6.3 KiB
description, icon
| description | icon |
|---|---|
| Go from a fresh install to a working personal assistant that knows your context, respects your boundaries, and acts only with your approval. | robot |
Create my personal AI assistant
Goal: a working assistant that has some memory of your world, replies in a style you like, and never takes a real-world action without your say-so.
This is the "start here" guide. It assumes nothing beyond a downloaded app.
Prerequisites
- OpenHuman installed on macOS, Windows, or Linux. If you haven't installed yet, do Getting Started first, then come back.
- 4 GB+ RAM (16 GB+ if you plan to connect very large mailboxes or run a local model).
- An account to sign in with (social login works).
- Optional: one account you'd like the assistant to know about (Gmail is the usual first one).
Privacy implications
- Signing in does not grant ongoing access to anything. Every integration is a separate, explicit OAuth approval you can revoke later.
- Your memory (the local database and the Markdown vault) is created on your machine. Raw source data does not sit on the OpenHuman backend.
- By default, chat/reasoning runs through the OpenHuman-hosted model router. If you want inference on-device instead, see Use OpenHuman with a local model.
- Full detail: Keep sensitive data private.
Steps
1. Sign in
Launch the app. The first screen is "Sign in! Let's Cook". Choose a login option. There is an Advanced panel for pointing at a custom core; most people ignore it.
2. Choose how AI runs
After sign-in you'll hit a runtime choice:
- Cloud: one click, and the hosted model router handles inference. This is the fastest path to a working assistant.
- Custom: walk through choosing your inference provider, voice, integrations (OAuth), web search, and embeddings yourself.
If you're not sure, pick Cloud. You can change any of this later in Settings.
3. Give it something to remember
An assistant with no memory is just a chatbot. Connect at least one source so it has context to draw on:
- Open Settings and connect an integration (Gmail is the common starting point). Each connection is a one-click OAuth approval.
- Once connected, auto-fetch starts pulling data into your Memory Tree on a schedule (the first Gmail tick lands within about twenty minutes).
4. Set your boundaries
Open Settings → Agents → Agent access. This controls how much the assistant can do on its own:
| Tier | What it means |
|---|---|
| Read-only | The assistant can observe and answer, but never acts (no sending, no file writes, no commands). |
| Supervised (default) | It can act, but any state-changing, network, install, or destructive action is parked for your approval first. |
| Full | Routine actions run automatically; network/install/destructive actions still ask. |
Leave it on Supervised unless you have a reason not to. Nothing with an external effect will happen in a chat without you saying yes. This is the Approval Gate, and it's on by default.
5. Shape its personality (optional)
How the assistant talks and behaves is defined by an editable prompt called SOUL.md (its mission and values live in a companion IDENTITY.md). You don't have to touch these (they ship with sensible defaults), but you can:
- Set a display name and short description in Settings → Personality.
- Edit the behavior itself via the Brain page (the raised center button in the bottom bar,
/brain), where memory, goals, and intelligence live.
OpenHuman also learns durable preferences from how you correct it over time. See Personalization & Self-Learning.
6. Run your first real request
Once a source has been ingested, try:
- "What do I need to know from the last 12 hours?"
- "What's waiting on me?"
- "Summarize what I missed today."
Success checks
You have a working assistant when all of these are true:
- The app is signed in (you're past the welcome screen and in the chat/home view).
- At least one integration shows as connected in Settings.
- A briefing prompt ("what's waiting on me?") returns something drawn from your actual data, not a generic answer.
- Opening the Memory tab shows summaries appearing (give it one auto-fetch cycle, up to about 20 minutes, after connecting a source).
- When you ask it to do something with an external effect (e.g. "draft and send an email"), you see an Approval Request card appear above the chat box rather than it silently acting.
Common failures
| Symptom | What it means | Fix |
|---|---|---|
| Sign-in returns to the welcome screen | The OAuth callback didn't reach the app | Follow Troubleshooting Sign-In |
| Connected a source but memory stays empty | First auto-fetch tick hasn't run yet, or the OAuth scope is too narrow | Wait one cycle (~20 min); re-check the connection in Settings |
| Assistant answers generically, ignores your data | It answered without recalling memory | Ask again and reference the source explicitly ("from my email…"); confirm the source is connected |
| It performed an action you didn't expect | Autonomy tier may be set to Full | Set Settings → Agents → Agent access back to Supervised |
Recovery
- Reset boundaries fast: if the assistant is doing too much, drop the tier to Read-only in Agent access. It takes effect on the next turn and blocks all acting immediately.
- Nothing you connect is permanent: revoke any integration from Settings; chunks already in your local memory stay (they're yours), and the next sync tick stops pulling that source.
- If the app itself won't start, see Recover from a failed installation. Your configuration is preserved by default.
Next steps
- Use OpenHuman with a local model: keep inference on-device.
- Connect OpenHuman to Obsidian: read and edit the memory by hand.
- Keep sensitive data private: understand exactly what leaves your machine.