Anything that can `curl localhost:51847` | Fallback path; auto-discovers a free port if the default is busy | | `[MONET]` UserPromptSubmit hook | Claude Code | Injects current `activeView` (canvas vs editor) into the agent's context every turn | ### Examples ```bash # Video editor editorctl get-state editorctl search-segments "terminal workflow" editorctl export /tmp/monet-export.mp4 high 1080p mp4 # Canvas — Paper.js, Matter.js, or raw HTML editorctl canvas-add-frame "Paper Layout" 1440 900 paperjs editorctl canvas-add-frame "Glitch" 1280 720 paperjs editorctl canvas-frames editorctl canvas-run-paperjs <id> "var c = new Path.Circle({ center: view.center, radius: 100, fillColor: '#7aa2f7' });" editorctl canvas-render-png <id> ~/Desktop/glitch.png editorctl canvas-export ~/Desktop/my-canvas.json ``` ## Local AI Stack Monet is intentionally local-first. - **Transcription**: local `faster-whisper` by default - **Embeddings**: OpenAI API key required - **GPT Image 2 generation**: the same OpenAI API key can be reused - **Export**: local `ffmpeg` - **Project files**: plain `.aiveproj.json` files plus autosaves If embeddings or transcripts do not exist yet, Monet should say that directly instead of pretending otherwise. ## Project Model Monet projects are structured, inspectable, and tool-friendly. That includes: - assets - sequences - tracks - clips - transcript segments - markers - captions - effects - exports This is what allows coding agents to operate the editor directly instead of scraping the UI. ## Current Scope Monet already supports: - timeline editing - preview and scrubbing - autosave / recover - local transcription - semantic search foundations - captions and markers - terminal-native agent workflows - packaged macOS app builds Monet is still evolving in areas like: - richer color finishing - deeper motion tooling - broader codec coverage - more advanced compositing - wider platform support beyond macOS ## Privacy Monet does not collect project content for analytics. That means: - no filenames - no prompts - no transcript text - no media content - no API keys Anonymous usage analytics are optional. Crash reporting is handled separately. Telemetry keys are injected from environment variables at build time and are not stored in the repository. ## Build ```bash npm run build ``` Other useful scripts: ```bash npm run dev npm run typecheck npm run build:cli npm run build:mcp npm run dist:mac ``` ## Repository Layout - `src/main` — Electron main process, project store, export, transcription, terminal, updater, analytics - `src/renderer` — React app UI - `src/cli` — `editorctl` - `src/mcp-server` — MCP server entrypoint - `resources` — screenshots and app icon assets - `release` — packaged macOS artifacts ## Status Monet is already usable as a real macOS app, but it is still best described as an early public alpha rather than a finished mass-market editor. If you want to help shape it: - open issues - test the app on real media - try the terminal agent workflow - report reliability problems before polish problems ## License Monet is released under the MIT License. ## Built with - [xterm.js](https://xtermjs.org) — terminal emulator powering the embedded shell - [Remotion](https://www.remotion.dev) — React-based video composition and rendering - [Paper.js](http://paperjs.org) — vector graphics scripting in Monet Canvas (paperjs frames) - [Matter.js](https://brm.io/matter-js/) — 2D rigid-body physics engine in Monet Canvas (matterjs frames) - [node-pty](https://github.com/microsoft/node-pty) — pseudo-terminal backend - [FFmpeg](https://ffmpeg.org) — video export and processing - [faster-whisper](https://github.com/SYSTRAN/faster-whisper) — local transcription runtime