Let Claude drive DaVinci Resolve
A safe, step-by-step setup for the community DaVinci Resolve MCP server, so Claude Code or Codex can inspect a project and run real timeline, media, Fusion, grading and render work. Every agent prompt from the guide, plus the one habit that keeps it from wrecking your edit. Checked against the project’s own package
Connect an AI agent to DaVinci Resolve and it can inspect your project and do real work: cuts and trims, markers, timeline versions, media bins, Fusion graphics, supported grades and render jobs, all from plain-English prompts. This is the whole setup — install, enable scripting, the one installer command, the diagnostic, and wiring up Claude Code or Codex — followed by every test prompt from the guide, ordered from a read-only inspection up to a silence cleanup and a render. It is the community davinci-resolve-mcp project, not an official Blackmagic tool, and it drives Resolve through Resolve’s own scripting API, which Blackmagic gates to the Studio edition. The discipline that makes it safe is one habit: inspect, then act, then verify, always on a duplicate.
An AI agent — Claude Code or Codex — wired to DaVinci Resolve so it can read your project and perform supported operations in it. The chain is simple: the agent talks to a local MCP server, and the MCP server talks to Resolve through Resolve’s scripting API. Claude Code / Codex → MCP server → DaVinci Resolve. The same community server works with Claude Desktop, Cursor, VS Code and a dozen other MCP clients.
This uses the community-built davinci-resolve-mcp server (by samuelgursky), which connects to DaVinci Resolve through Resolve’s own scripting API. It is not a Blackmagic product — don’t call it “Blackmagic’s official MCP server.” That matters for what works: the agent can only do what Resolve’s scripting surface and the server actually expose, and that varies by Resolve version.
What held up, on 1 October 2026
Checked against the davinci-resolve-mcp package on npm (v4.8.26) and its GitHub repository, and Blackmagic’s scripting behaviour.
Every prompt below follows the same shape: ask the agent to inspect and report first, make one small change second, and verify the result third — and never on the original. Duplicate the timeline, name the copy, work on the copy. An AI that can edit your Resolve project can also quietly break it, and this is the discipline that catches a mistake while it is still one marker, not one master export.
Install and enable
Install DaVinci Resolve Studio
On the same computer the agent runs on. Use Studio — external scripting is supported there directly. The project lists Resolve 18.5+ as supported; newer builds are preferable.
Install Python 3.10+
3.10 to 3.12 is the lowest-risk range. Verify it with the command below.
Install Node.js
The quick-start installer runs through npx, so install a current Node.js (20.9 or newer) and check both node and npm return a version.
Enable Resolve scripting
In Resolve Studio: Preferences → General → External scripting using → Local. Save, quit Resolve, reopen it, and keep it running while you test. Same-computer setups use Local.
Create a throwaway test project
Make an MCP_TEST_PROJECT, import a few disposable clips, and create Timeline 1. Never point the agent at a real client project first — this makes troubleshooting far safer.
Check your versions
Install the MCP server
One command installs it. The bootstrapper creates a managed installation, detects Resolve, and can configure your MCP client automatically — it may ask which client you are using, and can set up several at once.
Install
Then run the diagnostic. It checks the environment and the MCP configuration and helps pinpoint an install or Resolve-connection problem before you hand the agent anything real.
Diagnose
Connect your agent
Option A — Claude Code. If the installer detected it, it writes Claude Code’s MCP configuration into a project .mcp.json (you can also target it during a source install with the client id claude-code). Quit Claude Code, reopen it, open the directory that holds the configuration, and ask it to verify the connection with a read-only prompt.
First prompt — read-only
Option B — Codex. Codex uses ~/.codex/config.toml (or $CODEX_HOME/config.toml), with a TOML [mcp_servers.davinci-resolve] section, rather than the JSON many clients use. For a manual repository install the command is python install.py --clients codex, which adds that section while preserving your existing settings. Restart Codex after setup, then test.
First prompt — read-only
Test it safely
Don’t open with “edit my whole video.” Walk up from a read-only inspection to a single safe edit, then to work on duplicates. Each prompt below is one message to send the agent. If a step returns the right information, the connection underneath it works.
Inspect the selected clip
Add a marker
Edit a duplicate timeline
Never the original. Duplicate it, name the copy, and only touch the copy.
Natural-language silence cleanup
The agent lists what it would cut and waits. Silence detection needs ffmpeg on your PATH — if the agent says it can’t, that’s why.
Organise the Media Pool
A lower-third graphic
What’s controllable depends on your Resolve version and the exposed tool; some operations are easier than others.
Inspect the grade, then a new version
Don’t assume every Color-page control is reachable; it depends on Resolve’s scripting surface and the server.
Propose, add, then start the job
Never let the agent render over production files. Settings first, then the job, then start it — three separate messages.
Give it a standing rule set
Paste this once at the start of a session. It is what turns the setup from a demo into something you can trust on a real project.
What you can actually do with it
Once connected, the agent can help with media organisation, timeline inspection, cuts and trims, markers, timeline versions, subtitle and transcription workflows, supported colour operations, Fusion compositions, graphics and templates, project management, render setup and basic QC. The project describes support across editing, media, render, grading, Fusion, Fairlight and project management — but what actually works depends on the underlying Resolve API and your version. Treat the list as the ceiling, not a guarantee.
Troubleshooting
The whole thing, in twelve lines
Inspect → act → verify, and test on copies before you let an agent near a production project. Everything else here is plumbing; that habit is what makes handing your timeline to an AI a convenience instead of a gamble.