Rive's AI Agent or Your Own? Connecting Claude, Cursor and Codex to Rive

Almost the entire Rive Scripting Course so far was written by the agent built into the Rive editor. I counted before writing this: 18 scripts, about 4,600 lines of Luau. A star field, planets, a nebula, a sun, a speed gauge, a fuel bar, a lightning cannon, an asteroid field. I typed prompts, it typed code, and that is exactly what the course teaches.

Two days ago I was recording the next lesson. Hit detection on the asteroids: hover over a rock, a ring locks onto it. Same routine, prompt in, watch it work.

Two things happened while I waited. A small line appeared under the chat: "Credits are low ($9.45 left)". And the agent took long enough on our 800-line Asteroids script that I had time to read that line twice.

That was the moment I decided the rest of the course would be built differently. Not because Rive's agent is bad. It isn't, and every prompt in the course still runs on it. But I already pay for an AI agent, and it turns out I can point it at Rive.

What MCP actually is, in one paragraph

MCP is a standard way for an AI agent to talk to a tool. Rive built a small server into the desktop editor. Any agent that speaks MCP, Claude Code, Cursor, Codex, can connect to it and get the same abilities the built-in agent has: read the file, edit scripts, create ViewModels, wire State Machines, export. The editor stays the editor. What changes is who is driving.

Why I switched

One bill, not two. Rive's agent runs on credits you top up inside the editor. If you already pay for Claude, Cursor or Codex, that subscription drives Rive too. Nothing else to buy.

It's faster on a real project. Early in the course the scripts were fifty lines and everything felt instant. Now the Asteroids script alone is past 800 lines, and the bigger the file, the longer every change takes the built-in agent. A connected agent edits the ten lines that need changing and is done in seconds.

It remembers. Rive's agent starts every session blank. It doesn't know what we built last week, what values we tuned, what we tried and threw away. Your own agent keeps a project folder, a notes file, memory. It gets better lesson after lesson instead of starting over.

It's not limited to Rive. Both agents read the Rive file. Yours also reads your notes, Rive's docs, your app's code, and it doesn't stop when the edit is done. More on that below.

Same prompts, same craft. Describing what you want to see, saying what's wrong on screen, protecting what already works: the prompt skill you build with Rive's agent is exactly the one you use with your own. Nothing is wasted.

Not limited to Rive

The biggest advantage is not about Rive at all. Rive's agent works inside Rive. Your agent works everywhere: it can export a file, save it, and do whatever you do next with it.

The hero on the Scripting Course landing page is a good example. This is it:

When I built it, Claude did more than add the shaders to the project. When we were done, it exported the final file, uploaded it to the site, and published it. It even noticed the site was running an older version of the Rive runtime and updated that too.

That is the difference between an agent and a built-in agent. It knows about more than Rive, it can look at more than Rive, and it has access to more than Rive.

The honest limits

This is a young integration and it shows. Creating a new script is asynchronous, so a script can take a few seconds to appear. Screenshots taken from the editor don't run scripts, so you verify on screen or in a browser. The editor is sandboxed, so exporting to disk goes through a small workaround. Rive has to be open with your file open, and you start the agent after Rive so it discovers the tools. And once, mid-session, Rive crashed. None of this changed my mind. All of it is worth knowing before your first session.

Setting it up

There is nothing to install on the Rive side. Open Rive with your file, then ask your agent to connect to the Rive editor's MCP server. Claude Code, Cursor or Codex takes it from there. If you prefer the step by step, Rive's official guide has it, with two short videos:

Claude Code and Cursor are the two clients Rive documents. Codex is announced by Rive as supported, and since the server is plain local HTTP, any MCP client should work. I have only used Claude Code myself.

Takeaways

1. If you already pay for an AI agent, connect it to Rive. If you don't, Rive's agent is fine. Don't pay twice.
2. Give the agent a home: one folder with your Rive files, a notes file with your decisions and the values you tuned. That folder is the memory the built-in agent never has.
3. Prompt in features, not code. What's on screen, what's wrong with it, what you want to see, and what must stay the same.
4. Rive open first, file open, then the agent.
5. Verify on screen before you believe "done". Numbers that compile are not numbers that look right.

In the Scripting Course, the lessons from The Explosion module on are recorded this way, with a skill file for your own agent included so it starts every session already knowing how we work in Rive. The prompts are the same in both worlds. Pick the one you already pay for.