A professional editor handed GPT-6 Astra a raw 75-minute recording and let it drive DaVinci Resolve 21.1 through a new MCP server with zero manual editing, and while the AI produced a usable rough cut, color grade, and final render, the whole process took about an hour, proving AI-driven editing already works end to end but isn't fast yet.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're a video editor curious whether an AI agent can now drive a real NLE like DaVinci Resolve end to end, not just suggest edits.
You're deciding whether to build an AI-editing workflow around DaVinci Resolve's new MCP server instead of screen-based computer-use tools.
You want a realistic sense of how long AI-driven editing actually takes, versus what polished demo reels show.
SKIP IF…
You're looking for a step-by-step setup guide for the Resolve MCP server. This is a live test, not a tutorial.
You only edit in Premiere or Final Cut and have no interest in DaVinci Resolve specifically.
TL;DR
The full version, fast.
A YouTuber with two decades of audio and video editing experience handed GPT-6 Astra his raw 75-minute recording and let it edit the video entirely inside DaVinci Resolve 21.1, using Blackmagic's newly released Resolve MCP server instead of screen-based computer control. Working through OpenAI's Codex app, the AI transcribed the footage, cut 75 minutes down to a 22-minute rough cut across 96 edited sections, added a lower third, color graded the picture, corrected loudness to YouTube's target, and rendered the final 4K file. The catch: the whole process took about an hour, with some single edits taking minutes for what a human could do with one keystroke. AI-driven editing already works end to end. It just isn't fast yet.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Cold-open hook: he watches the AI cut and audio-dip his timeline in real time and calls it possibly AGI.
00:25 – 03:07
02 · Why I wanted an agent in my NLE
He introduces himself as a two-decade audio/video editor, then explains what pulled him in: a GPT-6 Astra hype post, DaVinci Resolve 21.1's new AI assistant announcement, and a tip from OpenAI's Brent about the first-party Resolve MCP server.
03:07 – 04:35
03 · Hooking up the Resolve MCP
He installs DaVinci Resolve Studio 21.1 and asks Codex to find and connect to Resolve's native MCP server. The handshake succeeds without any manual clicking.
04:35 – 06:19
04 · Asking for a rough cut
With both his program and screen-share footage loaded, he asks GPT-6 Astra to build a rough cut using only the MCP link, explicitly ruling out computer-use control.
06:19 – 07:12
05 · Waiting on the transcription
The agent runs an in-app transcription of the 75-minute recording. It takes nearly six minutes, prompting a joke about needing monkey-grinding-away b-roll to fill the dead air.
07:12 – 08:48
06 · Watching it edit, hands off
Once transcribed, the agent starts making real cuts on the timeline while he watches without touching anything, visibly excited that it's actually happening.
08:48 – 11:24
07 · 75 minutes down to 22
The agent works in bursts, pausing to self-check between passes, and lands on a 22-minute rough cut, matching the length of his own previously published edit of the same footage.
11:24 – 13:16
08 · Half an hour, start to finish
The full rough cut, all 96 matched sections with linked audio and 10 clean screen-share cutaways, is done at the 30-minute real-time mark.
13:16 – 14:39
09 · Fixing one thing
He asks the agent to tighten a single long pause. The fix works, but takes nearly two minutes for a trim he could do by hand with Shift-Delete in about a second.
14:39 – 15:59
10 · Adding an overlay
With no style guide given, the agent designs and inserts a five-second lower third introducing him by name and channel, choosing its own fade and accent color.
15:59 – 16:44
11 · Grading and loudness
He hands off color grading and YouTube loudness correction as the final two technical passes before render.
16:44 – 19:34
12 · What nobody tells you about this
The grading and loudness pass runs long. His context window fills and has to auto-compact, and he sits through the kind of dead time that never makes it into AI demo reels.
19:34 – 20:46
13 · Rendering it
The agent confirms the edit is technically ready, builds its own named export preset, and sends the graded timeline to the render queue through the MCP link.
20:46 – 22:10
14 · The finished file
He reviews the rendered 4K file and calls several individual cuts good enough that he, an experienced editor, wouldn't even retrim them.
22:10 – 23:05
15 · The verdict
He frames the real goal as freedom from vendor lock-in: choosing his own AI agent and NLE rather than being stuck with one company's built-in assistant.
23:05 – 24:14
16 · What I actually want next
He closes on what's still missing (AI b-roll, overlays, full publish automation) and asks viewers whether this workflow is one they'd actually use.
Atomic Insights
Lines worth screenshotting.
Blackmagic shipped a native MCP server in DaVinci Resolve 21.1, letting AI agents control the editor directly instead of clicking around the screen with computer-use tools.
GPT-6 Astra, running through OpenAI's Codex app, cut a raw 75-minute recording down to a 22-minute rough cut across 96 edited sections without any manual guidance.
The AI transcribed the full recording inside Resolve itself, but that step alone took nearly six minutes, versus roughly 20 seconds using a dedicated service like AssemblyAI.
The entire rough cut, from opening the raw files to a finished 22-minute timeline, took about 30 minutes of real time.
Fixing a single 2-second pause in the timeline took the AI agent nearly two minutes, a trim a human editor can do with one keystroke in about a second.
The AI built a lower third overlay from one plain-language instruction with no style guide, choosing its own fade timing and accent color.
Color grading and loudness correction together took over 12 minutes of agent working time for a single video.
The finished render hit -14.68 LUFS integrated loudness with a -1.04 dB true peak, both correct for YouTube's target, without the human giving any numeric spec.
Every edit decision, including which of two identical camera angles to show, was made by the agent inspecting screenshots of the video, not by reading metadata alone.
The AI agent's context window filled up mid-task and had to auto-compact, adding several minutes of dead time with no editing happening at all.
Every AI-made edit inside the real NLE stayed fully undoable with a normal Command-Z, so the automated workflow never left the editor's own safety net.
The editor's verdict was that AI can already run the full editing pipeline end to end in a real NLE, but fine-grained trims and QA passes are still faster done by hand.
Takeaway
AI can already drive a real editor end to end, just not fast
WHAT TO LEARN
An AI agent connected directly to professional software through an open protocol can complete an entire edit unsupervised, but the demos that make it look instant hide an hour of waiting, context resets, and single trims that took longer than doing them by hand.
02Why I wanted an agent in my NLE
A tool doesn't need built-in AI to become AI-native. An open connection standard like MCP can let any general AI agent control it directly.
A single competitor's protocol announcement is often the real trigger to test a new AI capability, not the marketing copy around the model itself.
03Hooking up the Resolve MCP
A native, first-party connection between software and an AI agent works dramatically faster than screen-based 'computer use,' because it skips visually locating buttons and cursors.
The setup step for hooking an agent to new software can be nearly instant once a maker ships protocol support, with no custom integration coding required.
04Asking for a rough cut
The vaguer and more complex the first prompt to an AI agent, the more it front-loads unglamorous prep work, like re-running transcription, before touching the actual task.
Being explicit about method matters. Specifying 'use the direct link only, not on-screen control' ruled out an entire class of unwanted agent behavior.
05Waiting on the transcription
A general AI agent driving specialized software inherits that software's slow paths. An in-app transcription step ran roughly ten times slower than a purpose-built transcription service.
Not every part of an 'AI does everything' workflow is actually fast. Agent orchestration can be bottlenecked by the mundane task underneath it.
06Watching it edit, hands off
An agent visually inspecting screenshots of its own work, not just reading file metadata, is what lets it make context-aware editing decisions.
Genuine hands-off execution is possible today for a real, non-toy task inside professional software, not just scripted demos.
0775 minutes down to 22
An AI editing agent can independently match a human's own previously published edit length, without ever seeing that earlier edit.
Filtering dozens of usable cuts and cutaway choices out of raw footage is squarely the kind of decision-dense task agents can now do unsupervised.
08Half an hour, start to finish
Getting from zero to a full rough cut, including all the false starts and thinking time, is a ratio worth benchmarking any AI editing tool against.
A small self-QA habit, like catching and correcting a trim-handling discrepancy before applying it, is a good sign an agent is trustworthy with more autonomy.
09Fixing one thing
Fine-grained fixes are where current agents lose their speed advantage. A two-second trim can take minutes for an agent versus about a second by hand.
Every AI-made edit inside a real application staying fully undoable with a normal undo command is what makes handing over control low-risk.
The lesson isn't to avoid the agent for small fixes. It's to know which tier of task (broad rough cut vs. single precise trim) the agent is actually faster at.
10Adding an overlay
Given one plain-language instruction and no style guide, an agent will invent reasonable defaults rather than stall for more specification.
An agent's unguided creative choices reveal its baseline taste, which is worth seeing before handing it real brand work.
11Grading and loudness
A combined technical pass, like grading plus loudness correction, can be the single slowest stretch of an otherwise fast automated session.
An agent hitting an exact platform-standard numeric target without being given the spec means it already knows common technical benchmarks for that platform.
12What nobody tells you about this
Publicly shared AI-automation demos are almost always the highlight reel. The actual session usually includes multi-minute stalls and a lot of watching a progress bar.
An agent's context window filling up mid-task is a real operational cost of long automated sessions, not just a theoretical limit.
13Rendering it
An agent describing its plan back to you, and confirming readiness, before it commits to an irreversible action like a render is a cheap verification step worth requiring of any automation.
14The finished file
A working professional judging individual automated outputs as ones they wouldn't even redo is a sign the bar for 'good enough AI output' has already reached a real practitioner's standard.
Objective technical numbers, not just gut feel, are how you verify an automated pass is actually ready to ship.
15The verdict
The real capability question isn't whether AI can drive existing software (it now can), it's whether you can choose your own agent and protocol rather than being locked into one vendor's built-in assistant.
Wanting complete freedom over your tools is a legitimate technical requirement to test for, not just a preference. It determines whether you can swap AI agents later without rebuilding your workflow.
16What I actually want next
The unsolved piece in an automated pipeline is often the surrounding skills, like asset generation and style guides, that still require manual assembly around the automated core.
Where the real long-term value sits is in how you configure and prompt your agent, not in which specific piece of software you point it at.
Glossary
Terms worth knowing.
MCP (Model Context Protocol)
A standard that lets an AI agent connect directly to an application's internal controls, instead of clicking around the screen like a person would.
Computer use
An AI control method where the agent operates software by moving a cursor and clicking on screen the way a human would, rather than through a direct API connection.
LUFS
A loudness measurement standard platforms like YouTube use to set a target volume level so videos don't sound inconsistently loud or quiet.
Rough cut
A first-pass edit that lays out the basic structure and cuts of a video before finer polish like color grading and sound design.
Codex
OpenAI's coding-focused AI agent app, used here to run GPT-6 Astra and connect it to DaVinci Resolve through MCP.
the cold-open thesis stated in one line, immediately undercut by the rest of the video→ TikTok hook↗ Tweet quote
05:37
“controlling by cursor is pretty slow and janky, whereas MCP is like a first-party thing”
the technical crux of the whole video in one sentence→ IG reel cold open↗ Tweet quote
10:05
“I am not editing. I'm an editor of like decades and now AI is doing it for me.”
punchy identity-shift line with an obvious hook→ TikTok hook↗ Tweet quote
13:58
“two minutes to make a two-second trim is not acceptable”
the video's counterpoint to the hype, in one number→ newsletter pull-quote↗ Tweet quote
21:21
“until someone completely automates that, I'm not calling AI video editing done yet”
the closing verdict line→ newsletter pull-quote↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogy
now look at this this is happening in real time hands off this it's really impressive this is the best bit of the video oh my goodness me i am not editing i'm an editor of like decades and now ai is doing it for me it's even look at this this is insane it's actually gone ahead and put audio dipping on the audio track look at that to make the edit nice and silky smooth maybe we've hit agi here guys Okay, here's the deal.
I'm Mike here and I'm wearing headphones today. My background is in audio editing and video production. So when GPT Astra came on the scene and basically said it's AGI for video editing, I got really interested.
Then I looked at posts like this post from Higgsfield saying that everything is being edited on an infinite autonomous canvas with GPT -6 Astra. Then I went along and I actually noticed that Blackmagic just announced DaVinci Resolve 21 .1, introducing, look at this, AI assistant integration. That really got my attention.
And then I got a message from Brent at OpenAI who says this may make things easier. He's pointed to the first party MCP server that just landed in DaVinci Resolve. So yes, I'm going to install DaVinci Resolve Studio 21 .1 and test out the MCP server and see if it truly is the Linux of video editing that I'm looking for.
Let's dive in. Quick caveat here, because yeah, Linux of video editing, well, that isn't DaVinci Resolve, because DaVinci Resolve is not free and open source. Diffusion HQ, actually, Diffusion Studio, that's what I've been using up until now to make most of my edits with AI agents.
But hey, maybe Blackmagic have made it super easy. All right, let's fire up DaVinci Resolve. 21 .1.
What's new? Look at this. Look at this.
AI assistant control. This is so exciting. Ask Claude or ChatGPT to organize media, create edits, and batch render.
This is what interests me here, and this is what we'll be testing out in this video. Right, so I'm going to create a brand new project, and I'm going to call this one OpenClaude2MCP, okay? And this will be one of the recent videos I actually published on my channel, started from scratch right here.
All right, that said, I've just dragged the clips in, changed the frame rate, and we're ready to go. Now, I'm not going to do anything. I'm going to see what my AI agent can do from scratch.
All right, so we've got Codex up here, and we're in the video editing project. It's a project I've created specifically for this purpose. Now, one thing I want to make absolutely clear is that the videos I've just dragged in are a raw recording of my face and my screen share for 75 minutes.
All the silences... All the mistakes, all the problems. If Astra from OpenAI truly is a GI for video editing, I want to experience it without any guidance from me.
Just going to open the file and show you. It's me clapping to sync there. 75 minutes of me going through recording and also looking at things like confused like that because things go wrong.
It's all there. All the mistakes and everything. So first of all, I'm just going to ask GPT -6 Astra.
Okay, I've got DaVinci Resolve running 21 .1. They say that it can integrate with AI agents via MCP. Are you able to find that functionality and hook it up so that we can control DaVinci Resolve, not with the cursor, but with the MCP server?
So I'm going to get Astra on the case with that and see if it can hook this up itself. Controlling a video editor by the cursor is not what I want to do. And the reason for it is that controlling by cursor is...
Pretty slow and janky, whereas MCP is like a first -party thing. So it's checking the support right now in my local setup and connecting. Okay, it's found it.
Your Resolve 21 .1 installation includes Blackmagic's native Resolve MCP. This is awesome. Checking the live connection and registering with Codex.
OK, you see, it's as easy as that. I've got the ChatGPT app or the Codex app, if you like, installed on my Mac. I asked ChatGPT 6 Astra to do the connection for me and it's done it.
So everything is connected up to Codex. This is awesome. Let's give it a go.
And I've just moved it over to the right so that my face is not blocking chat GPT. Connected, tested, okay, excellent. Without cursor control, it's all registered.
This is fantastic. To load it into Codex's tool menu, go to Settings, MCP, Servers, and Restart. Okay, so I'll do that.
Okay, so can you see my open DaVinci Resolve session? Are you able to insert both the program and the screen share onto separate video tracks and perform a rough edit for me using only the DaVinci Resolve MCP, please? Now, that's quite a complex first initial prompt.
You'll see I'm using GPT -6 Astra High, and it says I'll inspect the open project. So hopefully it's going to see this OpenClaw 2 MCP project and edit through Resolves MCP. I'm very specific here.
I don't want it to take over the computer. I don't want computer use. I want this all done via the MCP link to see how good it is.
Okay, hands off at the moment. It's thinking things through. It's made a change to a file in the project.
We'll see what comes back. Now it seems to be running a lot of scripting here. A lot of Python is getting created.
Resolve MCP client Python. I can access the session in both recordings. 75 minutes long.
This is correct. 4K 60 frames per second. This is good.
Screen recording at 5K. Yes, excellent. Preserving sync.
Oh, wow. Okay. Transcription initializing.
I did not do that. So clearly the MCP is doing something. Resolve has kicked into action over here.
This is great. And yes, so that was the first step, getting the transcription, what any good AI agent would do before editing. We're in a good place.
Now, this is going to take some time, but the best thing is it's included in my DaVinci Resolve Studio subscription. So all of this is done at nearly 10 times speed for free. And we see while this is going on, ChatGPT is still thinking here.
Resolve is returning the media correctly, but transcription timed out. Oh, dear. This is a first error.
It didn't time out. It's still going. But ChatGPT over here is getting impatient.
It's really interesting because the workflow that I've built that I use now to agentically edit videos uses assembly AI transcription, and that will transcribe 75 minutes in like... 20 seconds. So it's kind of painful sitting here watching Resolve take nearly six minutes to transcribe the same video.
And at this point, B -roll of a monkey grinding away, typing on a computer, transcribing a video would look absolutely epic on screen. Oh, nearly. Yes.
OK, we're transcribed. This is good. Let's go back over here to ChatGPT.
Hey, the transcription just finished. So, oh, wow. OK, hands off.
This is happening now. This is my program, my screen share, my program audio all here in the. Yes, this is this is kind of what I want to see.
This is amazing. OK, it hasn't actually done anything yet, but it definitely switched the scene here. Now I want it to edit.
Okay, what have we got here? The two video tracks are now in place. Transcript shows retakes and long waits so I can make useful content -based cuts.
Resolve couldn't confirm waveform alignment. Both files are the same recording, which I'm using. Wow, okay, things are happening.
Things are actually happening now. I'm not touching Resolve. This is interesting.
And I can see it's viewed an image. Yes, it's actually watching screenshots of my video as it edits, which is a great sign. come on GPT -6 Astra make your first edit wow okay so it's skipping around the video Look at this.
Here is my speech. You can see long gaps as I pause to change scenes or do something different. You'll see there's a lot of retakes here as well.
I'm hoping that GPT -6 Astra can see all of this and control it through the Resolve MCP. It says, I'm keeping your existing program composition as the main picture, clean screen share cutaways as well to help with readability, removing failed takes, production asides and long waits, and adding markers for the two claw cast inserts that need your narration.
All right, this is great. It has fully picked up context of exactly what I'm trying to do in the video. This is a great start.
But just for your information, the video you're watching is obviously edited, but I've been sitting here for 20 minutes and I still haven't seen a single splice in my timeline. This is not fast. Okay, what's happening here?
Okay, everything is gone, apart from one single clip here. That's interesting. What are you doing?
It looks like it's building a timeline plan. And we can actually see that over here. It's made this new timeline called rough cut.
All right, comes to 22 minutes down from 75 minutes. This is pretty cool. I caught a one frame difference in Resolve's trim handling and I'm correcting it before applying cuts across tracks.
Okay, now what is really cool is I already completed this video and published it on my channel. And when I edited it with my current agentic process, it made a video that's about 22 minutes in duration. So this is a really good sign.
Okay, what's happening here? It's blotted out the screen share. It's focusing on my face.
Now look at this. This is happening in real time, hands off. This is really impressive.
This is the best bit of the video. Oh my goodness me. Yes, maybe we've hit AGI here, guys.
So it did a load of stuff. It stopped and now it's continuing again. This is insane.
This is stuff that would take me hours post creating a video. It's literally being done in minutes here. Now what I can see is it's doing a few minutes at a time.
So it did the first three minutes. It paused. Then it did another three minutes.
It's pausing. So it's like maybe it's QAing. I can't tell if it's doing quality assurance right now.
It's running a load of Python files that seem to be controlling the Resolve MCP client, which is pretty awesome. I had absolutely no idea that this was possible. Now, I'd love to zoom in and see what's going on here.
It does seem to be cutting my video quite well, which is good. Wow, more edits are happening here in front of my fairy eyes. I am not editing.
I'm an editor of like decades and now AI is doing it for me. The thing that used to be the most painful thing for any content creator is getting solved by AI now. And that is huge.
And we can actually see now the whole timeline is here in place. We're at 21 minutes and 59 seconds. So that's pretty good.
It's actually placing some markers on now all the way through. Look at these markers that are placed on. This is really interesting.
I do want to know what kind of markers it's placing on the timeline. Rough cut review. That's the intro.
Very nice indeed. We've got another one. Chapter marker, managed hosting and plans.
Okay, so it's like chapter markering here. Choose an AI model. So it's basically placing chapters in blue.
Then here, orange. What have we got? Original source includes setup token.
Wow, it's actually telling me here about setup tokens that I should probably take off camera. I love this. Now look at this.
The 2159 rough cut is saved. All 96 sections have matching source times and linked video and audio. 10 clean screen share cutaways as well.
So it's making decisions on when to use my screen share for readability. This is really, really cool. It's QAing again, viewing more images like so and even more images.
I'm pretty impressed. Okay, I can tell you this has taken exactly half an hour, so that's the real time, even though you're watching the edited final polished cut of me here on YouTube. 30 minutes to get to the rough cut where GPT -6 Astra says it's done.
The full length source timeline is preserved, markers, flag, missing claw cast inserts, and everything else needed. Okay, the moment of truth. Let's play it back and check those edits.
Let's dive in. Okay, I'm going to use this open claw a little bit slow, but do you know what? I would accept that edit.
That would pass. I wouldn't need to retrim that. It's even.
Look at this. This is insane. It's actually gone ahead and put audio dipping on the audio track.
Look at that to make the edit nice and silky smooth. I'm not even going to touch that. I could as an OCD video editor, but I won't.
Let's check the next edit here the first time. Well, if you're not, that is perfect. That is 100 % perfect.
All right, let's look at this edit now since on Hostinger. One click deploy is yes. Oh my goodness, guys.
We're at AGI for video editing and audio editing, may I say. As someone who's taught audio editing for like nearly two decades on YouTube, that's done. If you want to see my other channel, that's an archive of how we humans used to do that stuff.
Couple more edits to see where we're at. What you want. either way when you choose a plan nice so it chose to use my screen share there for readability and then cut back to me with picture in picture at this point now this is interesting as i look along this timeline there's quite a big pause here so i'm concerned by this kvm2 here you can type in my okay that is a little bit of silence there that would be a little bit too long so that's at this particular time stamp here so i might go back to chat gpt 6 astra and see if it can make that one edit for me I'll just say there is a long silence at 1 .20 .11.
Please fix only this. And we'll see if GPT -6 Astra can make just that one little edit there and tighten this bit up. It's actually inspecting right now, keeping tracks aligned.
Of course, that's what we want. Okay, you can see there's a two -second pause between you and typing my coupon code. So it's going to shorten that now.
Again, hands off. And there we go. It's done, I think.
Fantastic. Let's see if we can play back that section and listen again. You can type in my...
Yeah, that's acceptable. That's okay. Now, full disclosure there, that took nearly two minutes for Codex to do using the DaVinci Resolve MCP.
Two minutes to make a two -second trim is not acceptable. It did remove 1 .67 seconds and kept all other tracks aligned, which is great. That's good, but it's not perfect.
If I hit Command -Z here, look, we get the trims back. Everything can be undone with the Command -Z. Now, of course, what took two minutes here can easily be done by doing Shift -Delete.
That took like one second. So in some cases, yes, getting the rough cut and doing the high -level edits, brilliant, making fine -grained changes in QA kind of things that human in the loop would catch. You probably still want to be doing that as a human, unless you've got loads of coffee and you just want to go off and drink it while the agent makes its small little edits.
So this is fun, and it's a great thing that can save me a lot of time, but can it do the polish? Okay, now you've done the rough cut, I'm happy with it. I wonder if you can insert an overlay at the start of the video introducing me as Mike Russell from Creator Magic.
Now, I'm just doing one simple thing, and it's got no house guide on how I do my overlays, so it's going to make it all up. But I'm wondering if, look, it says it can do this with Resolves MCP, if it could do this. Obviously, if I was making a full video, I'd put loads more overlays and b -rolls and cutaways, but this is just one test.
Let's roll back to the start and zoom in here and watch it happen in real time. Okay, it says I'm making a five second lower third with soft fade white name text and small teal accent. It will sit on an overlay track.
Nice. Okay, look at that. It made it.
It has taken nearly three minutes of working here, but it did make it a nice lower third here that I assume it's going to insert on the timeline. Okay, there we go. Yes, it's done.
Look at that. Wow, that's amazing. Let's rewind and play back.
open claw 2 .0 is here with 16 000 pull requests yeah about half of everything in the project okay nice fade in fade out for the lower third that is done it's just wrapping up it's doing qa here again looking at the image to check and it says yes i've added the five second lower third mike russell creator magic till accent obviously no style guide here i did not give it any style guide but the fact it's gone ahead and done that is pretty insane So I'm pretty happy that pretty much anything related to video editing can now be done in DaVinci Resolve.
Definitely we can get the rough cut. We can get lower thirds and overlays. I'm sure we could probably get B -roll in there as well.
Okay, I'm really happy with this rough cut. It's basically ready to go to YouTube. There are only a couple more things I'd like you to do.
First of all, I'd like you to color grade the video like a boss. You're the world -class expert in it, so do it for me. And then I'd like you to make sure the loudness is absolutely correct for upload to YouTube.
Ensure those two things, and then tell me if it's ready to render. And if you're able to render it yourself, let me know. Don't render yet, but just tell me if you can.
Now, this is the one thing that no one tells you when they're playing with video editing. First of all, video editing takes a lot of tokens. My context is now compacting.
And the other thing is, well, all those slick demos you see on X and other YouTubers doing that are edited tightly. And full disclaimer, I'm editing this video with AI, by the way. They show you that it happens in seconds or minutes, which it really doesn't.
I've been sitting here the best part of an hour running this DaVinci Resolve MCP via Codex using GPT -6 Astra. Can it do all this stuff? Yes.
Can it do it fast? It's getting better all the time, let's say. But, you know, right now I've been sitting here for three minutes watching context automatically compacting.
OK, after four minutes, we've done the compaction. Now let's get into action. Okay, it's already working like a pro, keeping the grade restrained so skin looks better without distorting software interface.
I love your thinking, GPT -6 Astra. Let's see what it does. For that moment, you'll see my face and the background pretty washed out.
I'm hoping that it makes some changes. It's also checking Resolve's YouTube loudness controls against the whole program. Okay, what I notice is a lot of Python files getting created here, so it doesn't seem fully like it's doing everything via the MCP.
A lot of stuff is codecs writing code and analyzing stuff before getting back to Resolve via the MCP. We've got images here of me. That's QA.
okay again through the mcp client making some changes here and we can see yes render settings are being locked in over here on the left with a preset it has created creator magic youtube 4k review now i didn't create that preset it seems that codex is doing that for me okay it's backed up the approved edit it's lifting the tone slightly to stop my excess orange skin and keep the purple studio lighting very nice audio analysis is running across the program cut some fades no video render started yet.
Okay, audio is already in good shape. We're at 14 .69 LUFS with a 1 .04 dB true peak, which I can confirm is absolutely correct. So this is really, really good.
Just so you know, on color grade and audio, it's been working over 12 minutes now, and it's still not quite done. Okay, look at this. The grade is applied.
Opening title looks clean. Checking the saved export settings now. Okay, this is looking really good.
One thing I will say is I don't see much change in terms of colour grade here. It looks pretty much like what we started with, but hopefully the audio will be better. And so you know this last step is chewing time.
15 minutes and we're still not done. Oh, here we go. We're done.
That's it. Just as I said that, the grade is applied and saved. Brighter, more natural skin tones.
I can't see much difference, but maybe it did make a difference. Audio is correct. Technically ready to render.
Obviously missing the inserts, which is fine. I can render it through Resolves MCP. So let's try this.
Render it. And this is the final step now. The final piece in the pudding.
And I can tell you as I type that command in, we're bang on an hour. So let's see. I'll render the graded timeline using the prepared 4K YouTube preset, then check the exported file.
Hopefully rendering should be pretty quick, and then we can guarantee we can work end -to -end in DaVinci Resolve using only an AI agent. And there we go. Yeah, it's going into the render and it has indeed popped something into the render queue here.
This is amazing. This is actually happening in real time right now. It's rendering my video.
So that guarantees the whole process can be automated end to end using your coding agent of choice. In my case, GPT -6 Astra on high mode, along with the DaVinci Resolve MCP that was just released today in version 21 .1. This is pretty insane.
The 4K render is running and I'm saving it as OpenClaw 2 .0 Creator Magic MP4 in your video editing folder. Okay, this is very nice stuff. And we are definitely pretty much there with video editing.
Yes, it can be done and automated by AI. Is working in DaVinci Resolve great? Well, yes, because I don't need to know the software.
I don't need to touch a button. This can all be automated by AI. Now, I'm an Adobe user and a Premiere user for years and years.
And yes, Adobe, I would like to see you do the same thing, not have a proprietary lockdown AI assistant that lives in and is controlled. by Adobe Inside Premiere, I want to be able to choose Codex or my Claude subscription to edit my videos. I want to bring in skills to make lower thirds and overlays.
I want to bring in other skills to make AI B -roll. And until someone completely automates that, I'm not calling AI video editing done yet. Now, let me explain to you a little bit about my video editing process.
I am looking for the Linux of video editing. I don't want to be locked in or locked down. I want complete freedom and sovereignty over the kind of edits that I do.
And up until a long time and still to this day, even the video right now that you're watching, it's edited in a Buzz channel. And I'm automating that using Claude Code on Fable 5 .1. It does the edits.
And then previously, I was eyeballing. them by pulling a timeline into DaVinci Resolve. I would love to do the same thing in Premiere, but it doesn't support timeline imports that my agent can make, which is a bit of a bummer.
And for the eyeball, I'm not even sure if I need to use DaVinci Resolve or Premiere. Maybe I can just use another open app that's Lots of them out there.
Chat cut is a big one, and it's gaining in popularity. I've been messing around with Diffusion Studio, which seems to cut out the bloat that Resolve and Premiere suffer with. So the jury is still out on how this process will look.
But in my opinion, the absolute meat, the brains, lie in how you automate this with your AI agent, the skills you create, and how you tell your agent to edit. That's where the moat is. It's not about doing the edit or what software you use.
to QA or eyeball that's basically solved and we could just use a simple video editor in the browser to check that stuff out we don't need these big editing programs anymore but Blackmagic and DaVinci Resolve have certainly laid down the gauntlet here by making a Resolve MCP available I have been recording this raw video for about an hour I'm about to hand this ProRes that's recorded on my Mac studio off to my AI agents that will edit and publish the video that you're watching right right now let me know your thoughts has this been solved in your opinion would you use this workflow what do you want to see what are the missing pieces what did I not mention that I should have let me know in the comments down below I'm thoroughly engaged in this topic right now I want to see this succeed and I want video editing overlays AI b -roll all of that I want it solved for creators so that literally my job is to come up with the ideas in collaboration with AI hit record in my studio stop the recording and that's where I finish as a content creator when I hit stop on the recording button my AI agents in an ideal world should take over
publish and have it ready for you to watch on YouTube. Hopefully that's coming soon. Thank you so much for watching.
And YouTube is showing a video on your screen right now that's been determined by the algorithm to be absolutely perfect for you. So you should probably pay attention and you should watch it next. Thanks.
The Hook
The bait, then the rug-pull.
The video opens with a promise of AGI: a professional editor sits back while an AI agent cuts and mixes his timeline in real time. What follows is an hour of testing that promise against a raw, unedited recording, and the truth turns out slower, glitchier, and more interesting than the cold open lets on.
Frameworks
Named ideas worth stealing.
01:03concept
The Linux of video editing
His shorthand for the open, non-proprietary editing stack he actually wants: any AI agent (Codex, Claude, etc.) plugged into any NLE through an open protocol, instead of one vendor locking its AI assistant inside its own software.
Steal forEvaluating any new 'AI-native' creative tool by whether it's open (any agent, any protocol) or closed (single vendor-locked assistant)
CTA Breakdown
How they asked for the click.
VERBAL ASK
00:00link
“Join the Creator Magic community: https://mrc.fm/cmc”
Not spoken or shown on screen. The only CTA lives in the video description, alongside the chapter list.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A same-prompt, parallel test of the new #1 leaderboard model against Claude Fable 5 across four real build tasks — a voxel game, a motion-graphics video, a synth pad, and a researched website.
One editor runs nine real tasks through DaVinci Resolve's new native AI integration to find out where it saves real time and where it's just slower than doing it by hand.
Team 2 Films runs Claude through four real edit jobs, project setup, a rough assembly cut, TIFF-cached rendering, and color grading, to see what Resolve's new AI agent integration can and can't do.
How to connect ChatGPT to free DaVinci Resolve with an open-source MCP server so it can transcribe, cut, and caption a highlight reel, a feature Resolve normally locks behind its $300 Studio license.
A video editor wires ChatGPT's Codex into DaVinci Resolve through an MCP server, then lets it build a full first-pass cut on its own, mistakes included.
A content director runs 336 unsorted vlog clips through a Claude Code + DaVinci Resolve Studio pipeline that classifies A-roll from B-roll, proposes cutaway placements against four editorial rules, and drops the picks onto a real timeline — then shows exactly where it still needs a human.
00:23
00:40
01:13
01:38
02:00
02:16
02:35
02:52
03:02
03:29
03:55
04:15
04:32
05:18
05:36
05:48
06:17
06:35
07:10
07:25
08:01
08:13
08:30
08:57
09:05
09:23
10:11
10:19
10:45
10:58
11:21
11:35
12:15
12:40
12:52
13:09
13:37
13:48
14:17
14:37
15:06
15:36
16:06
16:30
16:53
17:06
17:43
18:01
18:19
18:37
18:58
19:06
19:29
19:47
20:06
20:18
21:03
21:39
21:57
22:15
22:54
23:10
23:28
24:05
15:12
17:33
18:59
13:44
17:02
15:15