AI video editors managed through Codex and Claude Code

The editing repeated the fate of the code: first there were agents who write in the editor, then those who move clips on the timeline. By the middle of 2026, there were four classes of instruments. They are constantly confused, and they solve incompatible problems.

## Four classes

The choice is dictated by one question: where should the project stay? Everything else is details.

## Class 1. Agent editors

There is no installation program at all. The agent reads the source folder and cuts the ffmpeg.

### video-use

The browser-use team did the same for video as before for the browser. Throw raw takes into the folder, run the agent there, write "edit these into a launch video" - you get edit/final.mp4.

What happens without your participation:

• words-parasites, false starts and dead pauses between takes are cut out;
• each segment is colored (warm movie, neutral contrast or its own ffmpeg-chain);
• at each junction, an audio file of 30 ms is placed so that it does not click;
• subtitles are burnt - by default, chunky two words capsom;
• animated overlays are generated via HyperFrames, Remotion, Manim or PIL, each in a separate parallel subagent;
• the result is checked at each gluing point before it is shown to you;
• the session memory lives in project.md, so after a week the installation continues from the stop.

The installation is one prompt in the agent: it clones the repository, picks up the ffmpeg, registers the skill and asks for the ElevenLabs key.

The minus of the genre as a whole: there is no timeline, there is nowhere to correct the frame with your hands.

### kinocut

The same, but with a focus on quality control: quality gates, Video Receipts, release checkpoints. 196 MCP tools and 167 CLI commands. Strength – repurposing: from one record collects a pack Shorts/Reels/TikTok with vertical frames, subtitles and review artifacts.

You need Python 3.11+ and FFmpeg in PATH. It works locally, without cloud or credit.

### oh-my-cassette

Plugin and local MCP server on top of the Cassette platform. Can beat-synced slicing, auto-selection of music, PiP, credits, lower-third.

A valuable detail of the process: the agent first gives the digest timeline and contact sheet for approval and only after approval of the plan launches the render. Rantim checks duration, black frames and sound levels. You don’t need a GPU, but you need a Cassette account.

### video-editing-killing

Six Bash scripts, trim → jump cut → speed → caption → overlay chain. Only FFmpeg plus local Whisper for subtitles. No cloud, no API keys.

If the task sounds like “to remove pauses and burn subtitles” – this is enough.

## Class 2: MCP Bridges to Professional NLEs

The agent manages your actual installation program. The project remains editable, you pick up the mouse at any time.

### DaVinci Resolve

The most mature bridge in the category is davinci-resolve-mcp: full coverage of the official API, 93.6% tested on live Resolve. Two modes to choose from: 36 grouped tools to avoid inflating the context, or 353 granular – one for each API method.

Separately, there is an 18-tool Node server that does what the API can’t do at all: read and edit .drp / .drt / .drx ** without running Resolve**, that is, it works in the cloud. There is also an offline catalog of grade: mixing skintones inside the camera and between cameras, white balance on a neutral card, match-to-reference, import ASC CDL, broadcast-legal QC. Grade recordings are calibrated by Studio panel screen units and verified by reverse reading.

Sets one command npx davinci-resolve-mcp setup - the installer itself configures Claude Code, Codex CLI, Cursor, VS Code, Windsurf, Zed, Continue, Cline, Roo Code, OpenCode and JetBrains.

A separate story is bypassing the free version restriction. Blackmagic only lets external scripts in Studio, but the Workspace ▸ Scripts menu is not locked. Therefore, the repository has a script bridge that runs in Resolve and re-exports the resolve object through a loopback lister with HMAC signatures and disposable nonce. Confirmed at free 21.0.3.7, Studio 19.1.3.7, Windows 11 and Fedora 43. The authors frankly caution that this is a documented path, not a circumvention of a license, but Blackmagic could shut it down.

unofficial-davinci-mcp should look up the security model: any altering operation first returns the plan and requires a re-calling from confirm=true.

### Adobe Premiere Pro

There is no official MCP from Adobe. The community is engaged in an arms race, with hundreds of instruments.

They're all MIT. The practical choice is between the first two.

hetpatel-11 explicitly states Codex as Agent Skill:

npx skills add hetpatel-11/Adobe_Premiere_Pro_MCP --skill premiere-pro-mcp

Both have a production bridge – a CEP panel inside Premiere (Window > Extensions > MCP Bridge), UXP is experimental. Node 20+ and Premiere 2020+ are needed, agent and Premiere are on the same car.

### Final Cut Pro

There are two fundamentally incompatible approaches, and you will have to choose consciously.

SpliceKit is not just an MCP, but a platform: Command Palette over Cmd+Shift+P on local Apple Intelligence, an MCP server and an open plugin framework (Objective-C/C++, Swift, Lua, Python). The tools work through the internal Objective-C FCP runtime: timeline reading and editing, retyming, color, effects, rendering. Plus text editor with transcription, audio mixer by roles, silence removal, scene and bit detection, support for BRAW and VP9/WebM, OpenTimelineIO.

Payment: SpliceKit patches Final Cut Pro with a GUI patcher from the release and uses private APIs. It's a conscious risk, not a small thing.

fcp-mcp-server is the opposite philosophy: no patches, private APIs, and accessibility scripting. Everything through FCPXML: timeline health analysis, search for flash frames and unplanned gaps, rough cut, clearing parasitic words, phrase-based cutting, diff, validation. Sandboxed Writers, defusedxml, 182 security tests. Timings are stored by precise rational fractions, including NTSC frequencies. Live Mode gives FCPXML to the official Apple Open Document event.

The restriction is stated honestly: you can import XML software, but you can not unload the current timeline from FCP. Export XML will have to be done with your hands.

### After Effects

All MIT. In JUNKDOGE-JOE, the CEP panel raises the endpoint to 127.0.0.1:11488/mcp, and the Claude, Codex and OpenCode channels are built right into the panel. aftr has a self-contained “create → render → visually check → fix” cycle and a simulator that allows you to work without installed AE.

### Blender

ahujasid/blender-mcp - ★26 507, MIT. Formally 3D, but the agent executes arbitrary Python inside Blender, which means it pulls out both the Video Sequence Editor and the render. The most popular MCP for creative software in general.

### CapCut and

Separate Universe: CapCut stores projects as editable drafts, which means they can be generated programmatically and opened in a desktop application.

Two warnings. jianying-editor-skill requires Windows and ши is not newer than 5.9, and autoexport is disrupted by any mouse movement; CapCut International and mobile versions are not supported. capcut-ai-editor rules the project in place without a backup - README directly asks to make a copy.

## Class 3. Editors for People and Agents

The most interesting wave: a real timeline for a person and an MCP port for an agent in one application.

### Pireel Studio

The tagline is “an open AI video editor designed for humans and AI agents.” There are three ways to work at any given moment: by hand on a canvas and a timeline, by chat or by an external agent like Codex and Claude Code.

npx skills add pireel/pireel-agent

Open edition (pnpm install && pnpm dev) focuses on local editing; built-in chat and connected services – in hosting.

### OpenChatCut

Local conversational editor with professional multitrack timeline. Embedded agent and external MCP use one set of tools, and the project can be transferred to another agent** - a rare and very practical feature.

Of the possibilities: transcription to individual words and text-based cuts, generation of images, video, speech, music and SFX, editable motion-graphics templates, WebGL effects and custom shaders, segmentation of people and face-safe zones (subtitles bypass the face), reframe by object. Export to MP4, audio, subtitles, FCPXML and full project data.

TypeScript 6, React 19, Remotion 4, Electron 43. Media and projects are stored locally.

### FireRed-OpenStoryline

Full cycle agent: finds and downloads clips, understands the content, writes the script and voiceover, selects music and fonts, synchronizes with the bit, cleans the reservations through ASR-based rough cut, transfers the style of the text by reference. The finished pipeline is saved as a reusable Skill for batch generation.

AI-transfers go through third-party services, are expensive and give unpredictable results, and open libraries of music and fonts are limited by licenses.

### The Research Front

CutClaw - agent editing of hours-long videos to music: deconstruction of material into structured captions, scheduling shots, synchronization with bits and track energy, content-aware autoframe. Style is defined by one text instruction. The interface is Streamlit, MCP for Claude Code is only planned, the last commit is April 2026.

VideoAgent is an academic framework: video understanding, content questions, beat-synced editing, commentary-video, cross-language adaptations. Neither CLI nor MCP is declared, it is a framework, not an agent tool.

## Class 4: Code-first engines

Here, the agent does not need a special bridge: video is a TypeScript file, and the agent can write code by definition. Formally, this is not “video editor management”, but in practice, the most reliable route for agent video.

### Remotion

The tagline changed to "Video tools for the agent era." Three modes are officially declared - agent, interactive drag-and-drop and software, and React code remains the source of truth and can be switched at any time.

The repository contains the entire .agents/skills/ directory with openai.yaml configs – the project is being developed by its own agents. There are страница Agent Skills and каталог промптов. Render – locally, on Lambda, in Vercel Sandbox or on the client. Start: npx create-video@latest.

About the license separately: Remotion has special conditions, and in some cases, companies are required to buy a commercial one. This is the only project in the review that can not simply be taken and implemented.

The ready-made skill for Claude is haidrrrry/claude-remotion-skill (★96, MIT): 10 mandatory motion rules, 17 components and, most importantly, the render → извлечь кадры ffmpeg → проверить → исправить → перерендерить cycle.

### Revideo

According to README, “A scene is a regular TypeScript, so Claude or Codex can generate it from a prompt.” Scene - generator-function:

export default makeScene2D('scramble', function* (view) {
  view.fill('#0d0d12');
  const cube = createRef<RubiksCube>();
  view.add(<RubiksCube ref={cube} size={620} />);
  yield* waitFor(0.5);
  yield* cube().scramble(18);
});

Core without dependencies, headless render via renderVideo(), parallelization by workmen, React player for previews on the same project, frame-by-frame synchronization of sound. Runs wherever there is a Node and headless browser.

Remotion is often more convenient for Revideo agents because of MIT.

### Diffusion Studio Core

Browser video engine on WebCodecs and Canvas 2D, "game engine optimized for video". Nonlinear editors are built on top of it. Declarative timeline compositions, layers, subtitles, silence removal, effects, transitions, keyframes, masks, audio ramps, hardware-accelerated render.

The authors themselves say when not to take them: server render and composition through HTML/React → Remotion; low-level encode and muxing → Mediabunny, on which Core is built. Free with a watermark; the key is a one-time purchase with local signature verification, without a licensed server.

## The bottom layer: what the agent actually cuts

Almost all the projects above end up causing the same thing. Often that layer is the solution.

OpenTimelineIO is a timeline exchange format between FCPXML, AAF, EDL, Resolve and Premiere with Python API and CLI. The best choice is when the installation decisions are made by an agent, and the final fine-tuning is done by a person in his program.

Thin ffmpeg wrappers for MCP: video-creator/ffmpeg-mcp (★145, 8 instruments, macOS only), misbahsy/video-audio-mcp (★83), egoist/ffmpeg-mcp (★119, unlicensed, abandoned from March 2025).

Catalogue of skills under Claude Code - wilwaldon/Claude-Code-Video-Toolkit: Remotion, Manim, recording via Playwright MCP, YouTube-clipping, ffmpeg-post-processing.

Not to be confused with shorts generators: MoneyPrinterTurbo (★118 754 – a record in the theme) and ShortGPT (★7 897) make videos “from the theme”: script, drains, TTS, subtitles, assembly. Manage your WebUI, not your agent.

## What to choose

## There is nothing in this category yet

A sober list of restrictions is more important than a list of features.

There are no official MCPs from vendors. Neither Adobe, Apple, nor Blackmagic have released their servers. Everything is based on CEP/UXP panels, ExtendScript, AppleScript and Scripting API – layers that the vendor breaks with any update.

Free editions locked. External scripting Resolve is a function of Studio only. The workaround exists and is tested, but lives until it is closed.

Reading is worse than writing. Remarkable: fcp-mcp-server can give FCPXML to Final Cut, but you can't download the current timeline software. The asymmetry of “write easy, read hard” is typical of the whole category.

Locality is required. Almost all bridges require an agent and NLE on the same machine - the panel listens to the loopback. A remote agent will not connect to your Premiere.

Disruptive operations without backup. capcut-ai-editor rules the project on site. Good tone is dry_run from unofficial-davinci-mcp or checkpoints from JUNKDOGE-JOE.

Verification is the main bottleneck. The agent can't see the picture. Projects that solve this in earnest – self-evaluation at the junctions in video-use, render → inspect frames → fix in claude-remotion-skill, visual verification in aftr, contact sheet in oh-my-cassette – give a noticeably more useful result than those that just pull the API.

## Outcome

If the project should remain in your installation program, take the MCP bridge. The most mature davinci-resolve-mcp: full API, two granularities, offline mode without running Resolve. In Premiere, the choice of four servers is a compromise between the power of SpliceKit and the security of the FCPXML approach.

If the project is not needed at all, but you need a ready-made file - video-use defined the genre, kinocut added quality control.

If you’re making videos from data or templates, forget about NLE. Remotion and Revideo turn editing into code, and the code agent writes better than moving clips.

The general rule: the tools that check their own result win - extract frames, count levels, look for black frames, show the contact sheet before the render. A blind-mounting agent produces a video that will still have to be reviewed in its entirety.