mcp server
artdaddy
Agent-driven desktop video editor. Local ffmpeg rendering, MCP-native.
Set up
Runs on your computer. Pick your agent and operating system.
This project does not publish an install method we can turn into a config. Follow the README instead.
README
synced Oct 21, 2026The AI video editor your agent can actually drive.
Windows 10/11 · macOS · Linux x86_64 preview
ArtDaddy is a desktop video editor with an AI agent inside it. The agent edits through the same operations you do — place, trim, ripple, grade, caption, export — so its work lands in your timeline and your undo history, not in a separate preview it hands back.
That whole surface is also an MCP server. Point Claude, Cursor or Codex at it and they drive the real editor, on the real project, alongside you.
Everything renders on your machine
ffmpeg and whisper.cpp ship inside the app. Cutting, compositing, colour, captions, transcription and the final export all run locally — no upload, no round trip, no per-minute render bill.
Generation in the timeline
Images, video, voiceover and music generate straight onto tracks. Placeholders appear instantly and fill in when the job lands, so you keep cutting while a shot renders.
Built for agents, not bolted onto
55 tools covering the whole editor, with one rule that keeps agents honest: timeline positions are project frames, source positions are seconds, and the tools convert between them. No agent has to do frame-rate arithmetic, which is where most of them get it wrong.
MCP server
While the app is running it serves MCP on http://127.0.0.1:19787/mcp.
claude mcp add --transport http artdaddy http://127.0.0.1:19787/mcp
codex mcp add artdaddy --url http://127.0.0.1:19787/mcp
The app installs these for you. Open the Connect agent chip above the chat composer and pick your client — Claude Desktop, Cursor or VS Code / Copilot install in one click, and the panel shows the command to paste for anything else.
The bridge hands the agent the same tools the in-app agent gets, and refuses anything the catalogue does not offer.
Build from source
npm install
npm run fetch:sidecars # ffmpeg, ffprobe, whisper-cli, yt-dlp
npx tauri dev # desktop app
Requires Node 20+ and the Rust toolchain (Tauri v2). On Windows you also need WebView2, which ships with Windows 11.
npm test runs the suite and needs no server, so a change can be verified end to end offline.
Linux release
Run the Linux build workflow to produce an x86_64 AppImage and .deb on Ubuntu 22.04.
A normal dispatch builds and verifies Actions artifacts. publish: true attaches stable-name
Linux assets to the existing release for the current version; publish Windows first so that
release already exists. Enable enable_auto_update only when the existing updater private key
is configured as TAURI_SIGNING_PRIVATE_KEY; otherwise first-install downloads are published
without an updater signature.
Linux users can make the AppImage executable and launch it directly:
chmod +x ArtDaddy-linux-x86_64.AppImage
./ArtDaddy-linux-x86_64.AppImage
Claude Code uses the same local HTTP MCP endpoint on Linux:
claude mcp add --transport http artdaddy http://127.0.0.1:19787/mcp
claude mcp list
The Claude Desktop .mcpb remains Windows/macOS-only. Linux users should use Claude Code,
Cursor, VS Code/Copilot, or Codex.
FAQ
Editing, rendering and exporting are free — they use the bundled binaries on your own hardware. AI generation and vision run on hosted models and consume credits.
Yes. The app signs in before use, and AI features call a hosted backend that is not part of this repository. Once signed in, editing keeps working without a connection.
Windows 10/11, macOS, and an x86_64 Linux preview built against Ubuntu 22.04.
Code-signing certificates are in progress, so until they land Windows shows a SmartScreen prompt ("More info" → "Run anyway") and macOS asks you to allow the app on first launch. Updates themselves are cryptographically signed and verified already — that part is separate, and has been in place from the start.
No. This repository is the editor — the timeline, renderer, preview, tool implementations and the MCP server. The hosted service that brokers model calls stays closed.
Yes. Set VITE_API_BASE_URL before building, or change the server address in
About → Server. The API surface it expects is visible in src/api/.
It goes through the same mutation path the UI does, under one project lock and one undo history. Anything it changes, you undo the way you undo your own work.
Contributing
Issues and pull requests are welcome. Contributors sign a CLA; see CONTRIBUTING.md for the workflow and what the test suite expects.
License
GPL-3.0. Third-party components and their licences are listed in THIRD-PARTY-NOTICES.md.