A privacy-first personal AI assistant from Conway Research that runs entirely on your own computer (currently Mac with Apple Silicon, Windows beta), using a 27B reasoning model fine-tuned from Qwen3.8-27B. It handles email, calendar, meeting notes, dictation, shopping research, and errands, with all content—recordings, notes, emails, and learned memories—stored locally and encrypted (keys in Apple Keychain). Pricing is free, with no ads and no subscription; founder Sigil Wen argues on-device inference has near-zero marginal cost, so the company plans to earn a small commission on payment transactions made through the assistant via Stripe (similar to an interchange fee), backed by a16z and Stripe co-founder Patrick Collison. Critical caveats: it is in invite-only beta (not publicly available), the 27B model’s real-world capability has not been independently validated at scale, and the business model depends on users trusting the assistant with their payment cards.

Pros & Cons:

Pros: The local-first + free combination is genuinely unusual—unlike cloud assistants that charge subscriptions to cover GPU costs, Underdog runs on your hardware and claims it can stay free forever, with no ads and no data harvesting; the 27B model (Qwen-based) is larger than most on-device models, and the founder claims performance close to Claude Opus 4.6 on some benchmarks, which if true would be impressive for a local assistant; privacy controls are concrete: all data stays on-device, keys are encrypted, and a “Naval mode” can prevent the app from calling home except for essential servers; the investor lineup (a16z, Khosla, Patrick Collison) provides strong early validation.

Cons: It is invite-only and very early—no public availability, no large-scale user feedback, and the 27B model’s real-world quality (especially in noisy environments or complex tasks) is unverified by independent third parties; the business model is unproven—earning enough from payment commissions to sustain a free AI assistant requires enormous user trust and transaction volume, and it remains to be seen whether users will let an AI make payments on their behalf; it only runs when your device is on—unlike cloud agents (Muse, Dots) that work 24/7, Underdog requires your Mac to be awake and running, which limits its usefulness for always-on task delegation; the 27B model is still much smaller than frontier cloud models, and the founder’s comparison to Opus 4.6 is based on a specific benchmark, not a guarantee of general capability.

Who is it for?:Mac users (Apple Silicon) who want a private, local AI assistant for email, calendar, and meeting notes, and who are willing to join an invite-only beta; privacy-conscious users who refuse to send personal data to cloud AI services and prefer on-device processing; users who are comfortable with a free product backed by a payment-commission business model and are willing to let the assistant handle transactions. Not suited for Windows/Linux/Android users today (waitlist only), anyone needing a 24/7 always-on agent (it requires your device to be running), users who require independently validated model quality before trusting an assistant with personal data, or those uncomfortable with an unproven business model tied to payment transactions.