Wasserman's Unofficial DaVinci MCP

Let any AI agent run DaVinci Resolve for you — Claude Code, Codex, Cursor, Hermes, or any MCP client. Cut clips to the beat, match a reference look, tighten dialogue, hit a loudness spec, and build timelines — either by driving Resolve Studio live, or by generating clean import files for the free edition.

A flat log-ish grade (BEFORE) matched to a warm reference frame (REFERENCE) → the MATCHED result, baked to a 33-point .cube LUT. Real output from the color_match engine on a single desert-dance frame.

You describe the edit in plain language; your agent calls the tools. The server does the deterministic, numeric work an LLM shouldn't guess at — beat grids, LUT math, EBU R128 loudness, FCPXML — and hands back files and plans you can trust. No cloud rendering, no black boxes: every engine is stdlib + numpy + ffmpeg, and every mutating tool previews before it touches anything.

Two tiers, auto-detected

The server detects at startup whether it can drive a running DaVinci Resolve Studio (verified on Studio 21) or should fall back to interchange files. resolve_capabilities tells you which tier you're in and why.

Being honest about it: external scripting is a DaVinci Resolve Studio feature. On the free edition the live tools can't reach the app — but every creative engine still runs, and the interchange writers give you a timeline, LUTs, markers, and a premix you import in a single step. Nothing about the free path is a stub.

Related MCP server: DaVinci Resolve MCP Server

Quick start

Install straight from GitHub:

pip install "git+https://github.com/wassermanproductions/unofficial-davinci-mcp.git"

Optional extras — beat-tracking (librosa) and the push-to-talk voice bridge:

pip install "unofficial-davinci-mcp[beats] @ git+https://github.com/wassermanproductions/unofficial-davinci-mcp.git"
pip install "unofficial-davinci-mcp[voice] @ git+https://github.com/wassermanproductions/unofficial-davinci-mcp.git"

Or clone and install editable:

git clone https://github.com/wassermanproductions/unofficial-davinci-mcp.git
cd unofficial-davinci-mcp
pip install -e .

You also need ffmpeg on your PATH (brew install ffmpeg, or your distro's package). That's the only non-Python dependency.

Wire it into your agent

Claude Code — one line:

claude mcp add davinci -- unofficial-davinci-mcp

Any MCP client (generic mcpServers JSON):

{
  "mcpServers": {
    "davinci": {
      "command": "unofficial-davinci-mcp"
    }
  }
}

Hermes — ~/.hermes/config.yaml:

mcpServers:
  davinci:
    command: unofficial-davinci-mcp

If the console script isn't on your PATH, use python -m davinci_mcp as the command instead. Then ask your agent to call resolve_capabilities first — it reports your tier, Resolve version, ffmpeg, and optional dependencies.

What it can do

Reference color matching

Point color_match at a reference still (or a graded frame) and the shots you want to match. It samples both in CIE Lab, computes a per-shot transform (Reinhard mean/std, or a firmer Lab-histogram match), and bakes a Resolve-loadable 33-point .cube LUT — plus a before/after preview strip and the numeric ΔE convergence. The hero strip above is a real run: a flat grade pulled to a warm reference, ΔE 43.9 → 3.2. On Studio it can apply the LUT for you; on free Resolve you get the file to drop on a node.

A worked example, end to end

"Cut these clips to the music, give it a warm look, and set dialogue at −16 LUFS."

One request, seven tool calls: read the tier, probe the media, find the tempo, assemble a beat-snapped edit, write an FCPXML, bake a warm LUT, and plan the mix — every number above is real output from the running server.

Beat-aware music cutting

beat_grid finds the tempo, beats, downbeat, and onset strengths (librosa when installed, an energy-envelope fallback when not). cut_music uses that grid to trim a track to a target length on a musical boundary, then rings the tail out under an exponential fade — or lands on a hit with a "button" ending. No cuts mid-phrase.

Dialogue, loudness, and the mix

• tighten_dialogue detects dead air in a talking clip and returns a keep-range cut plan (with head/tail handles) — apply it live, or export the tightened edit as FCPXML.

• measure_loudness reports EBU R128 integrated LUFS, loudness range, and true-peak per file.

• mix_plan normalizes dialogue to your delivery spec, sets the music bed under it, derives ducking windows from speech, and renders a re-measured premix so you can trust the number.

Build the timeline

assemble_edit validates an edit plan against your probed media and snaps cuts to a beat grid. Then generate_fcpxml writes an FCPXML 1.9 timeline that imports cleanly into Resolve 19/20, generate_edl writes a CM3600 cut list, and generate_marker_csv writes a marker manifest. On Studio, resolve_create_timeline / resolve_append_to_timeline build it in the app directly.

The 23 tools

Every mutating tool defaults to dry_run=true and returns a plan; re-run with dry_run=false and confirm=true to apply it. Tier both = works in free or Studio; live = requires DaVinci Resolve Studio.

It knows editing, not just buttons

Operating the app is half the job; knowing what a good edit is is the other half. get_editing_knowledge serves the same editorial playbook the tools were built around — written like a seasoned editor briefing an assistant, with numbers your agent can act on:

• Loudness by destination. Dialogue sits at −16 LUFS for web / −23 LUFS for broadcast; the music bed rides 4–8 dB under it; ducking is −6 to −8 dB with a 0.3 s attack / 0.5 s release; the master limiter caps at −1 dBTP. mix_plan takes these as parameters.

• Cut to the energy, not the metronome. At 120 BPM a beat is 12 frames — so cut a moving shot 1–2 frames early and the motion lands on the beat perceptually; hold 8–16 shots per 8 bars in the chorus and only 1–3 in the intro. assemble_edit's beat-snap tolerance tightens to ±2 frames for the drop.

• Match, don't invent. For teal-and-orange, reference a warm-skin/cool-shadow frame with method="lab_histogram" at strength 0.5–0.65, and always re-check skin last. Never bake exposure or white balance into a look LUT.

Topics: color-looks, beat-cutting, dialogue-editing, music-editing, mixing.

Talk to your editor

An optional macOS push-to-talk voice bridge lets you hold a key, speak, and drop the transcript straight into your agent terminal — hands on the footage, not the keyboard.

It runs entirely on your Mac: faster-whisper transcribes locally, nothing is sent to the cloud, and editing jargon (LUFS, J-cut, sting, FCPXML) is biased into the model so it transcribes cleanly. Setup, permissions, and configuration are in voice/README.md.

pip install "unofficial-davinci-mcp[voice] @ git+https://github.com/wassermanproductions/unofficial-davinci-mcp.git"
python -m voice

Requirements & troubleshooting

• Python ≥ 3.10, numpy, and ffmpeg on your PATH (checked in /opt/homebrew/bin, /usr/local/bin, /usr/bin as fallbacks). Install ffmpeg with brew install ffmpeg on macOS or your distro's package manager on Linux.

• librosa is an optional extra ([beats]) for tempo-tracked beat grids; without it, beat_grid uses an energy-envelope fallback that still works.

• Studio vs free. Live tools need DaVinci Resolve Studio, running, with external scripting enabled (Preferences ▸ System ▸ General ▸ External scripting using: Local). If you're on the free edition or Resolve isn't running, the server drops to interchange automatically — resolve_capabilities explains why.

• "Studio required" / handshake failed. The scripting module loaded but couldn't reach the app: confirm Resolve is open, you're on Studio, and scripting is enabled. Everything in the interchange tier still works regardless.

• The voice bridge needs Microphone + Accessibility permissions — see voice/README.md for the walkthrough.

Disclaimer

Unofficial. This project is not affiliated with, authorized by, or endorsed by Blackmagic Design. "DaVinci Resolve" is a trademark of Blackmagic Design Pty Ltd. This is an independent, community tool that talks to Resolve through its public scripting API and standard interchange formats.

License & credits

Apache License 2.0 — see LICENSE. Free to use, modify, fork, and build on, commercially or otherwise.

Attribution required: per the NOTICE file (Apache 2.0 §4(d)), any use, fork, or redistribution must retain the NOTICE file and credit Sam Wasserman (wassermanproductions.com) in its documentation and about/credits surface.

Created by Sam Wasserman — wassermanproductions.com · wasserman.ai.