Public Beta — Free & Open Source
Open-source
AI-native video editing
AKARI Video is a save-data-centric video editing environment where Claude Code or Codex becomes your editing team. From analysis to cuts, captions and export, the AI finishes the first draft, so when you open it, it's almost done — you fix only what you want to change.
The assets are waiting in the Lab.
AKARI Video Lab is the official library of assets that work in AKARI Video as they are: music and sound effects, telops, LUTs, backgrounds, B-roll and 3D. It isn't plain file hosting. Every asset declares its meaning and its knobs, so AI understands and places it without explanation, and you just turn the knobs to adjust color, motion and text. Start free, and add only what you want as one-time purchases.
- Free tier
- Audio
- Telops
- LUT
- Backgrounds
- B-roll
- 3D
- 2D mockups
- Motion
- Avatars
- Thumbnails
- Templates
- Fonts
Don't wait for rendering.
Tools that build video from code need a render every time you want to check. At the center of AKARI Video is save data — edit.json. The moment the AI rewrites it, the picture in the viewer changes. Rendering happens once, at the final export. Try it with the buttons below.
Changes to the save data show up in the preview as they happen.
Drop in footage. The AI edits.
Open it, fix it.
Drop in your footage
Create a project with skills built in and drop in what you shot. The AI analyzes it automatically, local-first: transcription, scene detection and highlight candidates.
The AI finishes the first draft
It presents a report with thumbnails explaining why each cut was chosen, gets your approval, then edits. You get a first draft complete with cuts, subtitles and music.
Open, fix, export
Check it in the viewer, drag telops, double-click text to fix it. Everything writes straight back to the save data, and a verified export finishes the job.
Beyond "making video with HTML".
Telops and diagrams are really HTML / CSS, the language LLMs are best at. On top of that, AKARI Video made four bets.
Save data is the single source of truth
Cuts, telops, audio, 3D: every editing state is declared in one save file. When the AI writes, the picture changes; when you drag, the data changes. No waiting, no drift.
No presets
Instead of a template pack, the AI writes HTML / CSS / Three.js freely every time. Only the parts you might want to adjust are declared as "knobs", so you can turn them without reading code.
Beyond HTML, there's Blender
3D looks a browser can't produce are baked into video clips with headless Blender. The recipe is the source of truth: swap colors, camera and screen contents and re-bake as often as you like.
An app for checking and fine-tuning
Not a viewer for assembling timelines, but for checking and fixing a video that's almost done. Scrubbing responds instantly with thumbnails, and the real frame catches up when you let go.
Features
The Public Beta covers the full editing loop, from ingest to export.
Ingest and automatic analysis
Transcription, scene detection, speaker and face tracking and highlight candidates are done the moment you add footage. Processing is local-first.
Decision reports and approval gates
Before touching anything, the AI presents a report with thumbnails on why each cut was chosen. It acts only after you approve.
Japanese telops and subtitles
Subtitles that follow dialogue and telops that punch up key moments are designed as separate things. Line breaks by meaning and explicit Japanese fonts are part of the convention.
3D and Blender integration
3D overlays that place glTF models over footage, plus high-quality 3D clips baked in Blender. Recipe knobs let you swap colors and cameras freely.
Narration in your own voice
Generate a scratch narration instantly with free local TTS, lock the timing, then swap in a clone of your own voice. Music ducks automatically.
Verified export
Deterministic QA checks for breakage before export. Afterwards, ffprobe and keyframe checks verify the result, so "I think it exported" never happens.
And more:
Skills come built in.
Every editing step ships with the project as an agent "skill". Just talk to Claude Code or Codex and it runs.
| setup-library | First-time setup: checks your tools and suggests starter assets |
| create-project | Create a self-contained project with skills included |
| analyze-footage | Transcribe, extract keyframes, track and find highlights in footage |
| edit-plan | Decision report → approval → generate edit.json and overlays |
| overlay-authoring | Implement telops, diagrams, 3D, motion and thumbnails |
| bake-3d | Bake 3D scenes into video clips with headless Blender |
| edit-lint | Deterministic QA before export, with an HTML report |
| render-cut | Final export with ffmpeg and automatic verification |
| harvest-asset | Return expensive assets to the library for reuse next time |
| manage-connections | Manage API connections, model choice and cost-approval policy in one place |
Skills are just Markdown and scripts. All of it is public, so you can read, fix and extend them.
The app is free.
The brain, you already have.
AKARI Video has no editing AI built in. The Claude Code or Codex you already use becomes your editing team. No extra subscription, local-first processing, and no path that can charge you without asking.
¥0The app and all skills
The viewer and every skill are open source and free. Get them from GitHub right away.
¥0Local processing by default
Transcription, scratch narration, 3D and export all run on local tools like whisper.cpp, VOICEVOX, Blender and ffmpeg. A full editing loop costs nothing.
¥0Bring your own agent
The editing brain is your Claude Code / Codex. It uses the plan you already pay for, so there's no extra cost.
Opt-inCloud generation is optional
Even for paid cloud generation like voice cloning, the AI declares the cost and runs only after you approve.
From ¥0Assets start free too
Audio, telops, backgrounds, B-roll and thumbnails are available in the AKARI Video Lab. Start free and add only what you need. See the Lab →
Editing is one leg of the loop.
The goal isn't a one-off editing tool but a pipeline that runs from planning and research to post-publish analysis, so each video gets better than the last. Stages are numbered, and skills keep growing on top of them.
● Lit stages = what you can use today in the Public Beta
Posting and publishing (stage 70)
From descriptions, chapters and final thumbnails to scheduled posts. Connects the end of editing straight to YouTube and social platforms.
Analytics → next plan (stage 80 → 00)
Analyze retention and comments, turn them into improvements and feed them into the next plan. The loop closes.
Style learning
Learns your taste from your approvals and fixes, so first drafts get more accurate with every video.
Windows app and a bigger asset catalog
A Windows desktop app (the CLI already supports Windows / Linux) and a growing catalog of assets with declared knobs. A place where asset makers and users meet.
Let's light the lamp together.
Development is fully open. Impressions, bug reports and casual chat are all welcome.
LINE open chat
Questions, sharing what you made and update news land here first. Mostly Japanese, but everyone is welcome.
Get it
Get the app for your OS from the download page. The full source is on GitHub.
What you need
- A Mac with Apple Silicon, or Windows (64-bit)
- Claude Code or Codex, as the editing brain (either works)
- Local tools such as ffmpeg.
setup-librarychecks them on first run and tells you what's missing