Free knowledge sharing on all things Agentic AI

How to Claw?

A recipe-style guide to getting your own AI agent — from zero to a self-hosted assistant you can reach via email and Telegram.

Prep time Less than an hour
One-time cost ~€10 (VPS setup + first month, domain registration)
Going cost ~€25/month (VPS ~€4 + Ollama Cloud $20)

Ingredients

This guide walks you through every step of setting up a personal AI agent on a cheap VPS. By the end you will have an autonomous agent that reads its own email, chats with you on Telegram, and serves its own website. No prior server experience required — just follow the recipe.

1 Rent a VPS from OVH

Head to OVH and order their smallest tier: VPS-1. It costs a few euros a month and is more than enough for running an AI agent. Pick a data centre close to you and choose Ubuntu 24.04 as the operating system.

2 Register a domain and set up email

While you are in the OVH dashboard, register a domain name. OVH includes a free Zimbra email account with every domain. Setting it up is not entirely obvious — you need to go through three steps in the right order:

2a. Create an Organization

  1. In the OVH Control Panel, navigate to Web Cloud → Zimbra Mail.
  2. Click the Organization tab, then Add Organization.
  3. Give it a name (e.g. your project name) and a label, then confirm.

2b. Add your Domain

  1. Click the Domain tab, then Add a domain.
  2. Select the organization you just created.
  3. Pick your domain from the list and choose Recommended configuration — this automatically sets up MX, SPF, DKIM, and SRV records.
  4. After adding, click the three-dot menu next to the domain and select Diagnostics to verify all DNS records are green.

2c. Create the Email Account

  1. Click the Email accounts tab, then Create an account.
  2. Choose the address (e.g. agent@yourdomain.com), fill in the display name, and set a strong password.
  3. Note down the IMAP/SMTP details — you will need them later:
    • IMAP: ssl0.ovh.net, port 993 (SSL)
    • SMTP: ssl0.ovh.net, port 465 (SSL)
Tip: The full OVH Zimbra setup guide is available at help.ovhcloud.com.

3 Choose a model host

Your agent needs a brain. Go with Ollama Cloud for unlimited cloud model hosting at $20/month. The recommended model is glm5.1 — it is capable, fast, and well-suited for agentic workloads.

Tip: The Pro offer on Ollama Cloud is more than enough for heavy — but not crazy heavy — usage. You do not need the Enterprise tier unless you are running dozens of agents or serving a large team.

4 Install an agent framework on the VPS

Hermes is a mature, actively developed agent framework that makes wiring everything together straightforward. It is the simpler alternative to rolling your own agent stack and a solid choice for a first deployment. SSH into your VPS and follow the official Hermes setup instructions. Once Hermes is running, configure it to talk to your Ollama Cloud endpoint and point it at the glm5.1 model.

Note: Both Hermes and OpenClaw are completely free. You can comfortably start with one and learn there — switching to the other once you are a practitioner is straightforward since the concepts transfer directly. If you want to explore OpenClaw, their install guide will get you up and running just as quickly.

5 Set up Telegram first

Telegram is the easiest messaging channel to wire up and the one you should get working first. Even if you have never used Telegram before, it is worth installing the app on your phone just for this purpose — the bot API is simple, free, and has no approval process.

  1. Open Telegram and talk to @BotFather to create a new bot. You will receive an API token.
  2. Add the token to your Hermes configuration so the agent can send and receive Telegram messages.
  3. Send your bot a test message and confirm it responds.
Tip: You can move to WhatsApp at a later date once the core setup is solid. Telegram is a better starting point because it has zero friction for bot development.

6 Get the agent settled in

Now comes the fun part — you tell the agent to set itself up on the VPS. In theory you could give it all three tasks in a single prompt, but it is better to start practising with small, focused work items. This builds your intuition for how the agent handles instructions, and makes it easy to catch and correct mistakes early.

Give the agent these three prompts one at a time, waiting for each to complete before moving on. Substitute your own domain, email credentials, and IP address.

Prompt 1: Get email operational

Configure email access on this machine.

- IMAP server: ssl0.ovh.net, port 993 (SSL)
- SMTP server: ssl0.ovh.net, port 465 (SSL)
- Username: agent@yourdomain.com
- Password: (your email password)

Connect to the mailbox and verify you can list folders.
Report back when done.

Prompt 2: Test email

Send a test email from agent@yourdomain.com to my-personal@email.com.

- Subject: "Agent online"
- Body: "Email is working. Ready for further instructions."

Confirm the email was sent and delivered successfully.

Prompt 3: Set up a website with HTTPS

Set up a public website on this machine.

- Install nginx.
- Obtain a Let's Encrypt certificate using certbot,
  with agent@yourdomain.com as the contact email.
- Configure nginx to serve a simple status page
  at https://yourdomain.com.
- Verify the site is reachable and the certificate is valid.

Report back when the site is live.

The agent will install packages, configure services, and report back at each step. Once all three prompts are done, your agent is live.

What you end up with

An AI agent running on your own VPS that can be reached via email and Telegram, and that manages its own website with automatic HTTPS. From here you can layer on more capabilities — calendar access, file storage, API integrations — one prompt at a time.

Bonus: use Claude’s frontend skill

When it comes to making your agent’s website look good, the Claude Code frontend skill is a huge help. It lets you describe a page in plain language and get well-structured, production-ready HTML and CSS back. If you want your agent’s status page or dashboard to look polished without hand-writing every line of markup, this is the tool to reach for. You can read more about Claude Code skills in the official documentation.