Follow
Fix it in post

Edit

Resolve 21.1 Puts AI Agents in Your Timeline. Here's Where to Let Them In

Resolve 21.1's MCP server lets Claude or Codex act on your project. A decision rule for what to delegate and a security checklist first.

Cover illustration generated with AI for Fix It In Post. Control panel based on a photo by TaurusEmerald, CC BY-SA 4.0, via Wikimedia Commons.

With DaVinci Resolve 21.1, released on 7 September 2026, Blackmagic shipped a native MCP server, so an outside AI assistant can now act inside your project instead of just telling you what to click. That is a real change in how an NLE works. It also means an AI model now has write access to a client's project, which deserves more caution than most of the coverage has shown.

Our position: point the agent at the jobs that are tedious, rule-based and easy to check, and keep it away from editorial decisions and from anything you cannot undo. Here is what the feature is, where it is useful today, and a checklist to work through before you connect it to client work.

What actually shipped

Newsshooter's release report lists the MCP server among more than 100 new tools and controls in 21.1. MCP, the Model Context Protocol, is an open standard for connecting AI assistants to applications. Resolve exposes its functions through MCP, and the assistant calls them on your behalf.

You set it up under File > Setup AI Assistants, which detects compatible assistants on the machine and configures the connection. YMCinema's walkthrough of the manual shows Resolve detecting Claude Desktop, Claude Code, Codex in ChatGPT, Google Antigravity and Grok. RedShark also names Claude, Claude Code and Codex. We have not checked Blackmagic's manual first-hand, and the list of detected assistants may change between point releases.

Two practical limits. The feature is Studio-only (Studio is $295 on Blackmagic's site), and 21.1 also moves advanced scripting to Studio and drops Python 2. If you have old Python 2 utility scripts, they will stop working when you upgrade, whether or not you use an agent.

What the agent can do today

The examples in the coverage fall into three groups.

  • Housekeeping: renaming clips sequentially, adding coloured markers and exporting them to CSV, comparing settings between projects.
  • Batch operations: applying a LUT to a specific node across every clip (an S-Log3 to Rec.709 transform is the usual demo), reframing a timeline to 9:16.
  • Deliverables: rendering the same timeline as ProRes 422 HQ, 4K H.265 and 1080p H.264, and reportedly checking a project against Netflix delivery requirements.

RedShark also mentions building highlight edits from long-form material and removing unwanted clips. That is a different kind of task, and the one we would be most careful with.

The decision rule: verifiable, reversible, rule-based

Before you hand a task to the agent, ask three questions.

  1. Can I check the result in under a minute? A render queue with three outputs, yes. A 90-second highlight cut, no: you would have to watch the source to know what was left out.
  2. Can I undo it cleanly? Markers and a new timeline, yes. Renamed media-pool clips on a shared project with linked bins, sometimes not. Treat renames and metadata changes as destructive.
  3. Is the task fully described by a rule? "Apply this LUT to node 2 on every A-camera clip" is a rule. "Pick the best moments" is a judgement.

Three yeses: let the agent do it. Any no: do it yourself, or have the agent make a new timeline or a CSV that you review before anything touches the main edit.

In practice, that sorts tasks like this:

TaskHand to the agent?
Render a deliverables set from a locked timelineYes
Export markers and notes to CSV for the clientYes
Apply a technical LUT across a camera groupYes, on a duplicated timeline or a versioned grade
Make 9:16 versions of a 16:9 editAs a first pass only; reframing is editorial
Rename media-pool clipsOnly with a written naming scheme and a project backup
Build a highlight cut from a 40-minute interviewAs a selects timeline to review, never as the edit
Delete "unwanted" clipsNo

Security checklist before connecting a client project

Blackmagic's own warning, as YMCinema reports it, is the important part: scripts an AI assistant creates get the same disk, network and system permissions as the assistant. In other words, the agent can do whatever your user account can do, not just what Resolve allows. Before connecting it:

  1. Check your NDA and the client's AI policy. Connecting a cloud assistant means project metadata, clip names, markers and possibly transcripts go to a third-party model. Many broadcast and brand contracts now restrict this. Get it cleared in writing.
  2. Back up the project database first. Export a .drp, or snapshot the database, before every agent session.
  3. Use a restricted machine or user account. Give the assistant access only to the project's media volume, not your home directory, your other clients' drives or your network shares.
  4. Keep approval prompts on. Most assistants can run tools without asking. Turn that off for Resolve so you see each operation before it runs.
  5. Work on duplicates. Ask the agent to make a new timeline or a new grade version and work there.
  6. Keep a log. Save the assistant's conversation with the project. If something changes overnight, you need to know who or what changed it.

What it doesn't change

We have not yet run this on a full client job, and we will report back when we do. Based on what has been documented, the 21.1 agent is a very capable assistant editor for repetitive work: renders, batch grading operations, naming and logging. It is not an editor. It can assemble selects quickly, but deciding which 90 seconds of a 40-minute interview tell the story still takes someone who has watched the 40 minutes.

The takeaway

Resolve 21.1's MCP server is useful now for checkable, reversible, rule-based work, and that alone can save an assistant editor hours on a deliverables day. Treat it as a script with your full user permissions, because that is what it is: back up first, restrict access and keep approvals on. Leave the cut to people.

Keep reading