I make short videos for clients and for my own accounts, and for a while now I've made them the same way: I describe the reel to a coding agent and it writes the edit as code. Invokard Studio is that method, packaged. It is a skill for Claude Code, Codex, Gemini CLI, Cursor and Antigravity, plus the engines it drives. Version 1.1 adds a second engine that draws animation, and Studio now installs as a Claude Code plugin. It's free and MIT.
Client footage cut on the beat: 34 phone clips, the client's script on screen, a wipe on the logo, everything on a 115 BPM grid. No video was generated.
The agent never has to watch anything
An agent can read a file and look at a picture. It can't sit through 40 seconds of video and notice that a caption came in late. So the skill turns every judgement it needs into something it can read or look at:
-
Timing comes from the audio. One voice file per sentence gives the engine the pauses, and the captions and cuts follow from them, word by word. With a recorded narration or a song,
align.py(faster-whisper, running locally) gives the time of every word. - The edit is reviewed on stills. The whole reel is laid out as a contact sheet before any clip is generated, and the agent fixes framing and layers there.
-
The final file is measured.
media.py verifyreads the MP4 itself: its specs, its loudness (the target is −14 LUFS) and a contact sheet taken from the encoded video, not from the engine.
Ekilib, 45 s: nine sentences of voice, thirteen illustrations in the client's style, an animated glucose card, a numbered list and a morph.
Two engines, one script
The montage engine (Python, Pillow and ffmpeg) composes every frame of a reel: word-by-word captions with the key word highlighted, animated headlines and chips, charts, before/after wipes, cuts on the beat, end cards, and a soundtrack that ducks the music under the voice. The same script renders 9:16, 16:9 or 1:1.
The animation engine (Node.js) is for what is designed rather than filmed. A composition is a React component that draws one frame from its frame number. The engine opens it in the Chrome or Edge you already have, renders in several tabs at once and pipes the frames to ffmpeg: H.264 MP4, ProRes 4444 with transparency, or PNG frames. It has springs, interpolation, sequences, images, video clips and 3D with three.js. Google Fonts stay on disk after the first download, and the same input gives the same bytes.
A character drawn in SVG, a can in three.js, fruit falling with planck.js physics. They are three of the four animated ads in examples/animation; the fourth writes a gym's name with thousands of particles.
The two engines meet in reel.py. An animation can be drawn over the montage as a transparent layer, or cut in as a shot, with its numbers passed from Python:
import animation as an
an.render(ENTRY, 'Headline', rk.P('anim/headline'), transparent=True) # PNG frames with alpha
reel.ov(t0, t1, an.layer(rk.P('anim/headline'))) # drawn over the montage
an.render(ENTRY, 'Counters', rk.P('anim/counters'), transparent=True, props=COUNTS)
reel.shot(t2, t3, an.clip(ENTRY, 'EndCard', 'endcard')) # a drawn shot
The montage + animation example: captions on the voice, a kinetic headline and a counter card drawn over the montage, and a drawn end card. It renders from nothing, because it draws its own pictures and synthesises its own audio, and CI renders it on every push.
Install
In Claude Code:
claude plugin marketplace add Octonove/invokard-studio
claude plugin install invokard-studio@invokard-studio
Then ask for a reel or an animation. For the other agents, clone the repository and run python install.py: it puts the skill where Claude Code, Codex, Antigravity, Gemini CLI, Cursor and Windsurf look for it. You need Python 3.10+ and ffmpeg, and Node.js 18+ for animation. The first example renders a whole reel with no account, no API key and no download:
python examples/first-reel/reel.py
What it costs
The code is free and MIT. Rendering your own footage, photos and voice costs nothing. When a reel needs something that doesn't exist yet, like an AI image, a clip, a voice or music, the skill can use a media provider you connect, and it says so once instead of assuming you have one.
The repository has the three client reels, the four animated ads and the scripts behind them: github.com/Octonove/invokard-studio. The videos also play on its page: invokard.com/tools/invokard-studio.






Top comments (0)