Run a 1-Person Business with Claude Code (Full 2026 Setup)
Right now, three Claude Code agents are running on a home server in my office. One is triaging Gmail. One is generating invoices for a small business in Serbia. One is hunting leads. I haven't touched them in eleven days. If you're a solopreneur watching tutorials where someone builds a landing page in a chat window and calls it a business, you're being lied to. In the next forty minutes I'll show you the exact skeleton I use to run multiple Claude Code projects in parallel, autonomously. I'm Lazar, I ship these systems for clients every week. Let's start with the part nobody warns you about.
The part nobody warns you about is that Claude Code is not a chat tool. If you treat it like ChatGPT with a terminal, you will get a demo that impresses your friends and breaks the second you close the laptop. A one-person business does not run on demos. It runs on agents that wake up, do work, log what they did, and go back to sleep. So everything I'm about to show you is built around that one idea. Persistent, scheduled, boring. Boring is what makes you money.
Let's talk about the painful scenario first, because I want you to recognize yourself. You're running a small operation. Maybe you're a freelancer, maybe you're a two-person agency, maybe you sell a service to local businesses. Every morning you open Gmail and lose ninety minutes. You answer the same three questions from leads. You forward invoices to your accountant. You copy data from one tool to another because nothing talks to anything. By the time you actually start the work you're paid for, it's already lunch. You've tried hiring a VA, they ghosted. You've tried Zapier, it broke on the third edge case. You've tried plain ChatGPT, but it doesn't actually do anything, it just tells you what you should do.
The reason most existing solutions fail is that they're built for the average user, not for your specific workflow. Generic SaaS gives you ninety percent of what you need and the missing ten percent is exactly the part that would save you ten hours a week. Zapier-style automations are brittle because they're stateless. ChatGPT is smart but it can't run on a schedule, can't read your filesystem, can't talk to your CRM without you babysitting it. Claude Code closes that gap. But only if you set it up like an operating system, not like a toy.
So here's how I do it, end to end. First, the foundation. I run everything on a small home server. Old desktop, WSL Ubuntu on top of Windows, nothing fancy, the kind of machine you probably already own. I install Claude Code from the official CLI, log in once, and that's it for setup. The reason WSL matters is that you get a real Linux filesystem, real cron, real background processes. You can't run a serious agent stack on a sandboxed Mac app that closes when you sleep. You need a box that's always on.
Second, project structure. This is where ninety percent of tutorials skip the most important file. Every project I run has a CLAUDE.md at the root. That file is not documentation. It is the agent's brain. It tells Claude what this project is, what the rules are, what commands to run, what to never do, where the secrets live, and what the human expects when it finishes a task. If your CLAUDE.md is two sentences, your agent will hallucinate. If it's two pages of specific instructions, your agent behaves like a junior developer who already read the wiki. I keep mine boring and explicit. Project purpose at the top. Tech stack. File layout. Conventions. Commands to run for tests, for deployment, for sanity checks. A section called Never Do This. A section called Always Confirm Before. That last one is the difference between an agent that helps you and an agent that drops your production database.
Third, the AI operating system layer. I run multiple projects in parallel on the same machine. One handles invoicing for a small business client, one handles email triage and routing to a Telegram chat, one handles outbound lead research. Each project has its own folder, its own CLAUDE.md, its own subagents for specialized tasks like reading PDFs or calling APIs. They share MCP servers, which are basically little plug-ins that give Claude access to specific tools, a database, a Gmail account, a file store. Setting up an MCP server once and reusing it across three projects is how you stop reinventing the wheel every Monday. The mental model is simple. Project equals one business problem. CLAUDE.md equals the contract. MCP servers equal the hands. Subagents equal the specialists you call when the main agent doesn't know enough.
Fourth, let's build a real automation so you can see the pattern. We'll ship a Gmail-to-Telegram triage agent. The job is simple. Every fifteen minutes, check the inbox, look at unread messages, classify each one as urgent client, lead inquiry, invoice or receipt, newsletter, or noise. Route urgent client and lead inquiry to a Telegram chat with a one-line summary and a link back to the email. Archive noise. Tag the rest. The agent does this without you opening Gmail. To build it, I create a new folder, write a CLAUDE.md that describes exactly that behavior, hook up the Gmail MCP server and a Telegram MCP server, and write a single entrypoint script that Claude itself helps me generate. I tell Claude in the terminal, read the CLAUDE.md, build the entrypoint, write tests against three sample emails I paste in, and don't touch anything outside this folder. Forty minutes later I have a working script.
Fifth, make it autonomous, because a script you have to run manually is just a slower version of doing the work yourself. I add a cron job that runs the entrypoint every fifteen minutes. I add a logs folder where every run writes a timestamped file with what it did and why. I add a small alert rule that if the agent fails twice in a row, I get a Telegram message. That's it. That is the entire autonomy stack. No Kubernetes, no fancy orchestrator, no five-hundred-dollar-a-month SaaS. Cron, logs, alerts. Now when I sleep, the agent works. When I wake up, I scroll Telegram and I already know what happened in my business overnight.
Sixth, and this is the part most builders skip, you turn what you built into something you sell. Because a one-person business is not just automation, it's automation you charge for. Take the Gmail triage agent you just built. That same skeleton, with a different CLAUDE.md and different routing rules, is a service you can sell to a local real estate agent, a clinic, a law firm, anyone who lives in their inbox. Scope it as a flat setup fee plus a small monthly maintenance, deliver it by deploying the same code to their machine or to a small VPS, and use Claude Code itself to customize the CLAUDE.md per client. One skeleton, many clients. That is how a single person runs what used to be a five-person operation.
When you watch this running on my screen, what you see is not impressive on the surface. It's a terminal scrolling log lines. An email comes in, the agent classifies it, a Telegram notification fires, the inbox gets archived. That's the whole magic. Boring on the outside, ruthless on the inside. The result is two hours of my day back, every day, and a service I can resell. That's the actual one-person business stack. Not a landing page in a chat window.
Want more like this?
I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.
Subscribe to bizflowai.io on YouTube — never miss a new tutorial.
Planning an AI automation project or need a second opinion on your architecture?
Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.
Visit bizflowai.io for our services, case studies, and AI consulting.