Startup Conway Research has launched an invite-only beta of Underdog, an AI assistant that runs on users’ computers and plans to earn money from purchases rather than advertising. It pairs a privacy-first pitch with a distinctly practical question: what will users let an assistant do when it can act on their behalf?
Watch Desk analysis
What happened
Founder Sigil Wen told TechCrunch that Underdog currently runs on Macs and Windows PCs, with Linux, iPhone and Android versions planned. Its 27-billion-parameter reasoning model is fine-tuned from Qwen3.8 27B. Wen says the assistant can handle everyday tasks such as shopping research and maths questions, and says it compares favourably with Claude Opus 4.6 on some benchmarks. Those are his claims, not an independent assessment of the product.
The software runs on the user’s device, rather than sending requests to a cloud AI provider, and Wen says it encrypts keys for email and other accounts users authorise it to access. The app will initially be free and, he says, will never carry ads. Instead, Conway Research plans to take a small percentage of payment transactions made by the assistant through Stripe’s payment rails. TechCrunch’s report says the startup is backed by several investors, including Stripe co-founder Patrick Collison.
Why it matters
Underdog brings two consequential choices together: keeping AI processing on a person’s own machine, and making money when the assistant completes a transaction. Local processing could reduce the need to send sensitive information to an AI provider. But a shopping assistant that earns a cut when it buys something also gives its business model a direct stake in what it recommends and how often it acts.
The model size matters, too. Underdog is not claiming to match the largest hosted systems across the board; it is betting that a smaller model running locally can be capable enough for everyday jobs. That makes the real test less about a grand privacy slogan and more about which tasks work well, what permissions the assistant needs, and how its recommendations behave when a transaction is on the line.
Our read
A free assistant that runs locally and avoids advertising is an appealing proposition, especially when the alternative is handing personal context to a cloud service. But “private” is not a complete product review. We would want to see how permissions, account credentials, recommendations and payment approvals work in practice before trusting an assistant with the digital equivalent of the shopping trolley.
What to watch
- Whether the invite-only beta supports Wen’s claims about everyday tasks and benchmark performance.
- Which operating systems get support, and when.
- How Underdog discloses transaction fees, recommendations and user approval before purchases.
Discussion spark: Would you trust an AI assistant to make purchases if it ran on your own device, or does a transaction-based business model create too much incentive to steer what it buys?
Sources and evidence
- Silicon Valley's AI wunderkind launches Underdog, the most private Instinct/Muse competitor yet – TechCrunch (6 October 2026, 20:47 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.