A complete, step-by-step guide to building a private AI agent with OpenClaw — on your own computer or a cloud server, with the AI model you choose. Written from a real project, with every command shown in full and explained.
This guide didn’t start as a guide. It started as Telêmaco, the personal AI agent our founder built for personal use — running 24/7 on a free cloud server, chatting on WhatsApp, and looking after two email accounts and a Google Drive.
Every step, every setting and every mistake of that project went into these pages. The guide is generic: you choose your agent’s name, personality, AI model and where it lives.
An AI agent isn’t a chatbot: it acts. By the end of the guide yours will be running 24/7 and doing this for you:
One message a day with what matters across all your inboxes, sorted by urgency.
You hear from it only when something really needs your attention.
Emails classified with your agent’s own labels, without touching yours.
Polite, short drafts waiting in Gmail for you to review and send yourself.
Folder plans, receipts and invoices saved from email to Drive — always with your approval.
Talk to it with voice notes, set reminders, keep a searchable log of decisions.
Email is the most dangerous thing you can give an AI: anyone in the world can write to you, and an email can hide instructions meant for your agent. So the guide doesn’t rely on asking the AI nicely. It builds layers:
The Google tool is compiled in a hardened version where sending, deleting, forwarding and sharing are blocked inside the program. No email or instruction can switch them back on.
A full admin version of the tool lives outside the agent, in a temporary container that only you run.
The rules file belongs to you. The agent can read it, not change it.
Email content is marked as untrusted, and the agent never follows instructions, links or requests found inside an email.
Moving or renaming files requires a plan and your explicit OK — and every change is logged so it can be undone.
The control panel only answers on the server itself. The only open door is SSH.
No setup is 100% safe, and the guide says so plainly: it explains the risk that remains and how to reduce it further.
Chapter 16 is a playbook of 18 problems we actually hit. A few of them:
A default setting woke the agent every 30 minutes to re-read the whole conversation. One line turns it off; normal use went back to about $1/day.
Adding a second AI provider silently switched on a mode that hid the terminal tool — and the scheduled email reports stopped working.
The server clock runs on UTC, so the agent got “today” and “tonight” wrong until the time zone was set.
A plain restart doesn’t re-read the secrets file. The guide shows the command that does.
Google access expired every 7 days while the app was in testing mode. The guide shows how to publish it.
A brand-new key, and every message failing: OpenClaw was using another provider’s default model.
Written so you can follow it even if you’ve never used a server before.
Appendices: final configuration files · command reference · privacy policy template for the Google consent screen · glossary · final checklist.
The software is free. The real cost is the AI model — and the guide shows how to keep it low.
Reference values from September 2026. Prices of AI providers and servers change; you set your own spending limit.
Edition 1.0 · September 2026 · Tested with OpenClaw 2026.9.6 and gog v0.42.0
Choose your edition:
Secure checkout by Stripe. No subscription.
Not available for purchase in Spain or Portugal.
By buying you accept the license terms.
No. You need to be comfortable opening a terminal and copying commands. Every command is shown in full, with what it does explained line by line.
The one you choose. The reference project uses Anthropic’s Claude, and the guide also covers OpenAI, Google Gemini, OpenRouter and local models. You can switch later with one command.
Both are covered: a home computer (Mac, Windows or Linux) or a cloud server (VPS), Intel/AMD or ARM. You can start at home and move later.
The guide uses WhatsApp, but OpenClaw also works with Telegram, Signal, Slack, Discord and others. The guide explains the WhatsApp terms-of-service risk and recommends a dedicated number or Telegram.
The full walkthrough is for Gmail, Google Drive and Google Calendar. Chapter 10 covers other providers (IMAP) with a safer, read-only approach.
Right after payment you’re taken to a download page with your PDF. Bookmark it; if you lose it, write to support@rmaincai.com with the email you used at checkout.
Yes. Each edition — English, Spanish and Portuguese — is sold separately; buy each one you want.
No. The guide is published independently by RMAINCAI and is not affiliated with OpenClaw, Anthropic, OpenAI, Google, Meta or any other company mentioned.
Up to 5 times within 30 days of your first visit to the download page. If you need it again after that, write to support@rmaincai.com.
No. The guide is licensed for the personal use of the buyer, and each copy carries the buyer’s email on every page. The commands and configuration examples can be used freely in your own installations.
Three to four hours of setup. An assistant that works for you every day after that.
You are solely responsible for the access you grant to your AI agent and for everything it does on your behalf. The guide is provided “as is”, without warranties. OpenClaw, Anthropic, Claude, OpenAI, Google, Gmail, WhatsApp, Oracle, Docker and other names are trademarks of their respective owners.
Chat with Ava — live demo
Ava is the AI receptionist of Demo Studio, a sample business. Ask about services and prices, or book a time — this is the real rmAgentBooking.
Your name and number are used only for this demo: nothing is sent to your WhatsApp, and the conversation is deleted after 7 days. Replies can take around 30 seconds.