An AI agent edited this video on my MacBook: AutoCut and FrameAtlas are now open source
The video below was not edited by me. A Claude Code agent found the clips, wrote and spoke the Malayalam narration in my cloned voice, composed the music, rendered the title animation and exported the film, with almost all of the heavy media work running on my MacBook Pro. Today I am opening up the two tools that made that possible: AutoCut, a video editor built for agents, and FrameAtlas, a local catalogue that finds footage by what is in it.
The video is in Malayalam. Turn on captions (CC) for English subtitles.
The code
- AutoCut: github.com/moheshmohan/AutoCut. A fork of FilmCraft, the open-source, pure-Rust non-linear editor from the ArtCraft team, with a local AI pipeline added on top. AutoCut is an independent fork and is not an ArtCraft product.
- FrameAtlas: github.com/moheshmohan/FrameAtlas. A local catalogue and search for footage that lives in Google Photos, on Insta360 cameras and in local folders.
Why build this at all?
It started with a small, practical problem. Cloud agents are great at planning an edit, but every frame they look at and every file they read costs tokens from a weekly allowance. Most of the work in video editing is not thinking. It is watching footage, transcribing speech, generating a voice, rendering and checking. So the idea was simple: let the cloud agent plan and decide, and let the Mac do the media work with local models.
That idea grew step by step:
- An editor an agent can drive. FilmCraft already exposes every menu item, button and slider as a command with typed parameters, over a CLI, a JSON control channel and an MCP server. AutoCut builds on that.
- Local models for the heavy lifting. Scene detection, frame descriptions, speech recognition, voice generation, music, privacy blur and quality checks all run on the Mac.
- A way to find the footage. My clips are spread across Google Photos, an Insta360 cloud and local folders. FrameAtlas catalogues them by what is in each scene and when, so an agent can ask for "the yellow robot dog" and get back a file and a time range.
What it can do
- Prompt-driven editing:
filmcraft-cli autocut "remove the silences and add captions". Every edit lands in the undo history like a person's. - One-command pipelines: a quick Short from a folder, a beat-cut reel directed by a local model, a narrated showcase from a Google Photos album, Shorts cut from a long talk, a clean rough cut of talking takes.
- Google Photos in and out: pick clips (a local vision agent can drive the picker for you), and upload finished exports back into an album.
- 360° video: import Insta360 recordings, stabilize them, reframe with keyframed camera moves, tiny planets, whip pans and several angles at once.
- Voice: narration in a cloned voice (with your consent), checked by listening back with a speech model; voice isolation that pulls a voice out of a noisy hall.
- Music and sound: music generated from a description, sound effects, ducking under the voice, loudness for each platform.
- Generation with your own local models: pictures, short clips, gap filling, extending a shot, restyling a clip, adding an object to a real shot. Every generated shot can be labelled on screen.
- Animation: titles and explainer scenes rendered locally with Render Anything (RAEA).
- Privacy and delivery: tracked blur for faces and number plates, and a pre-flight check for YouTube, Shorts and Reels before anything is published.
Does it actually save anything? The measured run
To check, I started a fresh Claude Code session (Opus 5.5), gave it one short brief, "make a 60 to 80 second film called Machines in Motion from my footage, in my voice, with local tools doing the media work", and measured everything from the session's own transcript. The same brief had been given once a week earlier, which gives a comparison:
| 12 November | 20 November | Change | |
|---|---|---|---|
| Output tokens | 176,495 | 86,165 | ‚àí51% |
| All tokens, cache reads included | 40.7 million | 27.9 million | ‚àí31% |
| Model calls | 167 | 159 | |
| Correction rounds | 2 | 0 |
The 20 November run took 57 minutes 38 seconds of working time. For 39 minutes 12 seconds of that, local tools on the Mac were doing the work (finding clips, speaking the narration, composing music, upscaling an old clip, blurring faces, exporting and listening back) and no cloud tokens were being spent.
To be honest about what this does and does not show: it compares two runs of the same brief a week apart, as the tools improved. I did not run a cloud-only baseline, where the cloud model does the media work itself, so I am not claiming a fixed percentage saving for every job. Local work still costs hardware, memory, time and power, and the output of every model still needs checking.
Running it by hand
What you need
- A Mac with Apple silicon. Everything here was tested on a MacBook Pro M3 Max with 36 GB of memory. The editor itself also builds on Windows and Linux, but the local AI pipeline was only used on macOS.
- Rust 1.95 or newer, and Git.
- Optional, for the AI features: Ollama; Python 3.9+ for FrameAtlas; RAEA for animations; Insta360 Studio for 360° recordings in the Insta360 cloud.
Build and open the editor
git clone https://github.com/moheshmohan/AutoCut
cd AutoCut
cargo run --release -p filmcraft # the desktop app, with a demo project
cargo build --release -p filmcraft-cli # the command-line tool: target/release/filmcraft-cli
The first release build takes a while. After that you have a full editor: import media, cut on the timeline, grade, mix and export, with the panels and shortcuts you know from Premiere Pro.
Local models (optional)
ollama pull qwen2.5vl:7b # looks at clips, drives the Google Photos picker, privacy checks
ollama pull qwen3:8b # writes scripts, including Malayalam
ollama pull phi4-mini # classifies plain-language edit requests
Other engine models (speech recognition, voice, music, upscaling, the voice isolation model) are downloaded by the app into its data folder the first time you ask for them, and only after you agree. Check what is installed with commands such as filmcraft-cli exec transcript.models and filmcraft-cli exec music.neuralModels.
Edit with plain language
CLI=target/release/filmcraft-cli
$CLI --project edit.fcproj --save autocut "cut the first 3 seconds"
$CLI --project edit.fcproj --save autocut "remove the silences and add captions"
$CLI quick ~/Movies/expo --topic "Robots at the expo" --seconds 30 --export expo.mp4 --save-as expo.fcproj
$CLI direct ~/Movies/trip --brief "a 30 s reel of our trip" --aspect 9:16 --seconds 30 --export trip.mp4 --save-as trip.fcproj
$CLI --project expo.fcproj redact --kinds faces,plates --dry-run
$CLI publish check youtube expo.mp4
Each pipeline also writes a report, a frame sheet and a QA file next to the export. Read them before you share anything.
Connect your media (optional)
- Google Photos: you need your own Google Cloud OAuth client. The steps are in crates/gphotos/README.md. Then
filmcraft-cli photos authonce, andfilmcraft-cli photos pickto choose clips. - Insta360:
filmcraft-cli insta360 listandinsta360 importbring recordings in from Insta360 Studio's download folder; see docs/360-video.md. - Your voice: record a reference and make a voice from it as described in docs/voice-recording.md. Only clone a voice with its owner's consent.
- Generation tools: AutoCut bundles no generation model. You list the tools you installed in one file, as described in docs/generation.md. Read each model's licence before you publish what it makes.
FrameAtlas
git clone https://github.com/moheshmohan/FrameAtlas # next to the AutoCut folder
cd FrameAtlas
./run.sh serve # then open http://127.0.0.1:7318
FrameAtlas has no pip or npm dependencies and keeps its catalogue on your Mac (~/Library/Application Support/FrameAtlas/). It finds the AutoCut CLI on its own if the two folders sit side by side, or through FILMCRAFT_CLI. Once your media is catalogued, AutoCut can search it directly:
filmcraft-cli catalog search '{"query": "yellow robot dog walking", "limit": 8}'
Running it with an agent
This is the part the video is about. AutoCut is built so that an agent edits the same way a person does, through the same commands, and every result stays an editable project.
Claude Code
- Build the CLI (
cargo build --release -p filmcraft-cli) and, if you want to watch the agent work, open the app with its control channel:cargo run --release -p filmcraft -- --control 9876. - Open the AutoCut folder in Claude Code. Its
.mcp.jsonregisters two MCP servers:filmcraft, a bridge to the running app, andfilmcraft-headless, a headless engine with a demo project. - The agent's entry point is docs/autocut-agent-guide.md. It covers which pipeline fits which request, how to build an edit by hand, checks, export and delivery.
- Give it a brief. This is close to what the measured run got:
Make a finished short film with AutoCut. "Machines in Motion", 60 to 80 seconds, 1920x1080. Use my own footage, found with FrameAtlas: robots, a 3D printer at work, a drive, and one 360° shot reframed with a camera move. Malayalam narration in my cloned voice, told with feeling. Music and a few subtle effects made locally, one Render Anything title, sidecar captions, and no readable faces of bystanders. Local tools do all the media work. Put everything in a new folder and keep a ledger of every step.
A few habits help: ask for a 1080p review export first and the 4K final only after you approve it; have the agent listen back to its narration and run the privacy check; and keep the agent's project file, because everything it made stays editable.
Other MCP clients
Any MCP client can drive AutoCut and FrameAtlas. For example, in Claude Desktop's configuration:
{
"mcpServers": {
"autocut": { "command": "/absolute/path/to/AutoCut/target/release/filmcraft-cli",
"args": ["mcp", "--bridge", "127.0.0.1:9876"] },
"frameatlas": { "command": "/absolute/path/to/FrameAtlas/run.sh", "args": ["mcp"] }
}
}
FrameAtlas's MCP tools include catalog_search, catalog_asset and catalog_resolve; keep ./run.sh serve running so its queue is processed.
Licences and credits
- AutoCut keeps FilmCraft's licence (MIT or Apache-2.0). Huge thanks to the ArtCraft team and FilmCraft's contributors. FilmCraft's name and marks belong to them.
- FrameAtlas is MIT-licensed.
- No AI model is bundled with either project. The models I used have their own terms: MiniMax-H3-Image (pictures; not licensed for use in the EU, the UK, South Korea or the USA), LTX-2 (clips), Stable Audio 3 (sound and music), ACE-Step (music), IndicF5 and VoxCPM (voice), Whisper (speech), Real-ESRGAN (upscaling) and DeepFilterNet (voice isolation). RAEA is a separate install under its own licence.
Both projects are young and move fast, so expect rough edges. Issues and pull requests are welcome on GitHub, and if you want me to go deeper into any part of this, tell me in the YouTube comments.
Comments
Post a Comment