A commercial AI chat app, MIT-licensed and self-hostable

What it does Whirl is a full-stack AI chat application — the actual production code behind whirl.chat, published under MIT. It routes conversations to essentially every major model through OpenRouter, with adjustable thinking levels, and layers on long-term memory, live web search, MCP integrations with OAuth, and “living artifacts”: documents, charts, and interactive pages that stay editable in a side panel and can be shared by link.

The interesting bit The optional-services architecture is the quiet star: billing, search, memory, analytics, tracing, and email each run on their own third-party service, and if you don’t provide the keys, Whirl simply hides those features rather than breaking. Self-hosters get a working app with just Convex, Clerk, and an OpenRouter key.

Key highlights

• Locked chats are encrypted on-device with a password the server never sees, and answered only by zero-retention models
• MCP integrations for your own tools, plus installable “skills” that extend behavior
• Incognito mode, message queueing, voice input, image generation, and an installable mobile web app
• Monorepo includes an admin console, an Expo mobile app, and even Remotion promo-video compositions
• Real docs: self-hosting, configuration, and a message-lifecycle architecture guide

Caveats

• Self-hosting still requires Convex and Clerk accounts — it’s not a single-binary deploy
• The repo carries a legacy/ previous web app, kept for reference and unmaintained

Verdict Worth a look if you want a polished, self-hostable ChatGPT-style app you can actually read and modify. Skip it if you were hoping for something that runs without any external SaaS dependencies.