Once an agent is connected over MCP, it works inside your projects with the same capabilities you have.
There are 242 tools, shaped after the editor's own surfaces, so an agent can join a project at any stage — not only at the start. Broadly, an agent can:
Create and read — projects, folders, files, and any file's whole element tree.
Build designs — elements, text, SVG, paths, images and component instances, with styles, variants and design-system tokens.
Wire behavior — properties, states, events, conditions and logic.
Animate — timelines and keyframes.
Manage the design system — colors, fonts, images, sounds, spacing, radius and themes.
Set up data — CMS models, APIs and storage.
Work the database — tables, views, write actions, and the auth built on them.
Produce output — export code, PDF and video.
Check its own work — take preview screenshots, measure rendered elements and read back the generated code.
That last group is the one that matters. An agent that can screenshot what it built and read the code it produced can catch its own mistakes before you see them. And it goes past screenshots: an agent can synthesize a real click, hover or keystroke and read back what changed in your component's state, and it can step through the interpreter's action log. Paste exported React at it and it finds the design nodes that produced it.
The full inventory is in MCP tools.
While an agent works, activity marks appear in the editor — on file tabs, on rows in the explorer, on screen headers and on individual elements — so you can tell which file it is touching right now, and hover to see which tools it called and whether any failed.
An agent's edit is an ordinary undo step. If it does something you did not want, Ctrl/Cmd+Z takes it back exactly as if you had done it yourself — there is no separate agent history to reconcile, and no mode you have to leave first.
That is also why you can work alongside one rather than taking turns. An agent can read what you currently have selected on canvas and act on that, so "restyle this" means the thing you are looking at.
Describing where something lives is slow and error-prone. Right-click a file or element and choose Copy For MCP to get a reference that resolves to the exact file and element — the walkthrough is in Connect to MCP.
Review the export diff. An agent's design change shows up as a code diff like any other. Reading it is the fastest way to see what actually happened.
The agent's project is its own. It is chosen per session and independent of whatever you have open on screen. The two do not have to match.
Prompting habits — naming the layer, asking for one call, asking it to verify — are collected in Example prompts.
The two mostly overlap; the gaps run both ways.
Only an agent can capture a live website and rebuild it as a design, export a PDF or a frame-accurate video, or duplicate an element — the editor has copy and paste, but no Duplicate command.
Only you can drag mesh points or path anchors directly. Those are editor gestures with no tool behind them — an agent adds whole paths and captures keyframes, but cannot nudge an individual point.
Work while the app is closed. The MCP server is part of the desktop app; previews, exports and screenshots are performed by the app itself.
Reach your projects remotely. The endpoint is local and unauthenticated by design — see the security note in Connect to MCP.
Replace your judgment. It builds what it is asked for. Taste, hierarchy and whether the result is good remain yours.
Example prompts — things to try, in order of difficulty.
MCP tools — the full inventory, grouped by area.
Troubleshooting — when the connection or a tool call fails.