Videos for agents: how AI assistants make real videos

Your AI assistant already writes the launch post, the release notes, the onboarding doc and the code. Then someone asks for a video, and the work leaves the conversation: a separate tool, a timeline editor, an afternoon. Ozor closes that gap. Connect it to Claude, ChatGPT, Cursor, Claude Code or Codex, and the assistant builds the video itself: it plans the scenes, writes each one as animated code, adds narration, and hands you a link to watch it. This is the first post in our Videos for Agents series: what the idea is, how the loop works, and what it costs.

Why agents need a video tool

Agents are good at turning context into artifacts. Give one a changelog and it writes the announcement; give it a spreadsheet and it writes the summary. Video was the artifact they could describe but not deliver. A model can write a perfect storyboard, but a storyboard is not something you can post in Slack.

General video-generation models do not fill the gap for this kind of work. They produce impressive footage, but they cannot reliably put your real numbers, your product UI, your logo and your exact wording on screen, and they cannot change scene four without regenerating everything. Most business video is the opposite: explainers, product walkthroughs, metric recaps, release notes. The content is precise and the format is motion graphics with a voiceover.

That kind of video is closer to code than to film. Which is why the agent that already writes your code turns out to be a very good motion designer, as long as something handles the parts a model cannot: rendering frames, synthesizing a voice, timing scenes to the narration, and making sure what it wrote actually plays.

Who does what: your assistant and Ozor

The connector splits the job along a clean line. Your assistant does the thinking:it reads your brief, plans the story, writes every scene's code and narration, and decides what to change when you give feedback. Ozor does the machinery: it checks each scene before saving it, synthesizes the narration in one of eight voices, stretches the scene to fit the audio, stores everything in your account, renders frames on request, and exports the MP4 at 720p, 1080p or 4K.

Every scene is a React component rendered once per frame at 30 fps, using a shared library with Framer Motion, GSAP and its plugins, Recharts, D3, Three.js, Lottie and the Lucide icon set. That gives an agent real range: kinetic typography, live-looking dashboards, product windows, charts that build, 3D, maps, terminals. Ozor also ships 121 tested example scenes the agent can look up by topic ("kpi dashboard", "logo outro", "chat thread") and adapt instead of starting from zero.

Because Ozor never calls a language model to build these videos, the expensive part (the writing) runs on the AI plan you already pay for. Ozor charges only for what it adds: narration and storage. More on the numbers below.

The five-step loop behind every agent-made video

You never call these tools yourself; you ask for a video and the assistant runs the loop. Knowing the shape helps you ask for the right things.

Between steps 4 and 5 there is a safety net worth knowing about. Before a scene is saved, Ozor checks it three ways without any AI involved: the code compiles, it follows the renderer's rules, and its first, middle and last frames actually render in the engine that exports the MP4. A scene that fails is not saved or charged, and the agent gets the exact error to fix. We cover this in why agent-made videos don't break.

What agents make with it

The best agent-made videos come from context the agent already has. A few patterns we see working:

• →A release video from the changelog. In Claude Code or Codex, the agent reads the last sprint's merged pull requests and turns the three biggest changes into a 45-second narrated video with the real UI rebuilt in code.
• →A KPI recap from a data answer. After an analyst tool explains last week's numbers, the assistant puts each mover on screen as an animated tile with one sentence of narration.
• →An explainer from a document in the chat. Drop a PDF or a policy into Claude or ChatGPT and ask for a two-minute walkthrough for new hires.
• →A launch teaser in portrait. A 9:16 cut for social with kinetic type, a product window and a logo outro, using your brand kit.
• →A one-off answer to a support question. A 60-second visual answer to the question that keeps coming back, with a share link the agent drops into the reply.

There are fifteen more, each with a prompt you can paste, in 15 videos your AI agent can make with Ozor.

Which assistants work

Ozor's connector is a remote MCP server with OAuth, the standard that Claude, ChatGPT and the major coding agents support. In Ozor, open Developer → Connect AI assistants and pick yours:

• Claude on the web or desktop app: add a custom connector. Works on every Claude plan. Step-by-step guide.
• ChatGPT: add a connector with developer mode, available on paid plans. Step-by-step guide.
• Claude Code, Codex, Cursor and VS Code: one terminal command or one install link. Guide for coding agents.

Already in Ozor? The create page has Create with Claude and Create with ChatGPT buttons: type your idea, click, and the assistant opens with the request filled in. The first time, Ozor walks you through the one-minute connection.

What it costs, compared with the app

Same account, same credits, same editor; the difference is who writes the scenes and who pays for that writing.

Through the connector, one credit covers a whole video with up to ten narrations; every ten more cost one more credit. Rewriting a scene's code, previewing frames and exporting are free, so iterating does not burn credits; only new narration does. The free plan includes 6 credits, and paid plans are on the pricing page.

Frequently asked questions