代理應用與開發工具
Lumae Exposes the Video Editor’s Same Set of Undoable Commands Directly as MCP Tools
The macOS beta Lumae lets agents inspect the timeline, configure zooms around cursor clicks, preview frames, and export videos, with every change reusing the GUI’s atomic commands and undo history. Media processing remains local, but frames sent for model inspection still go to the connected agent service, and the product currently lacks an independent security audit.
Recently gaining attention in the developer community, Lumae integrates agents into the video editor’s data model instead of using mouse automation to operate the interface. Editing operations such as splitting, moving, trimming, speed changes, subtitles, voice-overs, and camera zooms are already implemented as serializable commands. The built-in `lumae-mcp` helper exposes the same commands over stdio to MCP clients such as Codex, Claude Code, and Cursor. `edit_project` can atomically group multiple commands into a single undo step, while `preview_edit` validates changes and returns the result without writing to the project, reducing the risk of an agent failure leaving the timeline in an inconsistent state.
Agents do not have to infer everything from the entire video alone. `get_project` returns each clip’s media time, timeline position, subtitles, zoom intervals, and the timestamps and screen coordinates of cursor clicks recorded during capture, allowing auto-zoom to be created directly around a specified action. When visual judgment is required, `render_frame` can retrieve either a composited or raw frame, while `sample_frames` arranges multiple timestamps into a labeled grid. Once editing is complete, `export_project` uses the same renderer as the GUI to generate H.264 or HEVC files. This design—combining structured state with inexpensive frame sampling—is better suited to iterative agents than placing an entire video into a multimodal context.
The privacy boundary still needs to be made clear. Recording, rendering, and transcription with Apple speech models are all performed locally on the Mac, and editing does not re-encode the original media. However, JPEG frames that an agent requests to inspect are sent to the model service to which it is actually connected; they do not always remain on the device. The product is still in beta and supports only macOS 26, while recording and screenshot operations are not yet available to agents. Its sandbox also restricts export destinations to `~/Movies`. In addition, the GitHub repository primarily hosts documentation and signed binary releases, so download availability should not be mistaken for full open-source availability. Teams working with unreleased product footage should still review the MCP client, the model provider’s data policies, and tool permissions before use.