Team Onboarding — Connecting Humans to the Swarm

The Last Mile Problem

You've built the agents. You've connected the APIs. You've deployed the MCP server. The system works beautifully — for you, the founder who built it.

But the rest of your team? They're still using Shopify admin panels, switching between 8 browser tabs, and asking you to "pull that number from somewhere." They don't know the agents exist. Or they think AI is something that writes blog posts.

This chapter solves that. In under 5 minutes per person, every employee on your team gets the same AI superpowers you have — without understanding a single thing about agents, MCP, or prompts.

What Employees Get

When a team member connects to the swarm, they can:

• Ask questions in natural language and get real answers from real data ("How much did we sell yesterday?")
• Read from the shared brain — every policy, process, and decision the company has documented
• Write to the shared brain — contribute knowledge that makes everyone smarter
• Talk to any agent — ask Ferland about P&L, Mafalda about tickets, Gala about store traffic
• Access every business tool — Shopify, Klaviyo, Google Workspace, Holded, Stockagile — without needing credentials on their machine
• Create and edit Google Sheets, Docs, Calendar — the full Google Workspace, through Claude

Zero API keys on employee machines. Zero training on business tools. Zero risk of credential leaks.

Architecture

┌─────────────────────────────────────┐
│    Team Member's Claude Desktop     │
│    (just a chat window)             │
└──────────────┬──────────────────────┘
               │ HTTPS (encrypted)
               │
┌──────────────▼──────────────────────┐
│    Cloudflare Tunnel                │
│    mcp.yourdomain.com/sse           │
└──────────────┬──────────────────────┘
               │
┌──────────────▼──────────────────────┐
│    MCP Server (your VPS)            │
│    44 tools · role-based access     │
│    all credentials server-side      │
└──────────────┬──────────────────────┘
               │
    ┌──────────┼──────────┐
    ▼          ▼          ▼
 Shopify    Agents     Brain
 Klaviyo    (6 AI)     (400+ docs)
 Holded     Aurelio    Policies
 GA4        Mafalda    Processes
 Meta       Ferland    Decisions
 Slack      Gala       Metrics
 GWorkspace Donatello  Knowledge
 ...        Olivia     ...

The critical design choice: employees never touch the infrastructure layer. They talk to Claude. Claude talks to the MCP server. The MCP server talks to everything else. If you change a tool, update an API, or add a new agent — nothing changes for the employee.

Setup Per Employee (5 Minutes)

Prerequisites

• Claude Pro subscription (€20/month) — the company can cover this
• A computer (Mac or Windows)

Step 1: Install Node.js

Node.js is required for the MCP bridge. It's a one-time install.

Mac:

curl -fsSL https://fnm.vercel.app/install | bash && source ~/.bashrc && fnm install --lts

Windows: Download from nodejs.org, install with defaults, restart.

Step 2: Add MCP Configuration

Mac — Open Terminal, paste:

mkdir -p ~/Library/Application\ Support/Claude && cat > ~/Library/Application\ Support/Claude/claude_desktop_config.json << 'EOF'
{
  "mcpServers": {
    "your-brand": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.yourdomain.com/sse"]
    }
  }
}
EOF

Windows — Create %APPDATA%\Claude\claude_desktop_config.json with:

{
  "mcpServers": {
    "your-brand": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.yourdomain.com/sse"]
    }
  }
}

Step 3: Restart Claude Desktop

Quit completely, reopen. The tools icon should show 44 tools connected.

Step 4: Test

Ask: "What were yesterday's sales?" — If Claude responds with real data, the employee is connected.

That's it. No API keys. No environment variables. No Docker. No terminal skills beyond copy-paste.

Access Control

Not every employee should have the same access. The MCP server supports role-based permissions:

Role	Can Read	Can Write	Example Employees
Admin	Everything	Everything (brain, files, shell)	Founder, CTO
Team	All APIs, brain, memory	Business APIs only	Finance, Marketing, Ops

Each employee gets a unique API key (lgm_{32-char-hex}) stored server-side. The key determines their role. Keys can be revoked individually without affecting anyone else.

What Team Members Cannot Do

• Execute shell commands on the server
• Write to the brain (read-only for now — can be unlocked per person)
• Access raw config files or credentials
• Modify agent behavior or prompts

The Shared Brain

The most powerful feature isn't data access — it's the shared brain. Every agent and every connected human reads from and writes to the same knowledge base.

Structure

brain/knowledge/
├── the-brand/          # Company knowledge
│   ├── finance/     # P&L rules, payment terms, tax info
│   ├── operations/  # Shipping, returns, warehouse processes
│   ├── product/     # Collections, sizing, materials
│   ├── retail/      # Store hours, staff, procedures
│   ├── marketing/   # Campaigns, segments, brand guidelines
│   └── team/        # Org chart, roles, contacts
├── platform/        # System documentation
└── projects/        # Active initiatives

How Employees Interact with the Brain

Reading (everyone can do this): - "What's our return policy?" - "Search the brain for wholesale pricing" - "What did we decide about the new store location?"

Writing (admin or unlocked): - "Add to the brain that Wholesale Account X requires 50% deposit upfront" - "Document that the flagship store closes at 20:00 on Saturdays" - "Update the brain: WELCOME15 code only works on first purchase"

Every write makes the entire system smarter. When an employee documents a supplier quirk, every agent — and every other employee — can find that information from that moment forward.

The Compound Effect

This is where the Compound Operations Model™ earns its name:

1. Week 1: Employee documents 3 supplier payment terms
2. Week 2: Finance agent uses those terms to flag a late payment automatically
3. Week 3: Another employee asks about the same supplier and gets an instant answer
4. Month 2: The brain has 50 new facts. Agents make better decisions. Employees get faster answers.
5. Month 6: The brain is the single source of truth for the entire company

No training needed. No wiki to maintain. No Notion pages to organize. The brain is just... there. And it gets better every day because the people who know the answers are writing them down in the same place the AI looks for answers.

Google Workspace Integration

Through the MCP server, every connected employee can work with Google Workspace using natural language:

Sheets

• "Create a spreadsheet with March sales by channel"
• "Add a row to the inventory tracker: Blazer Nara, M, 15 units"
• "Read the P&L sheet from the finance manager's Drive"

Gmail

• "Show me unread emails from suppliers"
• "Draft a reply to the last email from Supplier X"

Calendar

• "What's on my calendar tomorrow?"
• "Schedule a meeting with the finance manager for Thursday at 10am"

Drive

• "Find the FW26 sourcing document"
• "Search Drive for the brand guidelines PDF"

All of this works because the MCP server uses Domain-Wide Delegation — a Google Workspace admin feature that lets the server act on behalf of any user in the organization. The employee doesn't configure anything. They just ask Claude, and Claude does it.

Talking to Agents

Every employee can talk to any of the 6 agents through their Claude:

Agent	What to ask	Example
Aurelio	General operations, company data	"Ask Aurelio for yesterday's revenue breakdown"
Mafalda	Customer service, tickets	"Tell Mafalda to check today's open tickets"
Ferland	Finance, P&L, invoices	"Ask Ferland for the Q1 P&L"
Gala	Retail, store performance	"Ask Gala how Store A did this week"
Donatello	Digital marketing, ads, email	"Ask Donatello for Meta Ads ROAS"
Olivia	Inventory, merchandising	"Ask Olivia about sell-through rates"

The syntax is simple: "Ask [agent name] about [topic]" or "Tell [agent name] to [action]."

Rollout Strategy

Don't roll out to 20 people at once. Use this sequence:

Week 1: Power Users (2-3 people)

Pick the people who already use ChatGPT or Claude. They'll find bugs, figure out the best prompts, and become internal evangelists.

Week 2: Department Leads (4-5 people)

Finance lead, CS lead, retail manager, marketing lead. These people will use it for real work and give feedback on what's missing.

Week 3: Full Team

By now you have working examples, a list of "things to try," and 5 colleagues who can help troubleshoot. Send the setup guide and let people explore.

Ongoing

• Share "prompt of the week" — real examples that worked
• Encourage brain writes — the more people contribute, the more valuable the system becomes
• Track usage — the MCP server logs every tool call

Troubleshooting

Symptom	Fix
"No tools connected"	Check Node.js: node -v. If missing, reinstall.
"Connection error"	Server may be restarting. Wait 2 minutes.
First query is slow	Normal — first SSE connection takes 5-10 seconds.
"Tool not found"	Restart Claude Desktop completely (quit + reopen).
Wrong data	Check if the query is ambiguous. Be specific about dates and channels.

For anything else: contact the founder or check #ai-agents in Slack.

Security Model

• Zero credentials on employee machines — all API keys, tokens, and passwords stay on the MCP server
• Encrypted in transit — Cloudflare Tunnel provides HTTPS/TLS
• Role-based access — admins vs. team, individually revocable
• Audit trail — every tool call is logged with timestamp and user
• No PII exposure — employees access tools through the AI layer, never raw database connections

Cost Per Employee

Item	Cost	Notes
Claude Pro subscription	€20/month	Can be covered by company
MCP server	€0 incremental	Already running for agents
Cloudflare Tunnel	€0	Free tier
Total per employee	€20/month

For context: a single Shopify admin license costs €32/month, gives access to one system, and requires training. Claude + MCP gives access to everything and requires no training.

What This Changes

Before: - Employee needs X data → asks someone → waits → gets partial answer - New policy decided → lives in someone's head → forgotten - Tool credentials → scattered across 15 systems → security risk

After: - Employee needs X data → asks Claude → gets answer in seconds - New policy decided → written to brain → available to everyone forever - Tool credentials → one server → zero on employee machines

The swarm isn't just the agents anymore. It's the agents, the brain, and every human in the company — all connected, all learning, all getting smarter together.

---

This chapter represents the final layer of the Compound Operations Model™. The technology is ready. The agents are running. The brain is growing. The only thing left is to connect the humans.

AI Adoption Framework (L0-L3)

Based on the approach that achieved 99.5% active adoption in a production deployment:

L0 — Observer

• Uses generic AI chatbots occasionally
• Hasn't changed any workflow
• Expectation: exit L0 within first week. If no progress, manager conversation.

L1 — Active User

• Connected to the shared brain via Claude Desktop / MCP
• Uses brain_search before asking questions about the brand
• Has completed the initialization prompt
• Target: entire team at L1 within 30 days of deployment

L2 — Builder

• Has built something that automates part of their job
• Uses skills and tools autonomously
• Contributes by writing to the brain when they learn something new
• Shares what they build with the team
• Target: 25-30% of team at L2 within 60 days

L3 — Multiplier

• Builds tools/workflows that benefit other teams
• Creates new skills or improves existing ones
• Teaches others how to use the system
• AI reference person in their department
• Target: 2-3 people at L3 within 6 months

Measurement

• AI level reviewed in monthly check-ins
• Sustained L0 (>30 days without brain usage) → manager conversation
• L2+ achievements celebrated in All Hands
• New hires must demonstrate AI aptitude during selection process

Day 1 Mandatory Setup

Every new hire completes on their first day: 1. Install Claude Desktop + connect to MCP (30 minutes with manager) 2. Run initialization prompt (1 hour of guided exploration) 3. Manager verifies: MCP connected, 44 tools visible, brain_search works 4. First real task using AI assigned (due by end of Week 1) 5. First task shared in #ai Slack channel (public commitment)

The biggest surprise: it wasn't who built the most. It was how many people had been waiting for permission to build at all.

AI Adoption Framework (L0-L3)

Production-tested framework that achieved 99.5% active adoption in the reference deployment:

L0 — Observer

• Uses generic AI chatbots occasionally. Hasn't changed any workflow.
• Expectation: exit L0 within first week. If no progress, manager conversation.

L1 — Active User

• Connected to the shared brain via Claude Desktop / MCP. Uses brain_search before asking questions.
• Target: entire team at L1 within 30 days of deployment.

L2 — Builder

• Has built something that automates part of their job. Contributes by writing to the brain. Shares what they build.
• Target: 25-30% of team at L2 within 60 days.

L3 — Multiplier

• Builds tools/workflows that benefit other teams. Creates new skills. Teaches others.
• Target: 2-3 people at L3 within 6 months.

Day 1 Mandatory Setup

Every new hire completes on their first day: 1. Install Claude Desktop + connect to MCP (30 minutes with manager) 2. Run initialization prompt (1 hour of guided exploration) 3. Manager verifies: MCP connected, all tools visible, brain_search works 4. First real task using AI assigned (due by end of Week 1) 5. First task shared in team AI channel (public commitment)