Palmier Pro: Operating an Agent-Connected Video Editing Workflow
A practical operations review of Palmier Pro: local MCP networking, project safety, build gates, cost boundaries, and recovery from agent mistakes.
Operating Palmier Pro safely means treating the timeline as shared state and the local MCP endpoint as a privileged control surface. The project is a macOS video editor built for AI, but the operational question is narrower: can an agent make bounded, observable changes without turning a creative project into an unrecoverable sequence of guesses?
Separate the editor from the generation service
The editor, MCP server, and in-app agent chat are open source under GPLv3. Generative AI processing is closed and requires login and subscription. A deployment plan should therefore track two separate dependencies: the local Swift application and the hosted generation account. Free editing does not imply free generation.
Build gates before automation
Supported machines are Apple Silicon Macs running macOS 26 Tahoe. Contributors can use Swift tools 6.2 and run swift build, swift test, and swift run. Before connecting a client, test project creation, media import, save/reopen, preview, and export. These checks isolate editor failures from model or MCP failures.
Keep MCP local
The app exposes http://127.0.0.1:19789/mcp only while it is open. Its server binds to loopback and validates origin, content type, protocol version, and session state. It also limits and expires sessions. Do not publish this port through a reverse proxy as a shortcut for remote editing; that would require a new authentication and authorization design.
Recovery-oriented agent use
Start with read-only inspection: list projects, inspect the timeline, locate a clip, and read a transcript. Then permit one named mutation at a time, such as replacing a clip or adding a marker. Save a project copy before multi-step edits and require visual review before export. A clear request should include the target project, exact range, expected result, and rollback point.
Capacity and product limits
The repository is early-stage and does not claim feature parity with Premiere Pro or CapCut. Its FAQ lists transitions, masking, and graphics as missing. That makes it a strong candidate for AI-media iteration and workflow experiments, but a poor assumption as the only editor for advanced post-production.
Operational decision
Adopt it for a controlled Apple Silicon pilot with one short production video and measured agent actions. Record client configuration, project backups, model cost, export results, and failure recovery. Expand only when the human review loop remains faster and safer than manual context switching.