How to set up a 24/7 AI employee (ClawdBot managed my business while I partied in Davos🤖)

Hey tech unicorns! 👋❄️🦄

I was just in Davos, Switzerland, at the World Economic Forum, surrounded by billionaires talking about the future of AI… ☕💼
Meanwhile, my own 24/7 AI assistant is grinding from the trunk of my rental car.

Yes. The trunk. 🚙😂

(Ok…. I ALSO have it running on AWS, which is what I’l show you in 1 min how EASY and FREE it is to set up…)

🔥but if you prefer videos the FULL CLAWDBOT VIDEO DROPS TOMORROW**
Click here to not miss it!

Otherwise…read on below to see how this magic works and how to set it up.

Why I Did This…

Between ski boots and an emergency tire repair kit sits a mini-PC running Clawdbot, quietly handling my emails, planning travel, and being way more organized than I am. Honestly, iconic behavior.

I built my own personal AI. It works for me, lives with me and does whatever I want- whenever I want. Email triage, logistics, planning. All handled. Even from a parking garage.

This way I can live life and work fully remotely from my phone (I’m tired of lugging my laptop up the Davos hills tbh…)

---

What’s Under the Hood 🧠…

This is personal AI, but make it unhinged and practical.

Clawdbot uses Claude’s Constitutional AI approach, which means it actually reasons instead of guessing.

It breaks problems into steps, checks its logic, and stays aligned with clear rules. Less pattern matching, more “let me think this through real quick.”

That’s why it can handle complex tasks without freaking out or hallucinating its way through my inbox 📬

Plus, there was a big update recently — Clawdbot now has adaptive context pruning and cache-TTL optimizations. Translation? It slashed token usage on repeated requests by basically remembering what it already cached instead of re-processing everything like it’s suffering from short-term memory loss. Your API costs just got cheaper, and responses got faster. Win-win!

---

The Setup (It’s Shockingly Simple)

How Clawdbot Actually Works (Technical Stuff)

Here’s the thing nobody tells you: Clawdbot isn’t just “an AI bot.” It’s an actual architecture with four distinct pieces working together:

1. The Gateway (Your Control Plane)

This is the brain. A WebSocket server running on port 18789 that coordinates everything. Think of it as air traffic control for your AI.

The Gateway handles:

• All incoming messages from your 15+ connected platforms (WhatsApp, Telegram, Discord, Slack, etc.)

• Session management (so your conversations persist and don’t get amnesia)

• Cron jobs (proactive briefings, reminders, scheduled tasks)

• Device pairing (your phone, laptop, iPad all connect to the same Gateway)

• Security approvals (before Claude deletes files or does anything destructive)

It runs as a background daemon via launchd (macOS) or systemd (Linux), which means it stays alive even if you close your terminal or reboot. That’s how mine keeps running in my car trunk 24/7.

2. The Agent (Where Claude Lives)

This is the actual AI model doing the thinking. Each agent runs in what’s called an “agent loop” — a single, serialized run per session that:

• Loads your workspace context (SOUL.md for personality, TOOLS.md for instructions, MEMORY.md for long-term notes)

• Builds a custom system prompt for every run (not generic ChatGPT vibes)

• Calls tools (read files, browse web, send messages, execute code)

• Streams responses back to you in real-time

The loop is serialized per session to prevent race conditions. This means if you send two messages rapid-fire, they’re processed in order, not simultaneously causing chaos.

Agents can be multi-routed too. You can have one agent for personal stuff, another for work, and route incoming channels to the right one. My setup uses one “main” agent with different workspaces for travel, coding, and personal admin.

3. Skills (The Extensions)

Skills are basically plugins that give Claude superpowers. They live in ClawdHub (the community marketplace) and you can install them with a single command.

Some skills I use constantly:

• Web search — Brave Search API integration (way better than generic ChatGPT web browsing)

• Browser automation — Controlled Chrome/Chromium with CDP (Chrome DevTools Protocol) for filling forms, scraping data, navigating sites

• Gmail/calendar management — Reads my inbox, drafts replies, books meetings

• Code execution — Runs scripts, tests code, debugs errors

Skills are loaded into the agent’s environment at runtime and injected into the system prompt so Claude knows they exist. You can create custom skills too — it’s just a folder with a skill.json manifest and some tool definitions.

4. Memory (Persistent Context)

This is what makes Clawdbot feel alive instead of lobotomized.

Memory lives as Markdown files in your agent workspace:

• MEMORY.md — Long-term notes (your preferences, recurring tasks, important context)

• memory/YYYY-MM-DD.md — Daily logs (short-term context that gets archived)

• SOUL.md — Personality and tone instructions

• TOOLS.md — How to use specific tools

• USER.md — Who you are, how to address you

These files are loaded at the start of every session and injected directly into Claude’s context. So when you tell Claude, “book me a flight to Zurich next week,” it already knows you prefer window seats, hate early morning flights, and want Swiss Air when possible because you told it once three months ago.

Session compaction handles the context window limit. When a conversation gets too long, Clawdbot runs a silent pre-compaction memory flush that reminds Claude to write durable notes to MEMORY.md before truncating old messages. So you don’t lose important context even in marathon 6-hour debugging sessions.

Semantic memory search is also supported (with OpenAI embeddings) — Claude can search past conversations for relevant context instead of keeping everything in the active window.

The Setup (It’s Shockingly Simple)

Here’s what I actually did to get mine running:

Step 1: Install Node.js 22+

Head to the NVM (Node Version Manager) GitHub and grab the install script. Run it in your terminal, then install Node 22 and set it as your active version. Takes about 3 minutes. Your terminal will look busy for a second, then you’re good.

Step 2: Install Clawdbot

Open your terminal and run the global npm install command for Clawdbot (it’s on their GitHub). The onboarding wizard pops up automatically and walks you through:

• Setting up the Gateway daemon (auto-starts on boot)

• Configuring OAuth or API key for Claude (I use OAuth from my Claude Max subscription)

• Connecting your first messaging channel (I started with Telegram, then added WhatsApp)

• Installing starter skills (web search, browser control, calendar)

• Setting your workspace directory (mine is ~/clawd-trunk because I’m funny)

Step 3: Configure your mini-PC (optional)

I used an AWS EC2 instance from the free tier (lol!) but will likely upgrade to a Mac Mini M4 Pro with 48GB RAM (overkill, but I want headroom for browser automation + video processing). You can absolutely run this on:

• A $5/month DigitalOcean droplet (2GB RAM minimum)

• A Raspberry Pi 4 (seriously, people are doing this)

• A VPS from Hetzner, Linode, or Railway

• Your existing desktop/laptop (just run it locally- FULLY FREE FOREVER!)

The Gateway binds to localhost port 18789 by default (localhost only). To access it remotely, you can:

• Use SSH tunneling to forward the port securely to your local machine

• Use Tailscale Serve/Funnel (built-in, auto-configured, magic)

• Expose it directly with password auth (but don’t do this unless you know what you’re doing)

Mine runs on a mobile hotspot in my car trunk with Tailscale for secure remote access. The dashboard lives at localhost port 18789 and I paste the generated auth token into my phone’s Control UI to connect.

Step 4: Pair your devices

Every device (phone, laptop, tablet) needs to be paired before it can connect. The Gateway generates a challenge nonce, your device signs it, and you approve the pairing in the dashboard. After that, the device gets a token and can reconnect automatically.

Local connections (same machine or Tailscale network) can be auto-approved. Remote connections require explicit approval (good for security).

Step 5: Connect messaging channels

Each platform has its own setup quirks:

• WhatsApp: Uses Baileys (web-based WhatsApp client). Scan QR code, done. But use a dedicated phone number because WhatsApp doesn’t have a “bot API” and can ban you if they detect unusual activity.

• Telegram: Create a bot with @BotFather, paste the token, done. Zero friction.

• Discord: Create a Discord app, enable MESSAGE CONTENT INTENT in the Developer Portal (critical or it won’t see messages), paste the bot token.

• Slack: OAuth flow, a bit more tedious but works great.

I have WhatsApp, Telegram, and Discord all connected to the same Gateway. Messages from any platform get routed to the same agent, so I can start a conversation on Telegram and continue it on WhatsApp without losing context.

Or if you don’t feel like setting that up…there’s also Claude Cowork, Watch my video about it here

---

What makes Clawdbot different from other AI assistants:

• It CAN reach out to YOU first — cron jobs for proactive briefings, reminders, alerts. Your assistant actually remembers to check on you instead of just sitting there waiting for commands like a lazy coworker.

• Multi-channel inbox - Works across WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and WebChat. One AI, all your platforms. No app-switching nonsense.

• Browser automation: it can navigate websites, fill out forms, click buttons, and extract data. Think of it as having an intern who never sleeps and never complains about tedious web tasks.

• Voice mode - Always-on speech for macOS/iOS/Android with ElevenLabs integration. You can literally just talk to it while driving (or skiing, no judgment).

• Live Canvas - Agent-driven visual workspace. Your AI can literally show you what it’s working on in real-time instead of just texting you updates.

• Skills marketplace - Community-built extensions for web research, email management, calendar automation, developer tools, and more. If someone’s already solved your problem, plug in their skill.
  Why This Changed Everything

My data stays with me.
My workflows are customizable.
Nothing breaks because a pricing tier changed overnight.

Also, there is something very satisfying about knowing your AI is running in your car while people upstairs debate the future of AI on stage 🧊🧠

What else can Claude do?

• Developers running overnight code testing with automatic PR creation when bugs get fixed

• Transcribing years of WhatsApp voice messages into searchable knowledge bases (I’ve exported Whatsapp convos with my VA and asked clawdbot to make notes of all action items to send to Notion….SOOO POWERFUL!)

• Automating insurance claims and expense tracking (saving literal hours every week-especially since we kinda scratched the rental car and will make a claim lol)

• Health reimbursement tracking that actually remembers to follow up

• Multi-agent systems where specialized AIs for coding, research, and personal tasks collaborate autonomously

Some people are even running Clawdbot on Raspberry Pis for $5/month.

It’s basically the AI assistant Siri promised to be but never became.

---

Big Picture 🌍

Personal AI is already here. It’s practical. It’s powerful. And yes, it fits in a car trunk.

Once you experience owning your AI, it’s really hard to go back.

Let me know if you were able to set it up in the comments of tomorrow’s YT video about it!

More soon 🤍
Possibly from another parking garage.

Delia 🚀