The agent

The launcher bar bottom-left reads Describe what you want to build…; click it or press ⌘J. The panel resizes from its edges, maximises, minimises and remembers its size. Its header carries a model menu for the agent itself, a chat-history dropdown and Agent settings. Signed-out visitors are asked to sign in first — the agent spends your budget, so it needs to know whose.

On an empty canvas it suggests starting points — Redesign this UI screenshot, Find icons for my dashboard, Build a color system from this image — but any request works. Attach reference images with the paperclip, or drop them straight onto the panel.

Build the graph

Create any node type, connect nodes, delete them, and lay out a whole plan atomically — several prompts, their outputs and groups in one step.

Run it

Execute the workflow in dependency order, or run one node, and read the results back to decide what to do next.

Insert a prebuilt workflow

Add any of the maintained prebuilt flows — Press Kit, Brand Kit, App Screens — into the current canvas without disturbing what is already there.

Edit images for free

Remove background, trim, rotate, flip and resize run instantly and cost nothing; the result lands as a new node under the source.

Find icons

Search the Unicon icon libraries, fetch individual icons or a starter pack, and place them on the canvas.

Tidy the canvas

Re-arrange overlapping nodes so the graph reads left to right.

Export a press kit

Combine rendered press-kit pages into one landscape PDF and download it.

Edit code and screens

Inside the screen editor and shader studio, the agent edits the document directly — targeted find-and-replace edits or whole rewrites — and can set shader knobs. See Editors.

Select nodes and press ⌘C to copy a reference; paste it into the composer and it appears as a removable chip. Under the hood it is a [node:…] token the agent resolves to the exact node, so “make this one warmer” is unambiguous. Prompt nodes accept the same tokens via @, which is how one prompt can quote another’s output.

While the agent works, the send button becomes Stop and Esc stops the run without clearing what you have typed. Type instead of stopping and the composer reads Steer the agent — send to redirect it…: sending interrupts the current run and starts the next turn immediately with your correction. A stopped turn shows Stopped by you in the transcript; nodes it had already created stay, so you never lose finished work to a change of mind.

• History. Each canvas keeps its own conversations. Start a New chat, switch between past ones, rename or delete them from the dropdown. New chats title themselves from your first message.
• Private to you. Conversations belong to the person who had them, even on a shared workspace canvas — a teammate or admin opening the same canvas sees the nodes, never your transcript.
• Agent rules. Agent settings holds a free-text rules field — Always use TypeScript. Prefer functional components. — applied to every conversation you have.
• Long sessions. When a conversation grows past what the model can hold, earlier turns are compacted into a summary and the work continues. In the editors, the agent also keeps a small set of memories per document you can review and forget.
• Reasoning. The agent’s thinking collapses into one expandable block per turn; every message has a Copy action.

From outside the browser

The same capabilities are exposed to agent clients over MCP, so Claude Code, Codex or your own agent can read a canvas, add and connect nodes, run generations and workflows, and hand you a link to the result. See the MCP server.