Build Your Own AI Assistant
I run my entire operation through an AI assistant named Beckett. He lives in my Telegram, remembers every conversation, checks in when something needs my attention, and handles voice notes when my hands are on the wheel.
This page walks you through building your own. Free code, about 45 minutes, no coding required.
Here's the part that surprises people: you won't follow a 40-step manual. The setup guide is built into the code itself. Claude reads it and walks you through in plain English. You answer questions, it does the configuring. You're not the technician here. You're the client.
What you're building
- A private Telegram bot that only responds to you
- Powered by Claude, the same AI I use to run my businesses
- Real memory: it stores conversations in your own free database and recalls relevant history automatically
- Optional: voice message transcription, morning briefings, proactive check-ins
- Runs on your computer, on your accounts, under your control
What you need before you start
| Thing | Cost | Time to get |
|---|---|---|
| A computer (Mac or Linux; Windows works via WSL) | you have one | now |
| A Telegram account | free | 2 min |
| A Claude subscription (Pro) | $20/mo | 5 min |
| A free Supabase account (your bot's memory) | free | 5 min |
| An OpenAI API key (powers memory search, pennies/mo) | ~cents | 5 min |
The Claude subscription is the only real cost. Everything else runs on free tiers.
Step 1: Install the tools (5 minutes)
Open your terminal (Mac: press Cmd+Space, type “Terminal”, hit Enter) and paste these one at a time:
# Bun: the runtime that runs your assistantcurl -fsSL https://bun.sh/install | bash# Claude Code: the AI that will set everything up for younpm install -g @anthropic-ai/claude-codeIf npm isn't found, install Node first from nodejs.org, then rerun the second command.
Step 2: Get the code (2 minutes)
git clone https://github.com/tuni-kern/beckett.gitcd beckettNo GitHub account needed. You're downloading a copy; everything personal you add later (your keys, your profile, your bot's memory) stays on your machine and in your own accounts. It never goes back to GitHub.
(Developers: fork it if you want to track your own changes. Everyone else: clone and move on.)
Step 3: Let Claude take it from here (30-40 minutes)
This is the whole trick. In that same folder, type:
claudeThen say:
help me set up my assistant
Claude reads the setup instructions baked into the project and interviews you, one phase at a time. It asks for what it needs, configures everything itself, and tests each piece before moving on.
Here's what it will walk you through, so nothing surprises you:
Phase 1: Your Telegram bot (~3 min). You'll message @BotFather on Telegram to create a bot (it hands you a token), and @userinfobot to get your user ID. Claude saves both and sends you a test message. When it arrives, phase done.
Phase 2: Memory (~12 min). You'll create a free Supabase project and paste in two keys. Claude builds the database tables, deploys the search functions, and wires up semantic memory using your OpenAI key. This is what separates your assistant from a goldfish: it will actually remember what you told it last Tuesday.
Phase 3: Personalization (~3 min). Claude asks your name, timezone, what you do, and how you like to be talked to. Your assistant stops being generic here.
Phase 4: First conversation (~2 min). Bot goes live, you text it from your phone, it answers. This is the moment it clicks.
Phase 5: Always on (~5 min). Claude configures your assistant to run in the background, start on boot, and restart if it crashes. Close the terminal, it keeps working. One limit: your assistant is awake when your computer is. If the lid closes, it naps too. That's what the cloud server in the full build is for.
Phase 6: Proactive mode (optional). Morning briefings and smart check-ins: your assistant decides whether something is worth pinging you about. This is the difference between a chatbot and an assistant.
Phase 7: Voice (optional, ~5 min). Send voice notes instead of typing. Free transcription via Groq, or fully offline with local Whisper if you're privacy-first.
You can stop after Phase 4 and have a working assistant. Phases 5-7 are what make it feel like staff instead of software.
If something breaks
Tell Claude. Seriously, that's the troubleshooting guide: paste the error into the same Claude Code session and say “this happened.” It set your system up, it knows where everything is, and it will fix its own work. The three usual suspects are a mistyped bot token, the wrong Telegram user ID, or Bun not being on your PATH yet (close and reopen the terminal).
Want the printable version?
One-page setup checklist PDF: every account, key, and command from this guide in order.
No sequence-spam. You'll get the checklist and the occasional build note.
Where this goes next
The free version you just built is a real assistant, and for a lot of people it's enough.
Mine does more: it runs on a $5-10/month cloud server so it never sleeps, reads my email and calendar, asks permission before taking actions, and calls my phone when something's urgent. That's the full build, and I teach it step by step inside The Agent Lab, my free community: build logs of everything I ship, and direct help from me when you're stuck.
And if you'd rather have it done for you (set up on your accounts, personalized to your business, working by Friday), that's what I do. Talk to me.
Readers will glance at this page and think “neat.” Builders will spend 45 minutes tonight and wake up tomorrow to an assistant that already knows them.
Be the second kind.