melt-7 MCP Server

An MCP server exposing melt-7 (the MLT Multimedia Framework's CLI rendering engine -- the same backend Kdenlive uses) for agentic video editing/composition: build a multi-track timeline with clips, transitions, filters, and text overlays, then render it to MP4.

Requirements

• melt-7 and ffprobe/ffmpeg on PATH (Fedora: dnf install mlt)

• Python 3.10+

pip install -r requirements.txt

Related MCP server: CapCutAPI-Complete MCP Server

Running

python mcp_server.py

Serves MCP over streamable HTTP on http://0.0.0.0:8001/mcp (port 8001, to avoid colliding with the wan2_2_t2v server on 8000 if both run on the same host).

Running as a service

A systemd user unit is provided for running the server persistently:

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
cp melt-7-mcp.service ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now melt-7-mcp.service

Check status/logs with systemctl --user status melt-7-mcp.service and journalctl --user -u melt-7-mcp.service -f.

Project model

Each call to create_project creates a project directory under projects/<id>/ holding project.mlt (the authoritative MLT XML, hand-rolled to a flat schema rather than Kdenlive's nested per-track-tractor convention), meta.json (display name/profile bookkeeping), and a renders/ directory for rendered output. Every tool re-reads project.mlt fresh and atomically rewrites it, so state survives server restarts -- there is no in-memory project cache.

Tools

Notes on reliability

melt-7's process exit code is not a reliable success signal -- it can exit 0 while logging a "failed to load producer" error and silently substituting a blank clip. render_project verifies success independently: the output file must exist, be ffprobe-readable, have a duration close to the project's expected duration, and stdout/stderr must contain no known failure markers.