# Build Your First AI Agent in 20 Minutes (No Coding)

[Watch on tuber](https://tuber.guzus.xyz/video/rCkYVxOX-dI/summary) · [YouTube](https://youtu.be/rCkYVxOX-dI)

Matt shows how to set up a free, open-source AI agent called **Hermes Agent** (made by Nous Research) in about 20 minutes without writing code. He runs it on a cheap Hostinger VPS using a one-click template, gives it a model through Nous Portal (DeepSeek Flash), connects it to Telegram, and gives it a first "watcher" job that runs on a schedule. His case for Hermes over ChatGPT or Claude is that you can pick any model, you own the software and the server, and it can message you in apps you already use.

- **0:02-1:14** — **Intro and the five-step plan.** Matt says his own agent runs more than 20 jobs for him, and that Hermes itself is free, so you only pay for a small server and model usage. The five steps are: choose the agent, host it on a server with one click, pick a model, connect your phone, and assign a first job. The steps are timestamped so viewers can come back to them later.

- **1:17-2:05** — **How an agent differs from a chatbot.** A chatbot answers your message and then waits for the next one. An agent adds three things: **tools** (web search, reading and writing files, running programs), **acting on its own toward a goal**, including repeating jobs on a schedule, and **memory** that improves over time. The AI model inside is the same kind that powers ChatGPT and Claude. What makes it an agent is everything built around the model.

- **2:06-3:48** — **Why Hermes instead of ChatGPT or Claude.** Matt admits ChatGPT and Claude can now also run scheduled tasks, remember you, and handle email, so if that's all you need, you already have it. He names three things they can't offer:
  - **Model choice:** Hermes works with any company's models, so you can put cheap models on routine jobs and strong ones on hard jobs in the same agent.
  - **Ownership:** it's free software on your own server, so no company can impose weekly limits, throttle it, or remove features.
  - **Where it lives:** it gets its own Telegram, WhatsApp, or Slack account instead of being stuck inside one vendor's app.

  There's one more difference: instead of separate projects that each keep their own context, Hermes is one agent that learns your whole world. It saves what it learns as reusable "skills" (step-by-step notes it can reuse), so it gets better at your specific work.

- **3:50-5:01** — **Why it needs a VPS.** The machine has to meet three conditions:
  - **Always on**, so scheduled jobs can run at 3 a.m. A laptop in a bag isn't on.
  - **Separate from your personal data.** The agent runs commands and installs or deletes files, so there should be no photos, passwords, or logged-in bank sessions on it. Matt calls it a "practice workshop" you can wipe and rebuild in 10 minutes.
  - **Reachable from any device.**

  You can install Hermes on your own computer, but it then only works while that computer is awake and runs next to your files. For a first agent, he recommends renting a server.

- **5:04-7:16** — **Choosing Hostinger and buying the plan.** Most VPS providers would make you install Hermes from a terminal. Hostinger has a one-click Hermes template that also sets up a secure web address, a login page, backups, one-click updates, and a restart button. Matt warns that interfaces change over time, so follow the general idea rather than exact screens. In checkout:
  - He recommends **KVM 1**, the smallest plan. It runs one agent comfortably and can be upgraded later without reinstalling.
  - Pick at least a **12-month term**, because the discount requires it. There's a 30-day money-back guarantee either way.
  - **Untick the pre-checked "ready-to-use AI" add-on**, because you're bringing your own model.

  His affiliate link adds an extra 10% discount.

- **7:18-9:29** — **Account setup, deployment, and first login.** Register with Google, GitHub, or email, then pay. Hostinger then opens the Hermes setup form with an admin username and password. Matt stresses saving the password in a password manager, not a plain text note. Click Deploy and wait until the app shows a green "running" status. Docker Manager lists two containers: Hermes and Traefik, which secures the connection. Click "open" under the Access column, bookmark the login page, and sign in. You never touch a terminal on this route.

- **9:31-10:33** — **Dashboard tour and the "no brain yet" error.** The menu includes:
  - **Sessions:** every conversation on every device, all sharing one memory.
  - **Models:** where you choose the AI that does the thinking.
  - **Cron:** scheduled jobs.
  - **Gateway:** the link to outside messaging apps.

  Typing "Hello" in chat returns "no inference provider configured." The agent is installed and running but can't think until it has a model provider.

- **10:34-14:36** — **Connecting Nous Portal and choosing DeepSeek Flash.** Nous Portal is made by the same company as Hermes and is built in, so you connect it by logging in rather than pasting API keys. One account gives access to more than 200 models plus tools: web search, reading web pages, image generation, browsing, and speech. The price is **$20/month for $22 of credits**, and his link takes $15 off the first month, so month one costs $5.
  - Connect it on the **Keys** page via "login," then "connect."
  - On **Models**, change the main model to **DeepSeek V4 Flash**, or whatever the newest Flash version is when you watch. He picks it because it uses tools well and costs a fraction of a cent per message, which matters while you're learning.

  After the switch, the agent answers "Hello." Asked what tools it has, it lists a browser, terminal commands, files, code execution, web search, sub-agents, memory, and scheduling.

- **14:38-18:07** — **Connecting Telegram.** Telegram is the easiest option: setup is a QR scan, and it handles files, voice notes, and photos. Hermes supports more than 20 platforms, including Discord, Slack, and WhatsApp. The steps:
  1. Go to **Channels**, turn Telegram on, and restart the gateway. If it doesn't come back after a few minutes, refresh the browser.
  2. Click **"Create with QR"**, scan the code with your phone, and create the bot in the Nabu Casa Hermes bot chat.
  3. Go back to that setup chat (rescan if it's missing) and tap **"Connect this bot."** Your Telegram ID shows up as the allowed user. Click **Save and restart**.
  4. Message your bot. When asked, run **/sethome** so Telegram becomes the "home channel" where cron results and notifications arrive.

- **18:08-18:28** — **Safety notes.** Only your Telegram account is on the allow list, so strangers who find the bot get no response. Telegram bot chats are **not end-to-end encrypted**, so never send it passwords or banking details.

- **18:31-20:36** — **First job: a watcher.** A watcher checks something on a schedule and only messages you when something changes. Using the Hermes desktop app (the same agent as Telegram), Matt asks it to watch Barcelona-to-London flight prices. He tells it to build a reusable skill for the check, keep track of what it has already shown him, run every few hours, and message him on Telegram only when there's something new. The setup took **30-45 minutes**. In that time the agent researched sources online, built the scraping setup, wrote a skill, created a cron job that runs every 6 hours, and set up memory to track what it had already reported. Later it sent a change alert on its own. His other jobs include:
  - a bookkeeping check every Monday
  - birthday tracking with gift research early enough to order
  - trip planning twice a day that collects quotes
  - follow-ups on a negotiation

  Caveat: some websites block agents. If that happens, ask the agent to find another source.

- **20:38-21:11** — **Costs and outro.** The server is a fixed monthly price. Model usage varies per message and is tracked on the Nous Portal usage page. On his first day, the model cost a little over **8 cents**. The rest is an outro that points to his longer Hermes guide and repeats the Hostinger discount link.

**Practical takeaways:**
1. **Run your first agent on a separate VPS, not your personal computer.** It runs commands and deletes files, so keep it away from your own data and accounts. It also has to be on 24/7 for scheduled jobs.
2. **Start with a cheap model that uses tools well, like DeepSeek Flash.** You'll send many test messages while learning, and you can assign stronger models to harder jobs later.
3. **Uncheck the pre-ticked "ready-to-use AI" add-on at Hostinger checkout,** and pick at least 12 months to get the discount.
4. **Run /sethome in Telegram** so scheduled job results actually reach you, and never send the bot sensitive data, because bot chats aren't end-to-end encrypted.
5. **Describe what you want once and let the agent build it.** Ask for a reusable skill, memory of what it has already reported, a schedule, and "only notify on change." Expect the first setup to take a while.

## Audience signal

- No high-signal audience themes available.