What We Built in Session 1

Russ opened with the core philosophy: there are a thousand AI tools, but most of them are renting you a black box you'll never own. ChatGPT, Notion AI, Superhuman — you can give them your credit card. But then they own your workflow.

What we're building is different. It runs locally. It costs ~$20/month. You control it completely. And the skills you learn here transfer to any stack, any model, forever.

• Off-the-shelf AI: $20–200+/mo, resets every session, you wait for their integrations, they own your data
• What we're building: ~$20/mo, persistent memory, runs locally, you build exactly what you need, you own it forever
• The skills aren't tool-specific — they're infrastructure skills. If a better harness comes out, you'll know how to switch.

Every session follows the same structure: first Russ shows you the finished dish (what's already built and running), then you build it together live. This session was extra raw because it was Session 1 — the hardest install session. It gets smoother from here.

A lot of the magic — being able to voice note from your car, getting a morning brief before 7 AM — only works if your computer is always on and connected.

• You can absolutely start this on your laptop — Russ did
• But the dream setup is a dedicated always-on Mac (Mac Studio or Mac Mini) sitting somewhere in your home or office
• Old Mac works fine. We're not running local AI models. The compute requirement is low.
• Before Session 2: if you have access to a desktop Mac, redo this first install on that machine instead of your laptop

Your AI COS can live in any messaging platform — iMessage, Discord, Slack, Telegram. Here's Russ's reasoning:

• Telegram: Most secure default. Locked down. Hard to accidentally expose to others. Lives on your phone for voice notes.
• Slack: Useful because it's in your team channels. Pin can read the channel, create issues, post notes — but is locked to that channel only, not exposed to your whole team.
• iMessage: Higher accidental exposure risk. Not recommended as primary.
• You can run all of them simultaneously — same brain, different endpoints.

Great live question from the class. Russ's answer: running this locally is actually safer than running it in the cloud. Cloud setups have more configuration surface area for mistakes. Local is contained.

The main things you need:

• Closed WiFi network — basic hygiene, your home network should require a password
• Mac Firewall enabled — System Settings → Network → Firewall. Built-in, one toggle.
• Tailscale — this is the magic sauce. Free for up to 3 devices. Creates a private encrypted VPN between your phone, laptop, and home Mac. Nobody can access your machine even if they somehow found your IP.

The real risk question is: are you comfortable with Claude having access to your Gmail? For most people, yes. Your email, calendar, tasks — that's the point. You're not putting your SSN or banking credentials in here.

This is the mental model that makes everything click:

The harness (Hermes) is what makes your COS actually useful. It adds persistent memory, connects to Telegram/Slack, runs scheduled tasks, and lets you wire in any tool or API. Think of it as the operating system your AI lives inside.

Tabby hasn't kept up with recent macOS updates and is throwing security errors. Use Ghostty instead. It's cleaner, faster, and widely used by engineers.

PC users: Ghostty works on Windows via WSL2 installed first. When you hit errors, take a screenshot and paste into Claude web with the Hermes GitHub link — that's how you troubleshoot.

Two installs required. Hermes assumes Claude Code is already present — if it isn't, the installer will fail silently.

Install Claude Code first:

Go to Claude Code docs, copy the macOS install command, paste into Ghosty.

npm install -g @anthropic-ai/claude-code

Then install Hermes:

Go to the Hermes GitHub, scroll to Quick Install, copy the command:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | sh

Hit approve/yes on every popup. The installer checks for dependencies and installs anything missing.

During setup, pick Anthropic → API Key. Set up your API account at platform.claude.com — this is pay-as-you-go, no subscription required.

• Go to platform.claude.com and create an account
• Create a new key — name it whatever (your bot name works)
• Copy it — you only see it once, so save it in a password manager, not a Google doc
• Add $10-20 in credits to start. You can add more as you go.
• Russ spends ~$100/month using it heavily every day. For a getting-started budget, $20 is fine.

This is the step that connects your COS to your phone. Here's the exact flow:

• Download Telegram on your phone if you haven't
• Search for @BotFather — it's an official Telegram bot
• Send: /newbot
• Name it (this is your public display name — anything you want)
• Give it a username — must end in "bot" (e.g. mycos_bot)
• BotFather will give you a long token — copy it carefully
• Paste into Ghosty when Hermes setup asks for it (it won't show on screen — that's normal)

Lock it to only you:

• Search Telegram for @userinfobot
• Message it — it'll reply with your numeric user ID
• Copy that ID and paste it into Hermes setup when asked
• This ensures only your Telegram account can talk to your bot

When Hermes setup asks about the gateway — hit Y (capital). The gateway is the always-running background process that:

• Keeps your Telegram/Slack connection alive
• Runs scheduled tasks (morning brief, email triage)
• Lets your phone reach your Mac from anywhere via Tailscale

This was the most-repeated lesson of the night. Copy and paste is the #1 skill when building with AI. Most of what you'll do is: find a command online → copy it → paste it into your terminal → hit Enter.

• Terminals often don't show what you pasted (for security). Trust that it worked and hit Enter.
• In Ghosty, a glowing lock icon in the corner confirms a sensitive paste succeeded.
• API keys are one-time view — save them in a password manager (1Password, Bitwarden, etc.) immediately after generating.

Hit a wall? The move is: screenshot the error → open Claude web → paste the screenshot + context → ask it what to do.

Useful context to include in your Claude debug message:

• "I'm trying to install Hermes: [Hermes GitHub link]"
• "Here's the error I'm getting: [screenshot]"
• "I'm on macOS [version] / Windows / Linux"

Once your COS is running, it knows nothing about you. The first thing to do is introduce yourself. Russ demonstrated by typing:

I want you to become my Chief of Staff. Interview me — ask me 20 questions to learn everything about me, my business, my role, and my life. Start now.

The longer you spend on this, the better your COS gets. Hermes writes everything you tell it to persistent memory — so it actually remembers next session, next week, next month.

This is the main advantage over standalone ChatGPT — it accumulates context instead of resetting.

We ran out of time before wiring in Gmail and Calendar. This is your homework before Session 2. Once Hermes is installed, run:

hermes setup google-workspace

It walks you through a one-time OAuth flow. Then verify it worked:

hermes "How many unread emails do I have?"

When it answers correctly — your COS is inside your inbox.

Once Gmail is wired, run these two commands to activate the automation:

hermes "Set up 3-tier email triage every 30 min. Tier 1: board/investors/escalations. Tier 2: team/projects/customers. Tier 3: archive newsletters and cold outreach. Label each email."

hermes "Set up a daily 7 AM brief: unread Tier 1 and Tier 2 emails, today's calendar, my top priorities. Send it to my Telegram."

Your first brief should hit the next morning. If it doesn't, post in the community.

Russ briefly showed a Bland.ai phone call — your COS making and taking real calls. This is coming in a future session. The teaser: you'll provision a real phone number connected to your AI, which can make outbound calls to book appointments, screen vendors, and leave messages — while you're doing something else.