Repository avatar
AI Tools
v0.2.1
active

shotcut-mcp

io.github.NoBanks/shotcut-mcp

Edit video through Shotcut's MLT engine: timelines, cuts, filters, titles, crossfades and renders.

Documentation

shotcut-mcp

MCP server that edits video through Shotcut's engine, MLT, with no GUI. An agent can create a project, add and trim clips, split, move and remove them, apply filters, add titles and crossfades, read the timeline back and render it with melt.

The output is a native Shotcut .mlt project, not a locked video file. Open the agent's cut in Shotcut and keep editing it by hand.

Requirements

  • Python 3.11+
  • MLT melt on PATH, or set MELT_BIN to its full path. Installing Shotcut includes melt.
    • Linux: sudo apt install melt
    • macOS: install Shotcut, then set MELT_BIN=/Applications/Shotcut.app/Contents/MacOS/melt
    • Windows: install Shotcut, then set MELT_BIN to melt.exe inside the Shotcut folder. Windows rendering is not yet covered by CI.
  • ffprobe (from FFmpeg) on PATH for reading clip durations and checking renders. Override with FFPROBE_BIN.

Install

Run it with no install step using uv:

uvx shotcut-mcp

Or install it with pip:

pip install shotcut-mcp
shotcut-mcp

The server talks MCP over stdio.

Use it with goose

goose Desktop or CLI

  1. Run goose configure, choose Add Extension, then Command-line Extension.
  2. Name: shotcut
  3. Command: uvx shotcut-mcp
  4. Timeout: 600 seconds or more. Renders take longer than goose's default.
  5. Add the environment variables SHOTCUT_MCP_WORKDIR and, if melt is not on your PATH, MELT_BIN.

config.yaml

Add this to ~/.config/goose/config.yaml:

extensions:
  shotcut:
    enabled: true
    type: stdio
    name: shotcut
    cmd: uvx
    args:
      - shotcut-mcp
    envs:
      SHOTCUT_MCP_WORKDIR: /path/to/your/projects
    timeout: 600

Then ask goose something like: "Make a project called promo from clip1.mp4 and clip2.mp4, keep the first 5 seconds of each, crossfade them, add the title 'Launch day' and render it."

Use it with other MCP clients

Any MCP client that runs stdio servers works. For example, Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "shotcut": {
      "command": "uvx",
      "args": ["shotcut-mcp"],
      "env": { "SHOTCUT_MCP_WORKDIR": "/path/to/your/projects" }
    }
  }
}

Configuration

VariableDefaultWhat it does
SHOTCUT_MCP_WORKDIR~/shotcut-mcp-projectsWhere .mlt projects and their state files are saved
MELT_BINmeltPath to the melt binary used for rendering
FFPROBE_BINffprobePath to ffprobe, used to read source durations

Tools

ToolWhat it does
create_projectCreate a new project with a frame size and frame rate
add_clipAppend a span of a source video to the end of the timeline. Returns the clip's id
trim_clipChange a clip's source in and out points
remove_clipRipple-delete one clip
move_clipMove a clip to a new position on the timeline
split_clipSplit a clip into two at a point inside it
remove_segmentRipple-delete a span of the timeline
append_text_overlayBurn a text title onto the timeline between two times
apply_filterApply brightness, volume, crop, fadeInBrightness, fadeOutBrightness, fadeInAudio or fadeOutAudio to a clip
set_transitionCrossfade video and audio between two adjacent clips
get_timelineReturn the whole edit: clips, source in and out, timeline positions, filters, transitions and titles
project_infoReturn summary metadata for a project
renderRender a project to a video file with melt
list_projectsList the projects in the work directory

Clip ids stay stable across calls. The edit state is kept in a .state.json file next to each .mlt, and the .mlt is regenerated from it after every change.

Current limits

  • One video track. Multitrack video and audio is planned.
  • Edits made by hand in the Shotcut GUI are overwritten by the next tool call. Safe round trips are planned.
  • The style option on set_transition is accepted but only a dissolve is rendered today.

Development

pip install -e ".[dev]"
pytest

The render tests are skipped when melt is not installed.

License

MIT