Most AI assistants want to live in the cloud, where the computing power is abundant and the data flows freely. Pradeep Elankumaran and Kevin Li think that's precisely the problem.
Their new venture, Jo, pitches what might charitably be called a compromise—or what could turn out to be the architecture the privacy-conscious have been waiting for. The core conceit: sensitive information never leaves your Mac unless you greenlight it. Everything else? That runs on a dedicated server Jo provisions for you alone, with contractual guardrails meant to prevent your queries from becoming someone else's training data.
It's a split-the-difference approach, landing somewhere between Apple's device-first Intelligence framework and the cloud-native convenience of tools like ChatGPT or Perplexity. Whether that hybrid model represents the future or just another footnote in the AI arms race may hinge on a question of trust—specifically, how much users are willing to extend to Jo's infrastructure.
Two YC Veterans, Third Time's the Charm
Elankumaran and Li aren't newcomers to the startup grind. They've cycled through Y Combinator twice already. The first go was Kicksend, a file-sharing play from the Summer 2011 batch that eventually landed at Lyft. Then came Farmstead in Spring 2016, a grocery delivery service that tried to crack the fresh food logistics puzzle. Both founders logged time at Yahoo; Li also spent a stint at gaming company Kabam.
Now they're back in the YC fold with the Spring 2026 batch. The team is lean: two founders, possibly a handful more depending on how you count contractors and advisors. LinkedIn pegs the headcount somewhere between two and ten, which is typical for this stage but also reflects how much of Jo's promise rests on architectural choices rather than manpower.
One telling detail: Jo explicitly offers migration from OpenClaw, a local AI framework with its own niche following. That the founders bothered to build an import tool for "your OpenClaw folder and settings" suggests they've done the homework on where their early adopters might come from. It's the kind of move that signals focus.
What Jo Actually Does (and Where It Does It)
The product itself is a personal AI assistant that lives across your Mac, Telegram, and WhatsApp. A local model runs silently on your machine, observing browsing patterns and group chats, assembling what Jo calls "clips"—essentially contextual snapshots of your recent digital life. Planning a trip? Researching a project? When you eventually ask a question, Jo already has the thread.
Morning briefings pull in weather, calendar events, reminders. The assistant can search across messaging apps and email, draft responses (though it won't autonomously hit send—a small mercy, perhaps). It runs recurring automations in the background. The company's examples include a Thursday routine: "Send ideas for the weekend." Whether that's useful or mildly dystopian depends on your tolerance for automated nudges.
When a task demands heavier computing, Jo routes to reasoning models or spins up a browser instance with anti-fingerprinting measures to check flight prices or restaurant availability. The system auto-selects between fast and reasoning models. Simple queries, the company says, return in one to two seconds.
Users pick their poison for cloud inference: OpenAI, Anthropic, Grok, or Kimi. A local model handles the on-device processing. Integration spans Safari, Chrome, Notes, Slack, Messages, Photos, Reminders, and Finder on the local side, plus Gmail, Google Calendar, and Drive for cloud services.
It's an ambitious scope. Whether it feels coherent or sprawling in practice will depend on execution—something beta users will presumably help determine.
The Privacy Pitch, Unpacked
Here's where Jo plants its flag. Photos, local files, anything on your screen—all processed entirely on your Mac. No network calls. The local model handles transcription and compiles those contextual clips without phoning home. Your personal data, in theory, never escapes your hardware unless you explicitly authorize it.
For tasks requiring more muscle—inference, browser automation—each user gets a dedicated cloud machine. Jo asserts that cloud providers are contractually barred from training on user data, a claim that's become standard in enterprise AI contracts but still requires a degree of faith in enforcement. If you want to go fully local, the system supports that too, though some capabilities presumably fall away.
The company invites scrutiny. "Audit us," the website declares, pointing users to Activity Monitor, privacy documentation, and a security email address. The Terms of Service carry a recent update stamp—February 27, 2026, according to the site—which at least suggests the legal scaffolding is being actively maintained.
Still, contracts are only as good as the parties honoring them. And in a landscape where data provenance and model training practices remain murky even for major players, Jo's promises rest on trusting both the company's architecture and its vendor relationships. That's not nothing, but it's also not a cryptographic guarantee.
A Crowded Privacy Battlefield
Jo arrives at a moment when privacy has shifted from niche concern to competitive differentiator. Apple is reportedly building out Private Cloud Compute for its Intelligence suite, with architecture details that have surfaced in various tech circles. Perplexity launched what it called a "Personal Computer"—essentially a Mac mini running a persistent agent—earlier this year, generating buzz in Mac-focused communities.
Then there's the purely local crowd: Mochi, which markets itself as "100% Local · Always Private"; Jan, an open-source desktop app for local language models; Njn, leaning on Ollama for on-device inference. Each offers a different privacy calculus.
Fully local means maximum privacy but limited power—your laptop can only do so much. Fully cloud means access to frontier models but inherent trust issues. Jo is betting that the middle path—local for sensitive material, dedicated cloud for everything else—will feel like the right balance rather than an awkward straddle.
That's a gamble. Users who prioritize privacy above all might stick with fully local tools. Those chasing performance may not see enough reason to leave established cloud assistants. Jo needs to convince a segment that the hybrid model offers meaningful advantages on both fronts.
What's Next
Jo is currently in beta, free to use, no credit card required. The waitlist is open; the requirements are specific: Apple Silicon (M1 through M4) with 16GB of RAM. A native mobile app is "coming soon," whatever that means in startup timelines. For now, mobile access routes through Telegram and WhatsApp, which feels functional if not elegant.
One feature stands out, for better or worse: nightly self-improvement. Jo reviews its own notes overnight, attempting to get better at predicting what you'll need. How that works, what controls users have, and whether it feels helpful or invasive—those are questions that won't be answered until more people actually live with the product.
Jo is free during the beta, with no credit card required, and no public post-beta pricing has been disclosed. That's standard for this stage, though it also leaves open questions about how Jo plans to monetize. Dedicated cloud servers per user aren't cheap to operate. Whether the founders envision a subscription model, tiered pricing, or something more creative remains to be seen.
For now, Jo is a bet on architecture as much as features—a wager that in the ongoing privacy wars, users will value a thoughtfully designed middle ground. Whether that middle ground finds a market is the test that comes next.
