Google Labs turns Stitch into a voice-powered design canvas

First came vibe coding. Now, Google wants to make vibe design a thing.

Google Labs rolled out a massive overhaul to Stitch, transforming the experimental tool into an AI-native design canvas aimed at turning plain English into high-fidelity, clickable UI prototypes.

Talk it into existence

The redesigned Stitch centers around an infinite canvas where you can dump raw context—rough text prompts, reference screenshots, or existing code. Instead of drafting wireframes or picking components manually, you describe the user vibe, emotional tone, or business goals you are aiming for.

Behind the canvas sits a new design agent that tracks the full history of your project. An accompanying Agent Manager lets you branch off and explore multiple visual directions in parallel without losing your original work.

Google also added voice-driven critiques. You can talk to the canvas in real time, treat the AI as a sounding board, or verbally ask it to show three menu variants and swap color palettes mid-conversation.

From DESIGN.md to working prototypes

Stitch is also tackling the handoff problem. It can instantly string static mockups together into interactive user flows. When you click a button in preview mode, the tool can automatically generate the logical next screen to map the journey.

To keep styling consistent across tools, Google introduced DESIGN.md, a markdown format for storing portable design system rules. Users can extract design systems directly from existing URLs, import or export rules across projects, and bridge work into engineering pipelines.

Stitch also integrates with a recently released MCP (Model Context Protocol) server and SDK—which has already collected 2.4k GitHub stars—along with direct export hooks into developer tools like Google AI Studio and Antigravity.

Why it matters

For solo founders and fast-moving teams, this is a massive speed boost. Building clickable product demos used to require juggling Figma components and basic front-end scaffolding for days. Collapsing that process into a single voice conversation and an instant prototype makes early idea validation dramatically cheaper.

Google’s smartest architectural choice here is playing nice with existing developer stacks. Introducing DESIGN.md as an agent-friendly format and supporting MCP shows Google knows design cannot live in an isolated sandbox. It has to talk to code.

That said, vibe design still has limits. It looks great for rapid divergence and early iterations, but enterprise software demands edge-case handling, strict accessibility, and pixel-level precision that probabilistic prompts struggle to nail reliably. Even so, it puts serious pressure on dedicated UI tools to make agentic, conversational generation a core feature rather than a side plugin.