How Glidezo makes a video with your AI

You ask your AI for a short video. Your AI writes the plan and the scene code. Glidezo checks it, renders it in a locked sandbox, and charges credits only when the video is made.

Connect your AI

A video made this way

The steps

1. 1. You ask. In a chat with Claude (with Glidezo connected), say what you want: "a 20-second Short about why cats knock things off tables". Your AI asks for anything missing, such as the goal and the mood. The rest comes from your video settings. Videos are 15 to 30 seconds.
2. 2. Your AI writes the script. Glidezo gives it a writing guide (the structure, how many words fit the length) and checks the script it sends back. Your AI shows you the script, and nothing is rendered until you approve that exact version.
3. 3. Your AI plans the scenes. Each scene is either a ready-made block (a bold claim, a list, a chart, a quote, an end card) or scene code your AI writes itself: a small JavaScript program that draws the scene frame by frame, in 2D (Canvas) or 3D (Three.js), with Glidezo's starter characters and props if it wants them.
4. 4. Free previews while it works. Your AI can ask for still frames of a scene (a contact sheet) to see its own work and fix it, before anyone pays. Free, up to 4 a minute and 60 a day.
5. 5. A price, then a yes. Each saved plan gets a price in credits. Your AI is told to show it to you and wait for your yes before it renders.
6. 6. Glidezo renders it. The credits are held, the video is rendered and checked, and then the credits are charged. If anything fails, they are handed back. At most 3 renders an hour and 10 a day.
7. 7. You review it. Every video gets its own review page (/v/<id>, only you can open it): the video and its download, a free live preview of each scene, and simple edits (a line, a scene's text, swap a block, move or remove a scene). Your AI gets the link and can make the same changes. A new render reuses the scenes that didn't change when it can.

What Glidezo checks

• The order of the steps. No render without an approved script, and no charge without a price made for the exact plan.
• Every piece of text (topic, script, text on screen, text inside scene code) goes through a safety check (Llama Guard 3 on Cloudflare Workers AI). Some topics are refused.
• Scene code is read before it runs: only Glidezo's own kit can be imported, and things a scene may not use (the network, timers, clocks, code built from strings) are refused with a message your AI can act on.
• The frames. Blank, frozen or flashing frames are errors; a face, hand or key object cut by the frame's edge, or under the app buttons or the captions, is a warning. Your AI gets each one back with the time and the fix.
• The finished video. Its frames get an image safety check before it is yours.

The locked sandbox

Scene code written by an AI is treated as untrusted. It runs on Glidezo's render server (Cloudflare Containers, on CPUs) like this:

• the container has no internet: its only way out is Glidezo's own storage, for that job's files;
• no keys or passwords are inside it;
• the code runs in a locked JavaScript runtime (SES), as its own unprivileged process, with time, memory and size limits;
• every library it can use (Glidezo's kit, Three.js, the fonts) is bundled in the image: nothing is downloaded.

How the video is made

• Voice: Kokoro. Kokoro-82M, an open-source text-to-speech model (Apache-2.0), reads each line in the voice you chose, on the render server. Glidezo keeps each spoken line, so an unchanged line isn't spoken again.
• Word timing: Parakeet. NVIDIA's Parakeet TDT 0.6B v3 speech-recognition model (CC-BY-4.0) listens to the voice and finds when each word is said, so the captions light up on the word. The captions keep your script's spelling.
• Music and sounds: CC0. The background music and sound effects are CC0 (public domain) files from OpenGameArt.org. The music plays quietly and dips further while the voice speaks.
• Drawing and joining. Each scene is drawn frame by frame (1080p, 30 frames a second) by Glidezo's renderer, and FFmpeg joins the scenes, transitions, captions, voice and music. The voice is levelled to -14 LUFS. Every frame says "Made with AI · AI voice".

What it costs

Connecting, planning, scripts, still previews and the live preview are free. Your first video is free (Standard quality, no 3D scene, up to 20 s). After that a render costs credits, priced by the video's length and how much drawing it needs: a 20 s video with template scenes is 24 credits, scenes your AI draws in code add per second, and HQ costs 1.5x. Each video shows its own price before you say yes. See pricing for the full table and credits.

Honest limits

• English only, 15-30 second videos, Claude only for now (ChatGPT is next).
• The live preview on the review page is silent with estimated timing, and doesn't draw 3D scenes yet.
• How good a code scene looks depends on your AI's code. Glidezo's checks catch broken frames, not dull ones.