Most AI design tools work like a vending machine: you type a prompt, wait, and get a picture back. If it's close but not right, you type again. Nobody else on your team sees what happened, and the agent you already use every day for code isn't invited.
Floo starts from a different place: a shared canvas, where people and agents work side by side.
Agents are collaborators, not a text box
On a Floo canvas, an agent has a cursor, a status line and a task row, just like a teammate has a face in the top bar. You can watch Claude Code draft a confirmation screen while Codex syncs tokens to your repo and a colleague leaves a comment on the checkout button. When the agent finishes, its task turns done and lists the frames it touched; reply to it and it picks the work back up.
Any MCP client can join: Claude Code, Codex, Cursor, Claude, ChatGPT and others sign in once with OAuth, and headless agents use a key. They act as you, with exactly your permissions, and revoking the key stops them at once.
Every frame is real HTML
Frames render live HTML in sandboxed iframes. That means the thing you approve is the thing a browser will show: real type, real layout, real states. It also means agents work in the medium they're best at, and that handing a design to a coding agent is not a translation step.
Or let Floo do it, on your key
No agent of your own? The built-in Floo Agent designs from a brief, an Ask AI on an element, or an @mention in a comment. It runs on the Claude, OpenAI, Gemini or OpenRouter key you connect, so you pay your provider directly. Send the work through specialist roles — copy, brand, accessibility, a polish pass — and each one does one job well.
Review where the work is
Comments pin to elements. Approvals record the exact version someone signed off on, and drop back to needs review when the design changes. Prototypes wire frames together with real transitions, and anyone with access can click through from a link.
Floo is free for personal use. Teams are billed per seat. Start a canvas and invite an agent.