OpenHuman is the first agent harness packaged as a double-click desktop app, winning on accessibility rather than power.
Load tinyhumans.ai/OpenHuman and hit download. You get a .dmg, the standard Mac installer format, the same shape as downloading Slack or Notion. Drag OpenHuman into your Applications folder and open it. That is the entire install, maybe two steps, about 30 seconds.
Yes. You drop a meeting link in and it joins as a real participant: it hears everybody on the call, takes notes in the background, and can speak with a voice in real time. The docs only show Google Meet right now, and he is wary of using it in real meetings.
Simple routes model calls, web search and integration OAuth through their backend: one subscription, every model included, no API keys, token juice automatic, plus a dollar of free credit. Custom means your own OpenAI, Anthropic, search and Composio setup, with more setup and more to go wrong. He recommends simple.
Everything you connect, email, chats, documents, calendar, gets chopped into small markdown files, summarized by source, by topic, and one global summary per day. Go to Intelligence, the memory tab, click view folder, and Finder opens straight to your vault. You can edit the files or open the vault in Obsidian.
A compression layer baked into OpenHuman that shrinks what is sent to the model, so you pay far less per call. They claim up to 80% token savings, though reviews he has seen measure closer to 70%. It is why OpenHuman can ship one subscription instead of bring your own API keys.
Only partly. The memory of your data really does live on your machine, but the default install still routes model calls, web search and the OAuth handshakes through their hosted backend. Full self hosting is doable, swapping in Ollama for models, SearXNG for search and direct mode Composio, but it takes work.
It uses Composio underneath, so when you click connect Gmail you go through Composio's authentication flow. He calls that smart rather than a knock: they scale on day one instead of spending two years building 118 connectors themselves. Gmail, Slack, Notion, Drive, Linear and Calendar are all included.
Every five minutes the agent quietly wakes itself in the background, looks at the standing tasks you gave it and the state of your workspace, then picks one of three options: skip because nothing changed, act on its own, or escalate to you. Read only work runs silently, actions wait for approval.
On Linux the AppImage fails to launch on certain setups, so use the .deb package on Debian or Ubuntu. The MCP server is only partially built out, the OneDrive integration is shallow and not fully searchable, and the local model path through Ollama crashes on certain embedding requests.
His verdict is that OpenHuman invents none of the mechanics except the meeting feature. Markdown memory already existed in Hermes, integrations and channels in OpenClaw. The real innovation is packaging: the most opinionated, most beginner friendly harness today. Not the most powerful agent, but by far the most accessible.
“which honestly kills it for about 90% of people.”
“But OpenHuman, it is the first one in this category that literally anybody, even my mom, can actually run.”
“This thing can attend a meeting on your behalf.”
“I would be very weary of utilizing this inside of your meetings.”
“They say you cannot trust a memory that you cannot read.”
“what I think is the real innovation here is they just packaged the most opinionated, most beginner-friendly version of an agent harness that exists today.”
“I am Switching to OpenHuman....” https://www.youtube.com/watch?v=4xDTGazlYHM. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/i-am-switching-to-openhuman.html.