A motion graphics editor that runs on Claude Code instead of the API

Caleb Porzio launched a new Flux card component with a short, beautifully animated video. A few minutes later he showed what made it: a tool he built for himself, which he described as After Effects without having to know how to use After Effects.

From the screenshot, it's a storyboard. A strip of scenes runs along the bottom, each with its own length. The current scene plays in a preview with a scrubber. Beside it is a chat where you ask for changes, like "hold on the headline a full second longer", with undo and a version history for every scene. There's an art direction panel, a render view, and a check that compares the frames at every cut so a change in one scene doesn't cause a visible jump into the next.

The replies were what you'd expect: people want it. When asked to release it, Caleb's answer came down to cost. A tool like this calls Opus on every change, and at API prices that adds up quickly for anyone iterating on a video.

That makes sense for a tool you would ship to other people. It made us wonder whether a personal tool has to work that way at all.

The idea

We don't call the API to get work done. We run Claude Code, on a Claude subscription, and we drive it through Solo. Solo ships with an MCP server, a CLI, and a REST API that let one agent start another, hand it instructions, and wait for it to finish. That's how most of our work gets done, including our video work. We already keep a repository just for agent-made video, where Claude Code builds explainers in Remotion and renders and checks its own frames.

So we got to thinking, what if the editor's backend were a Claude Code session instead of an API client? The storyboard UI would stay the same. When you type "slide the card in from the right", the request goes through Solo to an agent working in the video's project. The agent edits the scene, renders it, looks at the result, and reports back. The editor shows the new version when it lands.

This will not be a product

We want to be clear about that up front. Now that Anthropic has seen how excited people get about Opus making motion graphics, we'd be shocked if a proper editor for this didn't show up inside Claude Desktop, or somewhere like Claude Design, within weeks.

We want to improve our own video production workflow, and I think building this could benefit us. It also lets us keep it vendor-agnostic. Right now, Opus 5.5 is the clear leader in stop-motion video. But things change fast in this industry, and if it works, this would let us test models against each other and swap whenever it makes sense.

What we expect to be hard

The loop has to feel fast. An API call returns a response. An agent session takes time to pick up a message, do the work and go quiet. That's fine for a big change and painful for "a touch slower". We expect to keep one long-running agent per video rather than starting one per message, and the first thing to measure is how long a small tweak takes, end to end.

Getting an answer back. Solo is built for orchestration: send an instruction, then wait until the agent is idle. An editor wants something closer to request and response: which scene changed, what the agent did, and where the new render is. Our guess is that the editor should watch the project itself (the files, the renders and the git history) rather than try to parse the agent's chat.

Versions for free, maybe. If every accepted change is a commit in the video's repository, undo and version history come from git instead of being something we have to build. That feels like the right shape, but it has to survive an agent that wants to tidy up after itself.

The agent has to see its own work. The best part of Caleb's tool is that it checks the seams between scenes. Our video agents already render frames and inspect them, so the pieces exist. Wiring them into every change without making each one slow is the open question.

It isn't free, just paid for differently. A subscription has usage limits, and iterating on video can churn through that usage pretty quickly. Part of the experiment is finding out how far a normal week of video work gets before that matters.

What success looks like

A local storyboard, pointed at one of our existing video projects, where we can pick a scene, type a change, and see the new render in the preview without leaving the page. That's all driven by Claude Code through Solo, with no API key involved. If it's good enough that we reach for it instead of briefing an agent by hand, it's worth keeping.

Whether we succeed or fail at this, we'll leave the repo public so that you can pull it down and iterate on it yourself.