Start with one useful assistant
Updated October 30, 2026. Educational guidance; use current official documentation and your installed version.
You do not need a server rack, 16 agents, or a dashboard. Start with one assistant that reliably performs one task you care about.
1. Choose where it will run, then install
First read Native, VM, or Docker?. For your first lab, use sample data and a restricted workspace without connecting personal mail, browser profiles, calendars, or production data accounts. Model-provider authentication may still be needed. A VM or deliberately restricted container can reduce exposure; native installation is an explicit option for native-app workflows, not blanket permission to access your normal account.
Follow OpenClaw installation for your operating system. Review downloaded installers before running them. Use the supported runtime version listed there, rather than a version copied from an old presentation.
Once installed, inspect the supported setup commands:
openclaw --version
openclaw onboard --help
openclaw onboard
onboard is the current guided entry point. The commands above were checked against Adam's installed OpenClaw 2026.10.9 CLI on October 30, 2026; this page is not a claim that a fresh installation on your device was tested. Do not use the older site's openclaw init or hand-written agent YAML as current installation instructions.
2. Choose a supported model and sign in
Use the provider choices offered by your installed setup wizard. Subscription authentication, API keys, and local-model endpoints are different routes with different permissions and limits. Do not assume a chat subscription pays for arbitrary API requests. Verify eligibility and terms with the provider; do not configure a paid fallback unless you intend to use it.
Adam currently uses OpenAI-backed Codex for normal work and Claude Code selectively for deeper coding and Riley's communication drafting. See model routing. You can use a different supported provider. Model names and prices change; this guide deliberately avoids a hard-coded model shopping list.
3. Define the assistant's job
Use the onboarding interview to supply your goals, preferred communication style, and boundaries. Keep credentials in the product's supported secret/authentication storage, not in personality or memory files.
Start with this task:
Help me prepare tomorrow's priorities from the information I provide. Separate confirmed commitments from suggestions. Produce a short private draft; do not message anyone or change my calendar.
Verify the result against your actual source. A confident answer is not evidence that a tool was connected.
A low-friction way to ask
Try dictating into the prompt field, then stop and edit before submitting. Windows uses Windows + H; on Mac use your configured Dictation key or shortcut. Voice in, voices out explains the shortcuts, cleanup habit, and distinct specialist voices. No custom voice service is needed to start.
4. Connect one tool
Add a single read-only integration you are allowed to use. Read a known item and verify its account, date, and contents. An account showing cached data is not proof that its provider is synchronizing. See integrations.
5. Automate one check
Ask the assistant to propose a bounded schedule, timezone, input scope, output destination, and stop condition. First run it manually. Verify an unchanged run stays quiet, a changed run produces a useful result, and an access error is reported as unknown rather than zero.
Prefer deterministic change detection before model reasoning. See proactive workflows and cost controls.
6. Add specialists only when useful
Use Paperclip and agent teams when you need persistent ownership, task history, or a handoff that survives a chat. Mission Control is optional. Neither is required to get value from a personal assistant.
First success: one verified task, one connected source, and one automation you can inspect and turn off.