How I Finally Got a Personal AI Agent Running (After Burning $800 Learning What Not to Do)

I came across this article on X and thought it was too good not to share here. If you've been curious about running your own AI agent, this is the guide I wish I'd had.

---

There is a version of this story where I spend a weekend, follow a guide, and have a working AI agent by Sunday evening. That is not my story.

My version involved AWS servers I did not need, API keys for the wrong services, three different AI models that each cost a fortune per call, and at one point — I am not exaggerating — eight separate Telegram bots running simultaneously, each with its own "brain," each confidently unaware the others existed. The context evaporated constantly. The costs did not.

By the time I found the setup that actually worked, I had spent roughly $800 on Anthropic API tokens alone, not counting everything else I tried along the way.

What I have now is simpler than anything I attempted during that period. A Mac Mini next to my TV. One agent. Always on. Reachable from my phone. No usage meters quietly burning through a credit card in the background.

This guide is the one I wish someone had handed me at the start.

---

The Most Expensive Mistake You Can Make (And How to Avoid It)

Before anything else: do not sign up for the Anthropic API console and start paying per token. This is the single biggest trap, and it is very easy to fall into because the console looks professional and the onboarding is smooth.

The problem is that AI agents are not like web apps. They call models constantly — for memory lookups, for reasoning, for generating responses, for tool calls. When you are paying per token, those calls add up faster than you expect, especially when you are still figuring things out and making mistakes.

Claude Pro ($20/month) or Claude Max ($90/month) gives you a flat monthly rate with far more runway than pay-as-you-go. The token you need comes from your Claude account, not the API console — and that distinction matters a lot. I will show you exactly where to find it during setup.

The flat-rate membership is the foundation everything else is built on.

---

Before You Touch Terminal: Get Organised

The single best thing you can do before starting is to have everything ready in one place. Open your notes app and collect the following four things. You will need all of them during the OpenClaw onboarding process, and hunting for them mid-flow is unnecessarily frustrating.

Your Claude token — from your Claude account (details below on how to get this)

A Brave Search API key — free tier available at brave.com/search/api, takes about two minutes to get

A Groq API key — free, gives your agent the ability to transcribe voice messages so you can speak to it rather than type. Worth noting: if you are reading this after mid-2026, Telegram voice support may be built directly into OpenClaw already, which could make Groq optional

A Telegram bot token — created through @BotFather inside Telegram, also free and takes about two minutes

Four things. Have them all before you begin.

---

The Hardware Question

You do not need anything impressive. The minimum that works is modest: 2GB RAM, a couple of CPU cores, 20GB of storage. Any machine that can stay on reliably.

An old laptop works. A spare PC works. Anything that does not need to be shut down regularly.

That said, if you are in a position to pick something purpose-built for this, an M4 Mac Mini is hard to argue with. Quiet, efficient, small enough to sit next to a television and be completely ignored. Mine runs 24/7 without a monitor or keyboard attached. It just sits there and handles things. Worth considering if you want a permanent home for your agent rather than a laptop you occasionally close.

---

The Setup, Step by Step

Getting Your Claude Token

Install Claude on your Mac by opening Terminal — find it in Applications → Utilities — and running:

curl -fsSL https://claude.ai/install.sh | bash

Follow the prompts to log in. Once it finishes, close Terminal completely and open a fresh window. Then run:

claude setup-token

Your token will appear on screen. Copy it into your notes immediately. One important detail: this token is frequently copied with a trailing space, which causes authentication failures that are genuinely annoying to diagnose. Paste it into your notes, then manually delete any space at the end before saving.

Installing the Prerequisites

In a new Terminal window:

brew install node

If Homebrew is not yet installed on your machine, visit brew.sh first and follow their one-line installation. Then come back and run the above.

Installing OpenClaw

npm install -g openclaw

That is the entire installation. One command.

Running Onboarding

Launch OpenClaw. You will be presented with a terminal interface — navigation works as follows: spacebar to select, enter to confirm, arrow keys to move around.

Accept the disclaimer, select Quick Start, and the onboarding process begins. It walks through everything in sequence.

When it asks which AI provider to use: select Anthropic. A secondary menu will appear with three options — choose Anthropic Token, not Anthropic API. Paste in the token from your notes. Double-check there is no trailing space before confirming.

Connecting Telegram

The communications section will present a list of options. Select Telegram.

If you need to create a bot token at this point: open Telegram, find @BotFather, and send it /newbot. Choose a display name, then a username — the username must end with the word "bot." BotFather will issue you a token when you are done. Paste it into the OpenClaw prompt.

BotFather is well-designed and walks you through this naturally; it tells you what it needs at each step.

Skills

You will be prompted to configure skills. Choose npm as your package manager. When the full skills list appears, skip it for now — none of it is needed to get started, and it is easier to add things later than to wade through options when you are still getting oriented. Same for the additional installs that appear afterward — ElevenLabs, Google Places, and the rest can all wait.

When the hooks prompt appears, select all three. These are required for the agent to function correctly.

Hatching

After the hooks step, your gateway starts up and the final option appears: Hatch in TUI. Select it.

This wakes the agent up for the first time.

If an HTTPS error appears at this point, the cause is almost always a space in your Claude token. Fix it, re-run onboarding, and it will clear.

The First Conversation

Your agent will introduce itself and ask for its name and yours. Introduce yourself, give it a name, and then move the conversation to Telegram:

Open the Telegram bot conversation (via the link BotFather provided), enter the pairing key when prompted, and the connection establishes. Your agent replies through Telegram.

That moment — when it responds on your phone for the first time — is the one that makes the whole thing click. Keep it.

---

After You're Running: The Final Additions

Once you are chatting with your agent on Telegram, go back and add the remaining keys you collected earlier:

• Brave Search enables web lookups
• Groq enables voice transcription (speak instead of type)
• QMD handles additional device pairing if you want it

None of these are required to get started. All of them are worth adding once the core is working.

---

Why One Agent Beats Eight

The multi-agent experiment was instructive, in the way that expensive mistakes tend to be.

Eight agents sounds like more capability. In practice it meant eight separate context windows that could not see each other, eight sets of costs, and eight entities that each believed they were the only one handling things. Coordination was impossible. Memory was siloed. The whole thing was dramatically less useful than a single agent with continuity.

OpenClaw is built around a different model: one agent, one persistent memory, skills that extend capability without fracturing context. It remembers what you told it last week. It connects to tools without splitting into a separate entity to do so. The simplicity is the feature, not a limitation.

The $800 lesson, distilled: complexity is not a substitute for coherence. A single agent that actually knows you is worth more than a fleet of agents that keep forgetting.

---

What Comes Next

Once your agent is running, the natural next steps are:

• Setting up heartbeat checks (periodic scans of email, calendar, and notifications)
• Voice control via Telegram voice messages
• Connecting additional channels — Discord is particularly good for searchable conversation history
• Giving the agent tasks it can handle autonomously while you get on with other things

I will cover each of these in future posts. For now: get the agent running. Everything interesting starts from there.

---

Suneet Malhotra is an engineering leader specialising in AI-driven quality engineering. He writes about agentic systems, test automation, and building software that actually holds up. Connect on X/Twitter or LinkedIn.