Editor while an AI agent works: a purple status bar with Follow, Background, Undo AI changes, Pause and Stop; timeline clips labelled with the AI's edits; the preview showing the timecode the AI is looking at (interface in Chinese)

Screenshot from the app's automated screenshot tests, using test-pattern media. The interface is in Chinese.

How it works

MCP, locally. The app runs an MCP server over Streamable HTTP on 127.0.0.1:39917/mcp (it picks a free port if that one is taken), protected by a bearer token and an Origin check against DNS rebinding. The bundled cf binary adds a stdio transport: cf mcp, for clients that launch a command, such as Claude Desktop.
53 tools, each with a JSON Schema and MCP annotations (readOnlyHint, destructiveHint): observe (project, timeline, clip, media, transcript, tasks), edit (import, insert, move, trim, split, delete, speed, volume, fades, transform, keyframes, color, filters, transitions, titles, auto-captions, ripple-delete ranges, chroma key, relink…), output (MP4, SRT, stills) and navigation (open projects, switch panels, move the playhead, play a range for the human). Results come back as text plus structuredContent.
The agent can see. render_frame returns a PNG of the finished picture at any time, and contact_sheet tiles evenly spaced frames with timecodes into one image. Both use the same renderer as export, so what the agent sees is what you'll ship.
One edit, one undo step. Each AI edit lands on the shared undo stack, labelled “AI: …”. apply_ops commits several operations atomically: all succeed or none apply.
No lost updates. Every change bumps a project revision. Edits carry the revision the agent last read; if you changed the project in the meantime, the edit is refused and the latest state comes back so the agent can re-read and retry. Clients can subscribe to cf://project/current and get notifications/resources/updated when you edit.
Headless too. The cf CLI works on project files without the UI: info, apply (a JSON list of operations, all-or-nothing, with --dry-run), render with platform presets (Douyin, Kuaishou, WeChat Channels, Xiaohongshu, Bilibili, YouTube), transcribe, frame, contact-sheet, call for any tool, and serve for a UI-less HTTP MCP server. Handy for scripts and batch jobs.

You can see what the agent is doing

A purple bar appears under the title bar while the agent works, showing the current step, with Follow / Background, Undo AI changes, Pause and Stop.
Follow moves the UI with the agent: the project it opens, the panel it switches to, the frame it renders. Background leaves your view alone while the highlights still appear.
The agent gets its own color: a lavender playhead on the frame it's rendering, “AI is looking at 00:00:04:18” in the preview, and a short label next to each clip it changes.
Pause keeps it read-only; Stop refuses every request and cancels the exports and transcriptions it started.
Undo AI changes restores the project to how it was before the agent's session began, as a single step you can redo.
Exporting to a new folder, overwriting a file, deleting media and downloading an extension all ask you first; once you allow one, similar actions in that session go through.

While the agent samples frames, evenly spaced lavender lines appear across the timeline and the status bar shows that it is skimming

Setup

Open ContentFactory and click the AI button [AI] on the right of the title bar. The panel shows the server status, address and token, plus ready-to-copy snippets like the ones below. The token is generated randomly on your machine and can be regenerated at any time.

Claude Code

claude mcp add --transport http contentfactory http://127.0.0.1:39917/mcp --header "Authorization: Bearer <your-token>"

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "contentfactory": {
      "command": "/Applications/内容工厂.app/Contents/MacOS/cf",
      "args": ["mcp"]
    }
  }
}

While the app is running, cf mcp forwards to it, so you watch the agent work live; when it isn't, the tools run headless against the project files. On Windows, cf.exe lives in the install folder (by default %LOCALAPPDATA%\Programs\ContentFactory); the snippet in the panel always has the right path for your machine.

Any other MCP client (Streamable HTTP)

{
  "mcpServers": {
    "contentfactory": {
      "type": "http",
      "url": "http://127.0.0.1:39917/mcp",
      "headers": { "Authorization": "Bearer <your-token>" }
    }
  }
}

Supported MCP protocol versions: 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05. Read-only resources include cf://schema/project (the project JSON Schema), cf://presets/text, cf://catalog/filters, cf://catalog/transitions and cf://presets/export.

Things to ask

“Cut the pauses and filler words from this talking-head clip and add captions.”auto_subtitles transcribes on-device → get_transcript returns pauses and fillers as time ranges → remove_ranges ripple-deletes them. You see each step and can undo any of them.
“Make it 9:16, put a title at the start, and show me how it looks.”set_project changes the canvas → add_text adds the title → render_frame checks the result, and the agent nudges the position if needed.
“Skim the whole edit and tell me where the pacing drags.”Read-only: get_timeline for structure and contact_sheet for pictures, then suggestions. Nothing in the project changes.
“Export it for Douyin to my desktop.”export_video with the douyin preset (1080 short side, 30 fps) runs in the background; the first write to your desktop asks you first, and progress shows in get_tasks.

What the agent can and can't do

Can

Read the project, timeline, every clip parameter and keyframe, and per-character caption timing
See the rendered output, frame by frame or as a contact sheet
Make almost any edit the app supports, including atomic batches
Export video, captions and stills, and download extensions (with your OK)
Drive the UI so you can follow along, or work quietly in the background

Can't

Connect from another machine: the server binds to 127.0.0.1 and requires the token
Write to a new location, overwrite files or delete media without asking you
Think for itself: there's no model in the app; you bring your own agent, and the edit is only as good as it is
Hear audio: it understands content through transcripts and rendered frames

Privacy

ContentFactory calls no AI service itself and sends no project content anywhere. The MCP server listens only on your machine and rejects cross-origin requests from web pages.
When your agent calls a tool, it receives the project structure, transcript text and any frames it asks for. Whether and how those reach your AI provider depends on the client you use and its privacy policy.
Every call (time, client, tool, arguments, result, duration) is logged locally to agent/calls.log in the app's data folder; open it from the AI panel.
To keep agents out entirely, switch the MCP server off in the AI panel.

Let an AI agent edit.Watch it work.