The technical stack is lean—built in Swift to avoid the overhead of Electron/Node.js—and leverages native macOS APIs like CoreML and SpeechAnalyzer for local processing.
Core AI Workflow Capabilities #
Connecting Claude or Codex to the local MCP server allows the agent to handle the mechanical "grunt work" that usually kills creative momentum. Here is what the agent can actually execute:
Media Management: Importing files from URLs or local paths and organizing them into folders.Semantic Search: Finding footage via local SigLIP2 embeddings.Timeline Manipulation: Performing operations on tracks, clips, and keyframes.Asset Generation: Creating music, SFX, captions, and video clips within the editor.Local Analysis: Running Silero VAD for silence detection and beat_this for beat-matching.
Deployment and Setup #
Since this is designed for macOS, you get native performance without the webview lag. The integration works in two ways:
-
Local MCP Server: Best for complex, automated workflows (e.g., replicating a specific podcast editing style across multiple raw files).
-
In-App Chat: Better for quick, low-latency iterations while you have the timeline open.
For those looking to build a custom AI video workflow, the ability to trigger exports and edits via an LLM agent transforms the editor from a manual tool into a programmable interface. It's particularly effective for rough cuts based on transcriptions or beats, where the AI handles the alignment and the human handles the final creative polish.
If you are running macOS, it's a great way to test how an LLM agent can manage a heavy professional application without the usual API friction.
[Next Loop Engineering: Stopping Agent Reward-Hacking →](/en/threads/2183/)