OpenHuman for Beginners:
2026 Q&A From Install to Data Sources

If OAuth, BYOK, and Memory Tree are new to you, that is normal. Run the minimum path first; you can read deep docs after something actually works.

QWhat am I actually installing?

A: Not a website and not an SDK—you are installing a desktop app with a local agent. It uses OAuth (a standard “sign in with Google/GitHub” style permission flow) to connect Gmail, Notion, GitHub, and similar services, then writes what it learns into a local memory tree and a Markdown vault you can open in Obsidian. Memory stays on your machine, but chat, model routing, and some OAuth steps may still use OpenHuman’s managed cloud—local-first ≠ offline-only. Do this now: confirm you are on the tinyhumans build before downloading.

QWhat do I need before install? Which download?

A: A daily-driver computer, stable internet, a few GB of free disk, and comfort granting read scopes on services like Gmail. Beginners should start with the official DMG (macOS) or installer from Releases; power users can use Homebrew (brew install tinyhumansai/openhuman/openhuman) or the official install.sh—on a primary machine, skim the script or stick to the GUI installer. Do this now: install until you see the sign-in screen.

QWhy does the OS block the app?

A: macOS Gatekeeper and Windows SmartScreen often flag apps that are not from the App Store. Use “Open Anyway” only after downloading from the official site, and match the build to the version on GitHub Releases. That trade-off is normal for indie desktop software—and it is why you should never install from random mirror links. Do this now: pass the OS prompt and launch OpenHuman once.

QAfter sign-in, what else must I configure?

A: Account OAuth sign-in proves who you are and activates subscription features; it is separate from connecting Gmail. By default, chat and integration tool calls may still route through OpenHuman’s managed backend (including Composio for connectors). You still need: ① pick a model in Settings; ② open Integrations and Connect each service. Do this now: sign in → model → integrations—hold off on long chats until those are done.

QHow do I choose a model? Do I need my own API key?

A: Not required. BYOK (Bring Your Own Key) means you paste your own cloud provider key and pay usage yourself. You can also use subscription + TokenJuice (compresses bulky tool output before it hits the model, saving tokens). Ollama runs models locally—on Apple Silicon, 24GB+ unified memory is a practical floor for comfort. Never commit keys to Git. Send a one-line “hello” to confirm the model responds. Do this now: pick one billing path and verify the model works.

QWhy authorize integrations separately?

A: Signing into OpenHuman and connecting Gmail are two different OAuth grants with different scopes—you can revoke either later. The project documents 118+ connectors; auto-fetch (automatic background ingest into the memory tree on a ~20-minute loop) is strongest on services like Gmail, Notion, and Slack—check the current GitBook for your connector before assuming full sync. Connecting everything at once also increases cost and privacy surface. Do this now: connect Gmail plus one source you already use daily.

QWhen does memory show up? How do I know it read my test data?

A: Active integrations refresh on roughly a 20-minute cadence; the first cycle often means waiting the full window. Acceptance test: send yourself an email with a unique subject → wait 20–40 minutes → search that subject in the vault → ask the agent to quote a detail only in that message. If the vault has the file but the agent cannot cite it, the bottleneck is retrieval or model config—not OAuth. Do this now: run those four steps before expanding integrations.

QFirst task and what if something fails?

A: Start read-only: e.g. “Summarize the last three emails in Gmail about [project X].” If it fails, ask in layers: app version and restart → sign-in and model health → integration status (Expired?) → did you wait 20+ minutes and see new .md in the vault → wrong Google account connected? Then check GitHub Issues; separate subscription charges, BYOK usage, and local Ollama RAM/CPU. Do this now: pass one read-only Q&A before adding more connectors.

→Running OpenHuman 24/7 on a Mac mini

Memory trees and auto-fetch reward a machine that stays on quietly: Mac mini M4’s unified memory helps local Ollama, macOS SIP and FileVault fit personal context, and idle draw is often around ~4W. If you want an agent that keeps ingesting mail and docs while you sleep, Mac mini M4 is a strong value starting point—see whether the specs match your vault size and model choice.