Anterra published Whirl, an MIT open-source AI chat app with every top model, long-term memory and self-hosting
Anterra published Whirl, an MIT-licensed AI chat app that routes to every top model through OpenRouter, carries long-term memory across chats, and can be fully self-hosted on Convex, with 281 stars after one day on GitHub.
Image: Anterra
Why it mattersA small team wanting its own ChatGPT-style app no longer has to glue together model routing, memory, artifacts and search from scratch, which means a working private instance in an afternoon rather than a sprint.
An in-house ChatGPT-style product usually means stitching model routing, chat storage, long-running memory, document rendering, tool calls and account billing together, each piece a decision. Anterra published Whirl on 2 October 2026 as the full working source behind the hosted app at whirl.chat, licensed MIT, and the GitHub repository at github.com/whirlchat/whirl reports 281 stars one day after publication, per the GitHub REST API on 4 October 2026.
What Whirl ships
The repository is a full-stack chat app built on Next.js 16, React 19, Tailwind v4 and Motion for the web, with Convex as the database and function host, Clerk for authentication, and OpenRouter for the models, wired through the Vercel AI SDK. Anterra's README lists one conversation that can switch between every top model, with an adjustable thinking level per message.
Beyond the model picker the README names living artifacts, documents, charts and interactive pages that stay editable in a side panel and share by link. It names long-term memory that carries preferences and projects across chats, live web search that reads today's pages, locked chats encrypted on the device with a password the server never sees and answered only by zero-retention models, voice input, image generation, file attachments, folders, incognito mode, message queueing, and an installable mobile web app.
Integrations are over the Model Context Protocol, OAuth included, so a user can connect Whirl to their own tools. The repository also ships installable skills that teach the app new behaviours.
The optional services and the self-hosting shape
Whirl needs three accounts to run end to end: Convex, Clerk, and OpenRouter. Anterra's README calls everything else optional and switches each feature on with its own key: billing through Autumn, web search through Exa, long-term memory through Supermemory, analytics through PostHog and Axiom, tracing through Braintrust, email through Resend, and a support agent through Median. Leave any of them out and the matching feature does not appear.
The repository layout keeps the production web app in apps/v2, an admin console in apps/console, an Expo mobile build in apps/mobile, a waitlist page and Remotion promo compositions in their own folders, and the Convex backend in packages/backend. A legacy earlier web app sits in apps/legacy, kept for reference and no longer maintained.
Anterra ships a self-hosting guide at docs/self-hosting.md that covers the Clerk JWT template Convex needs and the deploy to production, plus a docs/configuration.md for every environment variable and docs/architecture.md for how a message travels from the composer to the model and back.
What the licence allows, and what it does not
The licence is MIT, so Whirl can be run and modified inside a company with no fee owed to Anterra. The hosted services under it are not free: Convex, Clerk and OpenRouter all charge for serious use, and the extra integrations each have their own pricing. The managed chat is a separate product at whirl.chat.
The latest commit on main is from 3 October 2026, one day after the first public commit. The repository has 21 forks after one day, which is early.
Source
- Whirl GitHub repository, primary source: https://github.com/whirlchat/whirl
- Whirl hosted chat product: https://whirl.chat
- GitHub REST API for star and age, queried 4 October 2026: https://api.github.com/repos/whirlchat/whirl
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.