Video AI Editor
The media library is analyzed locally (Qwen3.5), and an agent assembles a rough cut from a phrase like “make a punchy one-minute clip from the sea shots over this track”.
- Format
- Local applicationPython backend + React; the media library stays on your disk
- Status
- In development
- Stack
- Qwen3.5 · Ollama · ffmpeg · librosa · FastAPI · React · MCP
- Direction
- Computer Vision
About the product
Video AI Editor is a rough-cut editor that first understands what was shot, then assembles the clip. Point it at folders of footage: a model describes every video and photo locally — what is in the frame, which scene starts at which second, how good the frame is — and finds the tempo and bars of the music. Then edit on the timeline yourself, or give the agent a brief — “exactly one minute, to this track, nature and streets only” — and get a draft with cuts on the beat, ready for a manual finish.
It earns its place where there is a lot of footage and no time to scrub through it: marketing and content teams with an archive of shoots for social media, production houses that need to find a shot by meaning (“a boat on the river at sunset”) instead of hours of scrubbing, and companies with a private archive that cannot upload video to cloud services. Instead of “watch everything and pick” — search by descriptions and scenes.
Media library
Folders of video, photos and music; scanned via ffprobe
Local analysis
Qwen3.5 describes frames and scenes; librosa finds the beats
meta.json
Descriptions, objects, quality, beat grid — next to the library
Editing
Timeline by hand, or the agent from a text brief
Preview and export
Proxy in the player, then ffmpeg to MP4, MOV or WebM
The key difference is editing from computed data rather than prompt wishes. The agent is not asked to hit the beat or avoid dark frames — it gets the beat times of the track and a quality grade for every scene, and places cuts on them. The same tools are open to external agents over MCP: Claude Code or Claude Desktop edit the project, and the changes appear in the UI right away. All media content is analyzed on your machine; only the text of the brief leaves it, and only if you use the agent. The code is open under GPL-3.0.
What it looks like
A real project: a one-minute cut of holiday footage, assembled by the agent to a track. Click a frame to enlarge.
Timeline and preview
File metadata
Clip properties
Export
Arrows, dots or drag with the mouse; in the enlarged view use ← → and Esc.
Who it is for
Marketing and content teams
Quick drafts from the archive for social media and presentations — from a text brief, then a manual finish.
Media and production
Find the right shot by meaning and timestamped scenes instead of hours of scrubbing the timeline.
Companies with a private archive
Media content is analyzed locally; only the text of the agent request leaves — and only if you use the agent.
How it works
- 1
Index the storage
Recursive incremental folder scan, properties via ffprobe, thumbnails, flagging of corrupted files.
- 2
Analyze
Frames are sampled at a step and downscaled to 640 px, Qwen3.5 describes content, scenes with timestamps and the quality of every frame locally; tempo and bars are computed for audio. Everything goes to meta.json.
- 3
Edit
Manually on the timeline or by a request to the agent: it finds material by meaning and works through the same operations layer as the UI.
- 4
Preview and export
A 640×360 proxy in the built-in player, then ffmpeg export in the chosen format and quality.
What the product does
Local media analysis
Qwen3.5 via Ollama: description, objects, style, mood; for video — scenes with timestamps so you can take a fragment inside a long file.
Frame quality as data
Every frame and file gets a grade — high, medium or low: dark, blurred, motion-smeared or noisy. The agent skips bad material by this field, not because a prompt asked it to.
Beat grid for music
Tempo, beats and bars, pauses and loudness peaks are computed once — a track is picked by tempo, cuts land on the beat, and transitions go only on phrase boundaries the music actually has.
Timeline editor
Three tracks (video, audio, text): trim, split, speed, fade, transitions, text layers, undo and redo.
Editing agent
An LLM with function calling assembles a draft on top of the same timeline operations; material is found by meaning via description embeddings.
Preview and export
Cached proxy render, built-in player; export to MP4/H.264, MOV/ProRes, WebM/VP9 with progress and cancel.
Loudness to standard
The final mix is normalized to −14 LUFS — the streaming-platform level — in both preview and export. Projects cut on differently mastered tracks don't jump in loudness relative to each other.
MCP server
The same editing tools for Claude Code / Claude Desktop: an external agent edits the project, changes appear in the UI right away.
Open source
GPL-3.0, sources on GitHub. The agent's editing rules live in a separate EDITOR_AGENT.md read on every run — the prompt is editable without touching code.
More products in this direction
Want to see the product on your data?
We will show a demo, walk through your scenario and suggest a deployment format.