GitHub Gave Away Its AI Prototyping Tool: Meet Replica

GitHub's design team built an AI tool that turns an idea into an interactive prototype, and on October 28, 2026, they open-sourced it under the MIT license. It's called Replica. GitHub says what sets it apart from generic AI prototyping tools is that it builds from your team's own components, patterns and documentation, so prototypes reflect how your product actually works.

What happened

GitHub announced Replica in its changelog on October 28, 2026, as "an AI-powered prototyping tool created by GitHub Design that helps anyone on your product team turn ideas into interactive prototypes using your own design system." The repository credits GitHub Design and Microsoft CoreAI Design.

Here's how it works, according to GitHub:

• Designers define the guidance. Your team's existing components, patterns and documentation shape what the AI generates.
• Anyone on the team describes what they want in natural language in the GitHub Copilot app. The Replica canvas, a panel in the Copilot app, shows the running prototype beside the conversation.
• Prototypes can be shared through Replica Hub, an optional, self-hosted service where teammates explore them and leave comments.
• It's yours to change. The tooling and guidance are open source and meant to be customized.

The repo ships two starters: Minimal, a small example app to adapt to your own design system, and CoreAI Sandbox, an app with a navigation layout and prebuilt interface elements from Microsoft's CoreAI design system.

The catch

• You need an AI model. The documented workflow runs in the GitHub Copilot app and requires a GitHub sign-in. You can use GitHub-hosted models with a Copilot plan, or connect your own model provider without a Copilot plan. That bring-your-own-key option is in public preview, and the provider's costs apply.
• Mac or Linux. Local canvas previews are supported on macOS and Linux.
• Team sharing means hosting. Publishing and comments need a Linux server, a dedicated domain with wildcard DNS and HTTPS, network access, and GitHub permissions. Hub keeps publishing and review activity in its own database for 90 days, and GitHub notes that a self-hosted Hub "does not mean that no data leaves your organization," since GitHub services and your AI provider have their own terms.
• It's young. GitHub says Replica is under active development.

The good news: you can prototype locally without Hub, Docker, a domain or even a GitHub repository. Viewing a finished prototype or reviewing it in Hub doesn't need a Copilot subscription.

Try it locally

1. On macOS or Linux, install the GitHub Copilot app and sign in, plus Git (with user.name and user.email set, for the installer's first commit), Node.js with npm (the version in the repo's package.json engines.node), Bash, curl, tar and lsof.
2. Download the starter installer and read it before you run it: curl -fLSs https://raw.githubusercontent.com/github/replica/main/scripts/install-starter.sh -o install-starter.sh
3. List the starters with bash install-starter.sh --list, then create one: bash install-starter.sh --starter minimal my-prototype.
4. In the Copilot app, open the new folder and start a fresh session there. Before you accept its setup, know that session startup can install Copilot CLI globally, install or enable the Replica plugin and run npm ci.
5. Turn on the canvas: under Customize → Installed, enable both the replica-prototype plugin and its canvas extension (they're separate switches). Start another fresh session, then open + → Canvas → Replica prototype.
6. Ask Copilot to run the app locally and show it in the Replica canvas.
7. Describe one screen you'd like to test, and refine it in the canvas until it feels right.

The local prototyping guide in the repo walks through every step, including adding your own design-system guidance.

Here's how you get paid

My opinion, no income guaranteed: plenty of product teams have a design system and a backlog of ideas nobody has time to mock up. Sell them a prototype sprint:

• Load their design system. Set Replica up with their components and guidance in a prototype repository they own.
• Build interactive demos of the ideas they're arguing about, in their own look.
• Test with customers before engineering builds anything, then hand over the repo so their team can keep going.

Agree up front on which AI model access they'll pay for (a Copilot plan or their own provider), and whether they want a hosted Hub.