Desktop one-click install · OAuth integrations · Six-step runbook · Local vs hosted boundary · 24/7 host selection
Trying to get OpenHuman (tinyhumansai/openhuman, ~29K GitHub stars) running in 2026 but unsure whether to pip install a model or grab a DMG? This guide walks the full path: Rust + Tauri desktop, Memory Tree local SQLite, 118+ OAuth integrations, and ~20 minute auto-fetch cycles. You get an environment checklist, a six-step runbook, first-request acceptance, an error table, and 24/7 host guidance—focused on OpenHuman alone (see the OpenClaw dual-framework guide for side-by-side deployment).
OpenHuman (TinyHumans, GPL-3.0) is a local-first desktop AI super-assistant: Memory Tree normalizes Gmail, GitHub, Notion, Slack, and other sources into Markdown chunks stored in on-device SQLite—not an opaque vector store. It ships automatic model routing (frontier models for reasoning, cheaper models for light tasks), web search, a Coder toolchain, voice, and a desktop mascot that can join Google Meet. Unlike ChatGPT in a browser, multi-week context stays on your disk, and integrations connect via one-click OAuth.
Five install pitfalls: treating OpenHuman as a CUDA digital-human project, ignoring the local vs hosted boundary, asking questions before auto-fetch finishes, expecting 24/7 sync on a closed MacBook, and building from source without Node 24 / Rust 1.93. Below covers official desktop install + first successful run only; for OpenClaw on the same machine, see the May dual-framework article.
| Environment | Minimum | Recommended (large inbox + local models) |
|---|---|---|
| OS | macOS / Windows 10+ / Linux desktop | macOS 14+ Apple Silicon or Ubuntu 24.04 LTS |
| RAM | 4 GB+ | 16 GB+ (official docs: large mailboxes/repos or local models) |
| Disk | ~2 GB app + growing SQLite | 50 GB+ SSD for Memory Tree and logs |
| GPU | Not required (default cloud routing) | Apple Silicon UMA when running local models |
| Network | Internet for login and OAuth | Stable link for Meet and auto-fetch |
Wrong pip/CUDA path: End users install via official packages or curl scripts from tinyhumans.ai/openhuman—not pip install -r requirements.txt on a digital-human repo.
Ignoring data boundaries: Memory Tree lives locally, but default login, some OAuth, and search proxy still use hosted services—Advanced settings can point RPC and models elsewhere.
Asking too soon: After connecting Gmail/GitHub, wait ~20 minutes for auto-fetch to populate the tree before prompts like "summarize last week's email."
Laptop as 24/7 server: Closing the lid breaks sync; Memory Tree compounding needs an always-on desktop host.
Source build missing deps: Contributors need Git, Node.js 24+, pnpm 10.10, Rust 1.93 (with rustfmt/clippy), CMake, and submodules.
Same OpenHuman binary—different channel and host change signed updates, uptime, and suitability for background auto-fetch. The table below targets production setups where Memory Tree should grow daily.
| Channel / host | Best for | 24/7 auto-fetch | Notes |
|---|---|---|---|
| Official DMG / MSI | macOS / Windows end users | Requires machine stays awake | Download from tinyhumans.ai/openhuman |
| curl / Homebrew / apt | Terminal-friendly macOS / Linux | Same as above | Scripts and package managers track README |
| Personal MacBook | Trial, light integrations | Stops when lid closes | POC only—not production memory compounding |
| Mac Mini M4 rental | Shared Memory Tree, Meet proxy | Datacenter SLA uptime | SSH delivery; pairs with Ollama/OpenClaw (see May guide) |
Install takes minutes; letting Memory Tree auto-fetch every ~20 minutes requires the right 24/7 host.
Official one-liners cover macOS/Linux and Windows PowerShell; Linux can add a signed apt source. If you plan to run OpenClaw Gateway and OpenHuman desktop on the same leased node, pick an M4 with 16 GB+ UMA and reserve disk quota for SQLite and logs.
Steps below target end-user desktop install (not source contribution). Data defaults to the local app data directory (SQLite Memory Tree and Markdown vault)—back up before major upgrades.
Download and install: Open tinyhumans.ai/openhuman for DMG/MSI, or run the official install script (macOS/Linux x64).
First launch and login: Click Sign in! Let's Cook and sign in with a social account; most users need not change Advanced RPC URL.
Connect first source: OAuth-link Gmail or GitHub on the onboarding screen (pick any of 118+ integrations) and confirm authorization succeeds.
Wait for auto-fetch: Keep the app foreground or host awake ~20 minutes so the scheduler ingests data into Memory Tree.
First-request acceptance: Ask "summarize my open PRs from last week on GitHub" or "group recent email by topic"—answers should cite your data, not generic chat.
Optional advanced: Review model routing, enable local models on high-RAM hardware, or add custom Composio / search credentials to reduce hosted dependency.
curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bash brew tap tinyhumansai/core brew install openhuman
git clone https://github.com/tinyhumansai/openhuman.git cd openhuman git submodule update --init --recursive pnpm install pnpm --filter openhuman-app dev:app
Local vs hosted: Memory Tree, Markdown vault, and workspace config stay on device; default login, model routing, some integration OAuth, and search proxy still use OpenHuman hosted services. For fully owned models and credentials, follow the custom setup path in docs.
These symptoms usually appear when the app is open but Memory Tree looks empty—work the table row by row instead of reinstalling, switching accounts, and wiping the database at once.
| Symptom | Likely cause | Fix |
|---|---|---|
| Cannot connect Gmail/GitHub after login | Network or blocked OAuth callback | Check proxy/firewall; retry OAuth; allow callback domains in enterprise policy |
| Answers ignore my data | Auto-fetch not finished | Keep app running ~20 minutes; confirm integration shows Active |
| Linux install script fails | Non-x64 or missing libs | Use official packages; see GitBook Getting Started platform notes |
| pnpm dev:app build fails | Node/Rust version mismatch | Align Node 24+, pnpm 10.10, Rust 1.93, and submodules |
| Local model OOM | Insufficient RAM | Disable local model and use hosted routing, or move to 16 GB+ host |
Large mailboxes or monorepo ingest benefit from disk IO and 16 GB RAM; low-memory hosts can reduce concurrent sources temporarily. Voice, Meet proxy, and Coder tools ship built-in—no extra "read file" plugins. When co-hosting with automation stacks (e.g. OpenClaw Gateway), snapshot-backup SQLite and log directories so concurrent cleanup does not delete the vault.
tinyhumansai/openhuman ~29K stars (June 2026), Rust + Tauri desktop stack.Use cases: Solo devs link GitHub + Notion for auto weekly reports; ops links Gmail + Calendar for meeting summaries; teams install OpenHuman on a rented Mac Mini and maintain via VNC/SSH while laptops stay closed. Laptops and home PCs pass install acceptance yet break Memory Tree compounding through sleep and cross-region IO; headless Linux VPS lacks the full official desktop experience. To treat OpenHuman like a colleague who remembers last week's mail, Mac Mini M4 rental bundles uptime and UMA into predictable OpEx.
Finished this runbook and need a 24/7 node? VpsMesh offers Mac Mini M4 cloud rental: SSH delivery, resource panel, and app-data backup before off-board. See Mac Mini M4 rental pricing, deployment help at the help center, and order at the order page.
Warning: Do not change host, wipe Memory Tree, upgrade a major version, and rewire every OAuth the same weekend—four parallel moves cannot be bisected. Stabilize first-request behavior for 24h before touching local models or Meet proxy.
No. End users install via official packages, curl | bash, Homebrew, or apt. Only source builds need Node 24+, pnpm 10.10, Rust 1.93, and related dev tooling—not a CUDA digital-human training stack.
Memory Tree database, Markdown vault, and workspace config default to local SQLite. Default settings still use OpenHuman hosted login, model routing, some OAuth, and search proxy; Advanced / custom setup can attach your own models and Composio credentials to tighten boundaries.
OpenHuman is a desktop app with the fullest GUI on macOS/Windows/Linux. For always-on auto-fetch and Meet proxy, Mac Mini M4 rental beats a closed MacBook; it can co-host OpenClaw (see the dual-framework guide). Plans on the pricing page, order at the order page.