GitHub rolled out a new feature for Copilot that might actually change how developers prototype tools. It’s called canvases, and the pitch is simple: describe an interface you need in conversational English, and Copilot builds you a working UI you can interact with and iterate on.

This isn’t code generation in the traditional sense. You’re not asking Copilot to spit out React components or HTML you then have to wire up yourself. Instead, you’re getting a live, functional interface that both you and the AI can use and modify together. Need a form to test API responses? A dashboard to visualize some data? A quick admin panel? Describe it, and the canvas appears.

The examples GitHub showed off include building data entry forms, creating simple dashboards, and prototyping workflows without touching a framework. You can update these interfaces by just telling Copilot what to change. “Add a filter dropdown” or “make that button trigger an export” and it happens.

Who this is for

This feels aimed squarely at two groups. First, developers who need throwaway tools constantly: test harnesses, data migration UIs, one-off admin panels. The kind of stuff that’s tedious to scaffold but too specific to justify a proper build. Second, people who can describe what they need but don’t want to fight with CSS or remember which props go where in whatever UI library is fashionable this month.

The canvas approach also sidesteps a common frustration with AI coding tools: the context-switching tax. When Copilot generates code, you still have to take that code, put it somewhere, run it, debug it, style it. Canvases skip that entire loop. The thing you asked for is just… there.

What it doesn’t do

This isn’t a production app builder. GitHub is positioning it as a prototyping and workflow tool, not a replacement for actual development. The canvases run in the Copilot environment, so you’re not deploying these things or integrating them with your existing stack without extra work.

It’s also limited to what Copilot can understand and generate on the fly. Complex interactions, custom business logic, anything that requires deep integration with external systems—those still need real code. Think of canvases as scratch space, not architecture.

The bigger picture

What makes this interesting isn’t the technology—LLMs generating UIs isn’t new. It’s the packaging. GitHub put this in a flow that developers already use, kept it conversational, and made the output immediately usable. You’re not learning a DSL or configuring a pipeline. You’re just talking to Copilot like you already do.

It’s also a reminder that AI tooling works best when it removes steps rather than adding capabilities. Canvases don’t let you do anything you couldn’t do before. They just let you skip the boring parts.

If you’re already using GitHub Copilot, this is available now in the Copilot app. If you’re not, this probably won’t change your mind—it’s an incremental feature, not a revelation. But if you’ve ever wished you could prototype a quick UI without opening a dozen files and importing half of npm, this might be exactly what you needed.