Claude Code + MCP: How I Build AI Infrastructure Without a Team

The Solo Founder Problem

I run an AI company called UmniAgents. We build AI agents for businesses — dental clinics, salons, restaurants, hotels, e-commerce stores. Each client needs a custom agent trained on their business, connected to their systems, deployed on their website.

That's a lot of moving parts. Backend APIs, frontend widgets, database schemas, deployment pipelines, client communication. The obvious move is to hire a team. A backend dev, a frontend dev, maybe a DevOps person.

I didn't do that. Instead, I built the entire infrastructure with Claude Code and MCP servers. One person. One machine. 20+ clients served.

This isn't a flex — it's a practical decision. When you're a solo founder in Eastern Europe, every $3,000/month salary is a bet. I needed to prove the business model first. Claude Code let me move at the speed of a small team without the overhead of one.

My Actual Stack — Claude Code + MCP Servers

Claude Code is Anthropic's CLI tool for coding with Claude. It's not a chatbot you paste code into. It lives in your terminal, reads your codebase, writes files, runs commands, and ships features. Think of it as a senior developer sitting next to you who never gets tired.

But the real power comes from MCP — Model Context Protocol. MCP lets Claude Code connect to external tools through standardized servers. My database, my messaging channels, my deployment scripts — Claude Code talks to all of them directly.

Here's what my setup looks like:

• Claude Code instances — separate instances for different workstreams (client work, infrastructure, content)
• MCP channel plugin — Claude Code receives client requests through messaging channels and responds directly
• PostgreSQL — all client data, conversation history, agent configurations
• WebSocket pipelines — real-time communication between AI agents and client websites
• Multi-tenant architecture — one server runs all 20+ client agents, isolated but sharing infrastructure

A typical workflow: a client sends a request through a channel. My Claude Code instance picks it up, reads the relevant code, makes the change, tests it, and deploys. I review the diff and approve. What used to take half a day takes 20 minutes.

What This Looks Like Day-to-Day

Monday morning. I open my terminal. Three client requests came in overnight — a dental clinic wants to add appointment booking to their AI agent, a salon needs updated pricing in their bot, and a restaurant wants their agent to handle catering inquiries.

I don't write the code for any of these manually. I describe what needs to happen to Claude Code. "Add appointment booking flow to Coldy Dent's agent. Use their existing calendar API. Confirm the booking with the patient and send a summary." Claude Code reads the existing codebase, writes the integration, and I review the result.

The salon pricing update? That's a 2-minute task. The restaurant catering flow? Maybe 30 minutes of back-and-forth with Claude Code to get the conversation design right.

By lunch, all three are deployed. No standup meetings. No PR reviews from teammates. No waiting for someone to context-switch to my request.

The math is simple. A mid-level developer in Bulgaria costs $2,000-3,000/month. Two developers plus a part-time DevOps person — call it $6,000-8,000/month. My Claude Code setup costs a fraction of that. And it works at 2 AM when a client's agent goes down.

Why This Works (And Where It Doesn't)

I want to be honest about the limits. Claude Code doesn't replace thinking. It replaces typing. I still design the architecture. I still decide which trade-offs to make. I still talk to clients and understand their problems. The creative and strategic work is mine.

Where Claude Code excels: repetitive backend work, CRUD operations, API integrations, database migrations, frontend components, debugging, writing tests. Basically, 70% of what a dev team does daily.

Where I still step in: system architecture decisions, complex debugging that requires deep domain knowledge, client conversations, and anything that requires understanding the business context beyond the code.

This approach works especially well for a solo founder in Eastern Europe. The talent pool here is strong but expensive relative to local revenue. Clients pay local prices, but developer salaries track closer to Western European rates. Claude Code closes that gap.

My infrastructure bill — server, database, Claude API, domain, hosting — runs under $500/month total. That serves 20+ clients generating recurring revenue. The unit economics don't work with a traditional team at this stage. With Claude Code, they do.

I'm not saying never hire. When UmniAgents hits 50+ clients, I'll bring on people. But they'll be working alongside Claude Code, not instead of it. The future isn't AI or humans. It's knowing which parts to give to each.