DaVinci Resolve 21.1 just levelled up in a big way with AI Assistance (via an MCP server), letting AI tools like Claude and Codex control the edit timeline. In this test, AI handled the cuts, titles and export, but improving the transcript made a noticeable difference to the finished edit.
Below is an AI-assisted summary of the key points and ideas covered in the video. For more detail, make sure to check out the full video above!
What DaVinci Resolve 21.1 AI Assistance actually does
DaVinci Resolve can now connect to an MCP server so an AI can drive real editing actions inside Resolve, not just give suggestions.
That means an AI can be prompted to:
- Create a new project and import footage
- Build an edit by choosing the best takes and trimming mistakes
- Add jump cuts and slow zooms (with consistent eye framing so cuts feel less jarring)
- Generate titles and graphics timed to the spoken content
- Duplicate timelines and make branded variants
- Export a finished file and verify settings (resolution, fps, audio, duration, file size)
This kind of “AI edits it for you” workflow existed in tools like ChatCut and Descript, but bringing it into a mainstream, pro-grade editor like Resolve is a massive step forward.
For a broader look at the editor’s built-in features, see our DaVinci Resolve AI tools guide.
What you need before this works
A couple of non-negotiables showed up in testing:
- DaVinci Resolve Studio (paid version) is required to access AI Assistance.
- DaVinci Resolve 21.1 is required.
- A pretty decent computer helps, because Resolve handles the editing and rendering on your computer.
- An AI tool connected via MCP, such as Claude or Codex.
- For the transcript workflow shown here: a Deepgram API transcript with word-level timestamps.
Setup: turn on AI Assistance in Resolve (MCP connection)
- Update to DaVinci Resolve Studio 21.1. Older versions won’t show the required AI Assistance option.
- In Resolve, go to File → Set Up AI Assistance. Resolve will walk through the connection process automatically.
- If Claude or Codex is already open, close and reopen it. That helps the AI “see” the new MCP connection.
- In the AI chat, confirm it can talk to Resolve. A simple prompt like “verify it’s working” is enough before running bigger tasks.
A prompt structure that gets solid edits (even if it’s basic)
The test used a simple talking-head recording with multiple takes, false starts, and notes spoken out loud (like “don’t use this one” or “actually do this again”). The prompt still guided the AI like a human editor would be guided.
Key instructions that helped:
- Use the last take for each section (even if the wording changes slightly)
- Keep pacing relatively quick, but allow a little breathing space when transitioning between sections
- Add jump cuts and zoom-ins
- Maintain eye framing consistency so zoom cuts don’t feel dodgy
One run using GPT-6 Astra (Medium) produced a base edit in about 15 minutes 30 seconds, cutting raw footage from 23:58 down to about 7:14.
The big flaw: Resolve’s built-in transcription can miss words (and mess up timing)
The edit looked great at a glance. Markers, “final take selected” notes, eye framing, and zoom moves were all there. But a repeat issue popped up:
- Some double-ups and false starts remained in the cut
- A few cuts left breaths/noise just before the line
- The underlying issue traced back to Resolve’s transcription skipping parts of speech in some scenarios, which lowers cut accuracy
This is exactly the kind of problem that makes AI editing feel “nearly there” instead of “hell yes, ship it.”
The fix: swap in a Deepgram transcript for word-level accuracy
In this test, Deepgram improved cut accuracy because the transcript isn’t just plain text. It can be a full JSON with:
- Gaps
- Word-by-word timestamps
- Timing data to help identify stumbles, repeats, and micro-pauses
Once the Deepgram transcript files were provided, the AI rebuilt a cleaner timeline in about 4 minutes 47 seconds, with noticeably better spacing and timing, without needing extra guidance beyond the original editing notes.
Deepgram cost details mentioned:
- Requires an API call
- Around $20 in credits covered “a lot of content” without needing a top-up
- Transcription for this type of file was estimated at $0.10–$0.15
After the base edit, the AI was prompted to add “key titles and graphics, animated where it would help.” No branding guidance was provided at first.
Results:
- Titles were animated in
- Text was timed to speech
- Key points were pulled directly from the transcript
- The style looked decent, even if it didn’t match any brand (because no brand rules were given)
Then the workflow levelled up: the AI duplicated the timeline and was instructed to match the branding of a provided webpage (fonts, colours, theme). It scanned the page and rebuilt the graphics in about 8 more minutes, producing a version that matched the page styling and even added a call to action at the end.
Tools installed on the system but intentionally not used for this test:
Export by prompt (plus verification checks)
The AI was told to export a 4K version to the desktop. It handled:
- Naming the file
- Rendering
- Verifying the export specs afterward
Verified output specs included:
- 4K
- 30 fps
- Stereo audio
- 7:06 duration
- 3.2 GB file size
What this test cost (estimates in USD)
These are estimates reported for this test, not a fixed price for every edit. Actual costs depend on the model, usage, subscription or API billing, and how many revisions you make.
For a full run using Deepgram from the start, plus one set of on-brand titles/graphics and export, the estimate came back as:
- Deepgram transcription: $0.10–$0.15
- Editing + review + generation work (standard API-equivalent): $13–$19 USD, with roughly 8–12 million tokens processed
- If using “fast mode rates,” budget roughly $30–$40 USD (not required here because speed was already strong)
Keep it repeatable: build systems and processes, not rigid rules
The real win isn’t a one-off edit. It’s building a simple process that fits how footage gets recorded.
To get more consistent results:
- Give the AI context (what the content is for, what matters, what to cut)
- Add flexible guidelines, not hard guardrails that box it in
- Record with a consistent structure so the AI can spot patterns
Get a cleaner edit, faster
AI Assistance in DaVinci Resolve 21.1 turns editing into a “say what you want” workflow, but transcript accuracy makes or breaks the outcome.
In this test, a Deepgram transcript improved timing, cut cleanup and pacing. Review the finished timeline yourself, especially around retakes and sentence boundaries. Let the AI handle titles, branding variants, and exports while staying focused on the message.
Done is better than perfect, especially when the AI can do the grunt work and check its own output.