OpenHuman from TinyHumans AI is a desktop-first personal agent: connect Gmail, calendars, and GitHub, then fold context into a local Memory Tree and Obsidian-style Markdown vault on SQLite. Optional Ollama routes workloads on-device. Many search results still describe pip install and Python checkpoints—that is not the 2026 shipping path.
This guide targets beginners and developers who need a clean install on macOS, Windows, or Linux without compiling Rust yourself. You get a pain-point list, a host matrix, signed package installs, and a six-step path through first launch, integrations, Ollama, and Memory Tree validation. After reading, you should know whether your data directory belongs on a laptop or an always-on Mac Mini.
01 OpenHuman install pain points: not the git clone plus pip workflow
OpenHuman ships as a Rust + Tauri + React desktop app (GPL-3.0, Early Beta). Unlike OpenClaw (terminal Gateway and Node plugins) or Hermes Agent (installer drops Python 3.11 and Node 22 under ~/.hermes/), end users do not hand-build a Python venv. Contributing from source is a separate track with Node 24+, pnpm, and Rust toolchains.
If you already run OpenClaw in production, treat OpenHuman as a parallel experiment on a second user account or rental host before you move Telegram tokens or cron jobs. The products share the “personal agent” label but differ on memory: OpenHuman compresses connected sources into Markdown chunks on a twenty-minute loop, while OpenClaw often depends on plugins to ferry context in.
Privacy-minded teams should document what stays on disk (Memory Tree, Obsidian vault, SQLite) versus what still transits vendor APIs when you use hosted model routing. Local Ollama reduces outbound prompt volume for supported tasks, but OAuth tokens for Gmail or GitHub still require careful key storage and rotation policies.
- Stale tutorials: Generic digital-human posts mention
condaandcheckpoints/*.pth; those templates do not matchtinyhumansai/openhuman. - Script installs lack integrity checks: The README warns that
curl … | bashcannot detect tampered script bytes; prefer Homebrew bottles, signed apt, or MSI. - Local plus managed: Memory Tree and workspace stay on disk, but default sign-in, model routing, and some OAuth flows use OpenHuman-hosted services unless you switch BYO models and direct Composio mode.
- Sleep breaks sync: Roughly twenty-minute auto-fetch loops need an awake host; laptop suspend stalls context ingestion.
- Linux Wayland: AppImage may crash on some Wayland setups; Debian and Ubuntu users should prefer .deb / apt (see upstream issue #2463).
Install can finish in under twenty minutes; uptime and vault backups decide long-term value.
Commands follow TinyHumans official docs; re-open the links after each release.
02 OpenHuman requirements and host matrix: where to run production
Official artifacts cover macOS (Intel and Apple Silicon), Windows 10/11, and Linux (Debian/Ubuntu amd64, Arch AUR). The table below is for host selection; RAM and disk should follow your Ollama model size.
| Host | 24/7 uptime | Install friction | Memory Tree / Ollama | Typical gap |
|---|---|---|---|---|
| MacBook | Sleep breaks sync | Low (Homebrew / dmg) | Strong on Apple Silicon | OS updates, lid closed |
| Windows desktop | Can stay on, reboots | Low (signed MSI) | NVIDIA optional | No macOS-native tooling |
| Linux desktop | High on servers | Medium (apt over AppImage) | Ollama supported | Headless GUI care |
| Mac Mini M4 rental | Datacenter SLA | Low (SSH + brew) | Unified memory for 7B–14B | Pick region and term |
Resource hints with Ollama: Hosted models plus light integrations may start around 8 GB RAM; keeping a 7B-class model hot with browser tools wants 16 GB or more. Reserve 10–30 GB SSD for Memory Tree, vault, and SQLite growth.
If you sync Gmail or corporate calendars, decide early whether the host sits in your office or in a datacenter you control. A dedicated Mac mini with documented wipe at contract end is easier to explain to security reviewers than a developer laptop with personal iCloud enabled.
03 Official OpenHuman install paths: Homebrew, apt, and Releases
Preferred order: download from the website or GitHub Releases, or use the OS package manager with signing. Use curl/bash only when packages are unavailable.
brew tap tinyhumansai/core
brew install openhuman
openhuman --versionDebian and Ubuntu: add the signed apt repository per README, then sudo apt-get install -y openhuman. Windows: install the signed .msi from Releases.
Arch users can follow the in-repo AUR recipe (openhuman-bin) once published. Manual .dmg, .deb, .AppImage, and .msi artifacts are also listed on the latest release page if your distro lacks brew or apt.
The README labels curl/bash as unverified because script bytes are not separately signed today. Prefer package managers until the project ships GPG-verified install assets (tracked in upstream issue #2620). After any install path, launch the UI once before wiring Ollama so onboarding can create the local workspace layout.
04 Six steps: first launch, integrations, Ollama, Memory Tree check
- Install and launch: Open the app from the OS launcher or run
openhuman --versionif CLI is on PATH; match the tag on Releases. - Finish onboarding: Default managed mode needs an OpenHuman account for routing and some OAuth; local experiments use BYO / Ollama per the Local AI doc.
- Install Ollama (optional): Pull models documented for OpenHuman (for example Qwen2.5 or Gemma3 families—verify on gitbook), then select the local endpoint in model routing.
- Smoke-test integrations: Connect GitHub or calendar first; wait about one auto-fetch cycle (~20 minutes) and watch Memory Tree for new chunks.
- Validate Memory Tree and vault: Confirm SQLite and
.mdfiles grow; if empty, fix OAuth and timing before reinstalling. - Plan 24/7 hosting: Move to a Mac Mini with sleep disabled (local or CALMVPS bare metal) and tarball the workspace before migration.
When testing Ollama on Apple Silicon, pull arm64-tagged models and confirm the daemon is running before you select the local endpoint inside OpenHuman routing. On Windows with NVIDIA, verify driver versions against Ollama’s platform notes—OpenHuman itself does not replace GPU drivers.
For messaging channels (Telegram, WhatsApp, Discord per official docs), complete OAuth or bot setup only after Memory Tree shows healthy ingestion; otherwise the agent answers with stale context. TokenJuice compression applies to tool output before LLM calls—watch token bills on hosted models even when local inference is enabled for subtasks.
On CALMVPS bare metal, pick a tier on the pricing page, SSH in, run the same brew commands, and archive the OpenHuman data directory before changing hosts.
05 Citable specs, common errors, and CALMVPS wrap-up
- Stack: Rust + Tauri v2 + React shell; Memory Tree uses local SQLite with Markdown chunks (official README / gitbook).
- Sync cadence: Active integrations pull on roughly a 20-minute loop (not real-time push).
- TokenJuice: Tool and scrape payloads compress before LLM calls; docs cite up to ~80% token reduction depending on workload—measure your own bills.
| Symptom | Likely cause | Fix |
|---|---|---|
| pip/conda guide fails | Wrong project docs | Use Homebrew, apt, or Release installers |
| AppImage crash on Linux | Wayland / sharun | Use .deb or documented env workarounds |
| Ollama silent | Daemon or model missing | ollama serve and ollama pull |
| Empty Memory Tree | OAuth or fetch window | Re-auth and wait one fetch cycle |
FAQ quick answers: You do not need a discrete GPU for hosted routing. Apple Silicon Macs are the smoothest local-Ollama path in 2026 practice. Slow model downloads are a network problem—pre-pull on a datacenter Mac before you cut over DNS or SSH tunnels.
To uninstall, remove the package via brew/apt/MSI, delete the local workspace and vault directories you configured, and revoke OAuth tokens at each provider. Keep a tarball if you plan to reinstall on CALMVPS or another bare-metal host.
A sleeping laptop breaks auto-fetch continuity; a cheap offshore VPS lacks the official desktop experience; skipping backups loses context on migration. For signed installs, Ollama on Apple Silicon, twenty-minute memory sync, and fast RAM upgrades, CALMVPS bare-metal Mac Mini M4 rental is usually the better production default on the pricing page.