jo, the Private AI That Acts Before You Ask

The pitch for most AI assistants goes like this: open the app, type a question, wait for an answer, close the app. jo wants to delete three of those four steps. Built by a two-person team out of Y Combinator's Winter 2024 batch, jo is a voice-first personal AI that runs on your Mac and starts working before you say anything at all - summarizing what is on your screen, checking your calendar, and turning the loose ends of a busy day into something closer to done.

It is a small startup with a specific opinion. The founders, Pradeep Elankumaran and Kevin Li, have built consumer software together for more than a decade. Their argument with jo is not that the world needs another chatbot. It is that the assistant model itself has been passive for too long - and that the interesting product lives in the gap between what you meant to do and what you actually got around to.

What jo actually does

jo describes itself plainly: a personal AI agent for life, work, and family. In practice that means it watches the things that matter to you - email, calendar, documents, the pages you browse - and does something useful with them without a prompt. It gives a morning briefing. It flags the message you forgot to answer. It notices you have been reading about a topic and clips related material. Ask it out loud, and it will draft an email (it stops short of sending), check availability across your week, or run a pre-filled search.

The company frames this against two familiar fictional assistants. jo, it says, is neither JARVIS nor Samantha - both of which mostly waited to be spoken to. The founders call their approach "a new type of assistance for a new age," and the operative word is proactive.

Who it is for

jo's early users are not enterprise IT departments. They are individuals and families trying to keep work and home from colliding. The testimonials on the company's site read like the inside of a group chat: a parent in Chicago keeping track of "moving pieces across work and life," a family in San Francisco using it for "practical items that slip through cracks like email and scheduling," a Toronto user calling it "the personal AI I enjoy using most."

That framing matters, because it tells you what jo optimizes for. Not the impressive one-off demo, but the boring, recurring texture of a week - the double-booked afternoon, the reply that never went out, the thing you meant to look up on Tuesday and remembered on Friday.

The privacy argument, which is the whole product

Most companies treat privacy as a line in the settings menu. jo treats it as architecture. Each user gets an isolated cloud machine, hosted per person rather than pooled. Raw personal data stays on the Mac and is not moved to the cloud without approval. Only summarized data lives remotely. And the AI models jo relies on are, per the company, contractually prohibited from training on user data.

This is the kind of design choice that is easy to write in a paragraph and hard to build. It also happens to be the clearest way jo separates itself from the large, general-purpose assistants that run everything through a shared cloud. When the product is a system that reads your email and browsing, "where does the data live" stops being a compliance footnote and becomes the reason to trust it at all.

Built native, tuned for speed

jo launched its macOS app on December 23, 2024, and it runs only on Apple Silicon. That is a deliberate constraint. The app is written in native Swift and, by the team's account, spent weeks cutting latency across the audio, network, and model systems that have to cooperate for a voice assistant to feel instant instead of laggy. The company says summaries, browser tasks, and calendar checks all come back as fast as you can speak.

Under the hood, jo combines local and remote LLMs with vector storage and tight native desktop code - a stack the founders describe as a set of chainable primitives for how people interact with AI. There is also a self-improvement loop: jo reviews its own work each night and, the pitch goes, wakes up a little smarter about you.

Relative emphasis by surface, per public product descriptions. macOS is the flagship; messaging apps extend reach.

The people behind it

jo is the second Y Combinator company Elankumaran and Li have built together. Before this, they co-founded Farmstead (YC S16), a consumer grocery venture; Elankumaran's founder history stretches back further to Kicksend (YC S11). The through-line is consumer products - software ordinary people use every day - which explains why jo obsesses over setup, latency, and trust rather than benchmark scores. Li, who serves as CMO, lists judo, gymnastics rings, and video games among his interests, which is either irrelevant or exactly the sort of detail that tells you this is a small team, not a committee.

The business, and the crowd it is in

jo is a consumer product - direct to the person using it, distributed by instant download with no waitlist and no onboarding gauntlet. That is a bet on friction: get the app open, let it prove itself in a morning briefing, and skip the enterprise sales motion entirely. The company is backed by Y Combinator, and at two people it is deliberately small.

The neighborhood is crowded and getting more so. Apple is pushing Apple Intelligence and a reworked Siri into the same Macs jo runs on. OpenAI, Google, and a wave of "personal memory" startups are all circling the idea of an AI that knows your context. jo's answer to the obvious question - why would you use this instead of the assistant already built into your machine - is the combination the big platforms are slowest to offer: proactive by default, private by architecture, and native enough to feel fast.

Whether a two-person team can hold that ground against Apple is the open question. But the shape of the bet is clear, and unusually coherent for such an early company: an assistant that acts first, keeps your data close, and meets you on the surfaces you already use. In a category full of chatbots waiting to be asked, jo's contribution is that it does not wait.