OpenHuman: Local-First Agent Tops Product Hunt at 11.6k Stars

For two years the AI agent pitch has been the same — a hosted box you can't see into, holding tokens to your Gmail, your Notion, your Stripe, your customers' files. The pitch was fine for solo founders shipping side projects. It was a non-starter for anyone who signs a BAA, a privilege agreement, or a 17a-4 retention policy.

OpenHuman shipped on May 15. By the end of the day it was #1 on Product Hunt with 11,600 GitHub stars in two weeks and a 150% week-over-week growth curve. The thing it sold was almost embarrassingly simple: the agent runs on your machine, the memory lives on your disk, and the 118 OAuth tokens it holds never leave your laptop.

What's actually in the box

Under the hood it's a local Python runtime, a persistent vector store for memory, and a one-click installer that handles the OAuth dance for the integrations most owner-operators actually use.

• Gmail, Calendar, Drive — the Google trio that runs most small businesses
• Notion, Linear, Jira — for teams that ship product
• Slack — meeting notes, channel summaries, draft replies
• Stripe — refund triage, dunning follow-up, churn analysis
• GitHub — issue triage, PR review, release notes

You can swap the model. Run it against Claude, GPT, or a local Llama if you're fully air-gapped. The repo ships with sensible defaults so you're not reading docs for a weekend to get a calendar summary.

Why this matters for the verticals everyone else ignores

I've spent the last six months talking to law firms, accounting practices, financial advisors, and a couple of small clinics about AI tooling. The conversation always ends the same way — "we love what it can do, but we can't send a client name to your API." Compliance officers say no. General counsels say no. Insurance carriers ask a question nobody can answer and then say no.

That's the wedge. Not "better than ChatGPT." Not "cheaper than Zapier." Just: it never leaves the building. For the buyers who care, that's the entire purchase decision.

The opinion

The hosted-agent companies will spend 2026 negotiating BAAs and SOC 2 reports they should have had two years ago, while OpenHuman and its forks quietly eat the regulated verticals. The moat in agent software was never going to be the model — every model is a commodity by Q3. The moat is going to be where the data sits when the agent is thinking.

For SMB operators evaluating tooling, the question to ask is no longer "is this AI any good." It's "where does my customer list end up after I click install." OpenHuman answers that question with one word, and the answer is good enough to sell. (If you want a structured way to compare your current stack on that dimension, our AI stack audit walks through it.)

What to actually do this week

• Clone the repo, point it at a junk Gmail, see how the memory layer behaves over three days
• If you sell to regulated verticals, draft one paragraph about local-first agent setup and put it in your next proposal
• If you're a hosted-agent vendor, start writing your data residency story now — your competitor just shipped theirs

"The agents everyone is paying for live in someone else's data center. The agents that actually get deployed in 2026 are going to live in yours."