Underdog on-device AI assistant: free, paid by checkout fees

Thiel Fellow Sigil Wen launched Underdog, an on-device AI assistant pitched against Instinct and Meta's Muse, in an invite-only beta on October 6. The model runs on the user's own Mac or Windows PC. It is free, with no ads. The money comes from a small cut of purchases made through Stripe's payment rails. That makes it one more shopping agent your store will meet, with a business model that only works if it can check out.

What actually happened

Per TechCrunch:

• Local model. Underdog runs a 27-billion-parameter reasoning model fine-tuned from Qwen on Husky, Wen's own inference engine. Wen says it matches Claude Opus 4.6 on some benchmarks. That is his claim, not an independent result.
• No server-side data. User data stays on the device. Encrypted keys handle access to the email and accounts a user authorizes.
• Platforms. Mac and Windows now. Linux, iPhone, and Android are "coming soon."
• Business model. Free and never ad-supported. Revenue is a small percentage of transactions paid through Stripe, which TechCrunch compares to interchange.
• Backers. The company, Conway Research, is backed by Andreessen Horowitz, Khosla Ventures, and others, with Stripe's Patrick Collison and Vercel's Guillermo Rauch among the angels.

Why it matters for your business

Instinct and Muse run in someone's cloud. Underdog runs on the customer's laptop. That changes two things for a store owner.

First, you can't see the profile. A cloud agent at least lives on a platform with policies you can read. A local agent's context about your customer never leaves their machine. That is good for them, and it means your personalization has to come from what the customer tells you, not from a data partner.

Second, the agent gets paid when it buys. A transaction-cut model rewards the agent for finishing checkout, not for browsing. Stores where an agent can find a product, confirm stock, and pay without a captcha wall will get the order. Stores that hide prices behind JavaScript or force account creation will not.

What to do this quarter:

1. Test your checkout with an agent. Can a bot add to cart and reach payment without a human solving a puzzle?
2. Publish structured product data. Price, stock, and shipping in schema.org markup or a feed, not just rendered HTML.
3. Keep guest checkout. An on-device agent will not want to create an account on your site.

Sources: TechCrunch.