DaVinci Resolve MCP Bridge

Control DaVinci Resolve from an AI assistant over MCP. Works with local models (LM Studio, Ollama) as well as Claude, Cursor and others. Includes a local Whisper + vision suite, so a local model can "hear" and "see" your footage without any cloud service.

This is a personal fork of samuelgursky/davinci-resolve-mcp (MIT). Changes in this fork: a local_ai tool (Whisper transcription + LM Studio/Ollama vision), a separate motion-graphics MCP server (Remotion + Three.js templates, no Resolve needed), the installer always installs the Resolve bridge, the installer can set up LM Studio directly, promo for the upstream's paid app is removed, and this README is shorter.

Install

git clone https://github.com/Thefrizzy1/davinci-resolve-mcp-bridge.git
cd davinci-resolve-mcp-bridge
python install.py

The installer runs these steps in order:

1. Finds Resolve and asks two questions:

   • Which edition do you have, Free or Studio? Free means the server talks to Resolve only through the in-app bridge, never through external scripting (a Studio feature), and Studio-only actions are removed. See Free vs Studio.

   • Which tools should the AI see? Everything, or Resolve only. Resolve only drops local_ai, editor and knowledge, which keeps the tool list small for local models.

2. Creates a venv/ with the dependencies.

3. Always installs the in-app bridge into Resolve's Scripts/Utility folders. You need it on the free edition. On Studio it does no harm.

4. Asks whether to install the local AI extras (faster-whisper). If you say yes, you pick a Whisper model and the installer downloads it right away, so the first transcription doesn't stall on a download.

5. Asks whether to install the motion graphics extra (Remotion + Three.js). The default answer is no, since many people already have their own setup.

6. Asks "Install straight into LM Studio?", which writes ~/.lmstudio/mcp.json. After that you can pick any other clients.

7. Checks the connection to Resolve (through the bridge on the free edition).

To run it without the prompts:

python install.py --lmstudio --edition free --toolset resolve   # free edition, Resolve tools only
python install.py --lmstudio --local-ai --whisper-model small    # LM Studio + Whisper, download "small"
python install.py --motion                       # add Remotion + Three.js motion graphics
python install.py --clients lmstudio,cursor      # several clients
python install.py --clients manual               # just print the JSON config

Before connecting:

• Studio: set Preferences ▸ General ▸ External scripting using to Local.

• Free edition: restart Resolve, open a project, then run Workspace ▸ Scripts ▸ resolve_bridge. The server finds the bridge automatically.

Related MCP server: davinci-resolve-mcp-free

Editing with a local model (the fast way)

The full tool list costs about 36,000 tokens before your model reads a word, which is more than most local models can hold. Pick Lean editor in the installer (or run python install.py --toolset lean). The model then sees only the editor tool, about 700 tokens. If you install motion graphics, that's a separate server with one more small tool. Replies are compact text instead of pretty-printed JSON sent twice.

1. Before you chat, with your project open in Resolve (and the bridge running on the free edition): venv\Scripts\python scripts\prep_footage.py (on macOS/Linux: venv/bin/python scripts/prep_footage.py) This transcribes all speech clips once. Transcripts are cached, so later chats skip this step.

2. In the chat, the model works through five calls:

The model only picks sentences ("c3#4", or "c3#4-6" for several in a row) and says which part each starts. The server converts seconds to frames, adds 0.1 s before and 0.3 s after each sentence so no word is clipped, and places everything with no gaps.

A prompt that fits this workflow:

Make a 55-65 s video from my footage with the editor tool. Story: Hook (the result, on my own RTX 3060) → Problem (everyone uses cloud tools, I wanted local) → Workflow (what I built) → Catch (slow on 12 GB) → Get it (free on GitHub). Each line must follow from the last with a "but" or "so". Open with 3-5 s of the fox story, then my best complete sentences (no false starts or crew talk). Add 4-6 covers of 2-5 s over the lines they match, never over my first sentence, plus the music. Colours: talking Blue, story Green, 3D Orange. Put a part on the first line of each section. Do a dry run first, fix any problems, then build, and tell me the sentences you chose.

What's what

Local AI: transcription and vision

local_ai(action="status")
local_ai(action="transcribe", params={"clip_name": "A001_C003", "write_srt": true, "add_markers": true})
local_ai(action="describe",   params={"path": "/footage/interview.mov", "frames": 4})
local_ai(action="analyze",    params={"clip_name": "A001_C003"})   # both at once

• Target: use one of path (a file), clip_name, or clip_id (a Media Pool clip).

• Transcribe: uses faster-whisper on CPU, or on CUDA if it's available. You get back text and timed segments. Options:

  • write_srt saves an .srt next to the media.

  • import_srt also brings that .srt into the Media Pool.

  • add_markers puts one marker per spoken line on the clip.

  • translate: true outputs English.

  • The installer downloads the model you choose and makes it the default. base and small are quick on a CPU. medium, large-v3 and turbo are much more accurate but take minutes per clip without a GPU.

• Long jobs run in the background. Each call waits up to 25 seconds. If the work isn't finished by then, the call returns a job_id and a progress value, and local_ai(action="job_status", params={"job_id": ...}) collects the result (it waits too). This means LM Studio's tool timeout never cuts a transcription off. Jobs run one at a time, so you can start several clips and collect them in order. jobs lists them all.

• Describe: pulls a few frames (with ffmpeg, or OpenCV as a fallback) and sends them to the vision model loaded in LM Studio (default http://127.0.0.1:1234/v1) or Ollama. Load a vision model first, for example Qwen2.5-VL, Gemma 3 or LLaVA. If you don't name a model, it picks a loaded vision model automatically.

• The built-in media_analysis pipeline also uses faster-whisper as its first-choice transcription backend.

Install the extras later with venv/bin/pip install -r requirements-local-ai.txt (on Windows, venv\Scripts\pip).

Motion graphics: Remotion + Three.js

This is an optional extra. Install it with python install.py --motion (needs Node 18.18+).

It's a separate MCP server, motion-graphics (src/motion_server.py), with a single tool called motion. It doesn't need DaVinci Resolve or the bridge. It starts in under a second, and its tool description is about 300 tokens. The installer adds it to each client next to davinci-resolve.

motion(action="list_templates")
motion(action="still",  params={"template": "Text3D", "props": {"text": "SEASON 2"}})      # frame shown in the chat
motion(action="render", params={"template": "LowerThird",
    "props": {"name": "Ada Lovelace", "title": "Mathematician"}, "open": true})          # plays when done
motion(action="job_status", params={"job_id": "..."})                                    # if it said "running"

Renders are saved to Videos/Resolve MCP Motion. To put them in Resolve, use the Resolve server's editor(action="import", params={"folder": "renders"}). They then show up in editor footage and can be placed with editor build.

• Every template accepts durationSeconds, fps, width and height.

• Captions takes segments or an srt_path (for example one written by local_ai transcribe with write_srt).

• Every call replies within 20 seconds. Long renders return a job_id, and job_status waits up to 20 seconds for it. Asking for the same render again returns the same job instead of starting another. Renders run one at a time, and one that takes longer than 15 minutes is stopped with the error message.

• studio opens Remotion Studio so you can preview and tweak in the browser. You can also run cd motion-graphics && npm run studio.

• Claude Code and other shell tools can call the renderer directly: node motion-graphics/render.mjs --template TitleCard --props '{"title":"Hi"}' --out title.mp4

• Already have your own Remotion project? Skip the extra and set DAVINCI_RESOLVE_MCP_MOTION_PROJECT=/path/to/project. The tool will list and render your own compositions with your own Remotion version.

• To add a template, put a component in motion-graphics/src/templates/, register it in src/Root.jsx, and add an entry to templates.json.

• Licence: Remotion is free for individuals and companies of up to 3 people. Larger companies need a Remotion company licence. Three.js is MIT.

Codex

python install.py --clients codex writes [mcp_servers.davinci-resolve] into ~/.codex/config.toml (%USERPROFILE%\.codex\config.toml on Windows). It includes startup_timeout_sec = 120 and tool_timeout_sec = 600. Codex's defaults are 10 s and 60 s, which is too short for this server to start, especially on Windows. When a server doesn't start in time, Codex drops it silently and the tools just don't appear. On Windows the entry also passes SYSTEMROOT, APPDATA and similar variables, because Codex starts servers with a trimmed environment. Values you've set yourself in that table are kept. After installing, restart Codex and open a new chat.

LM Studio tips

• The installer adds the server as davinci-resolve to ~/.lmstudio/mcp.json. Any other servers already in that file stay as they are. Restart LM Studio after installing.

• Use a model that's good at tool calling (for example Qwen 2.5/3 or Llama 3.1+ instruct) with a context of 16k or more, because the tool list is large.

• You can run a text model for chat and a separate vision model for local_ai describe at the same time.

Free vs Studio

The removed actions are the Neural Engine and Studio features: Resolve's own transcription and subtitle generation, scene-cut detection, magic mask, smart reframe, voice isolation, IntelliSearch, slate analysis, speech generation, motion deblur, audio classification, Dolby Vision analysis and stereo 3D clips.

They're blocked on purpose rather than just failing. On the free edition, such a call makes Resolve show an upgrade dialog, and every later call fails until you click it away. For transcription and subtitles, use local_ai instead.

The choice is saved in ~/.config/davinci-resolve-mcp/profile.json. Re-run the installer to change it. You can also override it with the environment variables DAVINCI_RESOLVE_MCP_EDITION=free|studio and DAVINCI_RESOLVE_MCP_TOOLSET=full|resolve. To hide any other tools, set DAVINCI_RESOLVE_MCP_DISABLED_TOOLS=a,b.

Free edition (in-app bridge)

Blackmagic only allows external scripting on Studio. The bridge is a small script that runs inside Resolve (Workspace ▸ Scripts ▸ resolve_bridge) and passes commands from the server over loopback, using HMAC-signed requests with one-time nonces. install.py always installs it. To reinstall it by hand, run python scripts/install_resolve_bridge.py. Once it's running, the server uses it automatically. Set DAVINCI_RESOLVE_BRIDGE=1 to force the server to use only the bridge. The connection to the bridge is kept open between requests. Older versions opened a new connection for every API call, and on Windows a busy session ran out of ports and the bridge seemed to "die". After updating, run python install.py (or python scripts/install_resolve_bridge.py) again and restart the bridge in Resolve so the copy inside Resolve is updated too.

Things to know

• Your source media is read-only. Analysis writes only sidecar reports and Resolve metadata. It never re-encodes or changes your camera files.

• Free edition, 21.1 and later: Blackmagic moved Python scripting to Studio, so the bridge reliably works only up to 21.0.x on the free edition.

• macOS with the free edition: Resolve finds Python only via PYTHON3HOME or /usr/local/bin/python3. If the bridge doesn't show up in the Scripts menu, run: launchctl setenv PYTHON3HOME "$(python3 -c 'import sys; print(sys.prefix)')" then restart Resolve. The resolve_bridge_canary script in the menu tells you whether Resolve can see Python.

• Safety: Risky edits use a plan, review, then confirm flow. Grade copies that would overwrite hand work are refused unless you pass acknowledge_trap: true.

More docs

• Install and client config

• AI skill reference

• Media analysis guide

• API coverage

• Control panel

• Security

Requirements

• DaVinci Resolve 18.5 or later (Studio, or the free edition through the bridge)

• Python 3.10–3.12 is the safest choice. 3.13+ works on recent Resolve builds.

• Optional: ffmpeg, Node 18.18+ (for motion graphics) / 20.9+ (for the advanced offline server)

License

MIT. Original work © DaVinci Resolve MCP Contributors (Samuel Gursky et al.). See LICENSE.