In 2026, OpenHuman (TinyHumans, GPL-3.0) turns personal AI into a local-first desktop agent: a Rust + Tauri shell, SQLite Memory Tree, Obsidian-compatible knowledge vault, and 118+ one-click OAuth integrations for Gmail, Notion, GitHub, and more. Misled by outdated pip + digital-avatar tutorials, or stuck on signed installs, first sync, or Ollama local models? This README-aligned runbook gets you from download to a working agent, and compares MacBook / VPS / Mac Mini M4 monthly rental for 24/7 always-on hosting.
OpenHuman is not a Python project that uploads a video to generate a digital avatar. It is an always-on desktop agent harness: UI-first onboarding, account connections and Memory Tree build in minutes, background auto-fetch every ~20 minutes for email, calendar, repos, and more, then TokenJuice compression before model routing. Compared with terminal-first OpenClaw or Hermes, it emphasizes fewer API keys, more OAuth, memory in local SQLite + Markdown vault.
This article follows a full install outline: requirements, signed install, first verification, Ollama/integrations, then host sizing. If you already run the Hermes step-by-step install guide or OpenClaw production deployment, treat OpenHuman as a desktop layer with a face and memory tree, combined with a NodeMini remote Mac.
Pain point: Following old tutorials with pip install -r requirements.txt, which does not match the current desktop release.
Pain point: Installing via curl | bash with no way to verify script integrity (official docs flag the risk).
Pain point: Gmail/Notion connected but the agent still has no memory — first sync not finished or vault path wrong.
Pain point: AppImage crashes on Linux Wayland (community issue #2463).
Pain point: Laptop lid closed — auto-fetch and Meet proxy stop.
What this delivers: A six-step checklist from signed package install to Memory Tree verification.
OpenHuman is a native desktop app and does not require an NVIDIA GPU; the Ollama path is common on Apple Silicon. The table below combines the official README with Early Beta community feedback (not an SLA):
| Item | Minimum | Recommended (24/7 + many integrations) |
|---|---|---|
| OS | macOS 12+ / Win10+ / Ubuntu 20.04+ amd64 | macOS 14+ (Apple Silicon) or Ubuntu 24.04 LTS |
| RAM | 8 GB (cloud routing only) | 16–32 GB; 32 GB+ for local Ollama |
| Disk | ~2 GB app + cache | 50 GB SSD+ (Memory Tree + vault growth) |
| Network | Reach tinyhumans.ai and OAuth endpoints | Stable low latency (auto-fetch and Meet proxy) |
| Build from source | — | Node.js 24+, pnpm 10.10, Rust 1.93 (contributors only) |
"The hardest part of installing OpenHuman is often not clicking Install — it is picking a Mac that will not go offline at 2 a.m."
# Option A: official DMG (primary Mac; verify developer signature) # https://tinyhumans.ai/openhuman # Option B: Homebrew (bottle hash chain) brew tap tinyhumansai/core brew install openhuman
sudo apt-get install -y --no-install-recommends gnupg2 curl ca-certificates curl -fsSL https://tinyhumansai.github.io/openhuman/apt/KEY.gpg \ | sudo gpg --dearmor -o /etc/apt/keyrings/openhuman.gpg echo "deb [signed-by=/etc/apt/keyrings/openhuman.gpg arch=amd64] \ https://tinyhumansai.github.io/openhuman/apt stable main" \ | sudo tee /etc/apt/sources.list.d/openhuman.list sudo apt-get update && sudo apt-get install -y openhuman
Not the default path: curl … install.sh | bash has no independent signature verification. The official README recommends DMG / Homebrew / apt / signed MSI first. If you must use the script, wait for the GPG verification flow in #2620.
Download and install: Get the platform package from tinyhumans.ai/openhuman or GitHub Releases; on Windows use a signed .msi.
First launch: Complete account login (default uses OpenHuman-hosted model routing and Composio OAuth proxy; you can switch to BYO keys in settings).
Connect 1–2 integrations: Start with Gmail + GitHub or Notion and confirm the OAuth wizard succeeds.
Wait for the first auto-fetch: Roughly a 20-minute cycle; Memory Tree / Obsidian vault should show Markdown chunks of about 3k tokens or less.
Verify in chat: Ask to summarize this week's calendar and unread email highlights — answers that cite local context mean sync is working.
(Optional) Local models: Settings → Model Routing → Ollama; ensure ollama serve is running (see our Ollama on rented Mac guide).
| Error / symptom | Cause | Fix |
|---|---|---|
| AppImage crashes on launch | Wayland / sharun compatibility | Use the .deb package or documented env var workarounds |
| Integration connected but no context | First sync not complete | Wait for an auto-fetch cycle; check vault writes |
| OAuth keeps failing | Clock drift / proxy blocking | Sync NTP; allow tinyhumans.ai and Composio domains |
Local inference: OpenHuman can route some on-device work through Ollama so sensitive steps never leave your machine. Cross-agent memory: If you already run agentmemory, set memory.backend = "agentmemory" in config.toml to share storage with Cursor, Codex, and others (see the official GitBook).
To use OpenHuman as an always-online personal chief of staff, the host must handle background sync, voice, and Meet proxy. The table below is a sizing reference:
| Platform | 24/7 availability | Memory Tree compounding | Best for |
|---|---|---|---|
| Personal MacBook | Stops when lid closes | auto-fetch interrupted often | Trial / demo |
| Linux VPS without GUI | High uptime | No native desktop or Tauri UX | Not recommended as primary |
| Mac Mini M4 monthly rental | Often online in datacenter hosting | GUI + Ollama UMA + local vault disk | Many integrations + background agent |
Remote Mac tip: After renting a dedicated Mac Mini, use VNC or screen sharing for first-time OAuth; day to day you can maintain Ollama and disk over SSH. Install steps are the same as on a local machine.
.md vault, roughly 3k tokens per chunk.For a weekend desktop mascot trial, a local install is enough. Once Gmail, calendar, and repo context must compound automatically every day, lid sleep and home network drops turn a five-minute onboarding into weeks of re-feeding data. Rent for 30 days to validate your workflow before buying a Mac Mini — lowest decision cost.
Low-spec VPS without GUI, noisy neighbors, and Raspberry Pi disk IO are all poor primary homes for OpenHuman — users only see an agent that forgot who they are. For production that needs a stable 24/7 desktop agent, local vault, and Ollama unified memory, NodeMini Mac Mini M4 cloud rental is usually less painful than a sleeping laptop and manual OAuth restarts: you focus on memory compounding, not fixing auth at midnight.
No. TinyHumans OpenHuman is a Rust/Tauri desktop agent installed via DMG, Homebrew, or apt — not git clone + pip install to train avatar weights. Use github.com/tinyhumansai/openhuman.
NodeMini offers dedicated Mac Mini M4 monthly rental. Pricing is on the rental rates page. OpenHuman subscription and third-party LLM fees are still paid by you to TinyHumans and model providers.
Yes. Default cloud model routing works out of the box; the local path can use Ollama with Apple Silicon Metal. For SSH and remote desktop questions, see the Help Center.