Business Agent OS: a free business operating system for Claude Code, Codex and Grok

I’ve published a free agent skill called Business Agent OS. It sets up a business operating system for an AI coding agent such as Claude Code, from one folder on your own computer. It’s about 99% of the agent operating system I run my business on, the same memory, commands and habits I use every day.

Who it’s for

It’s built for business owners, founders and creatives who know enough about code to read it, even if they don’t write it, and who like having access to the source of everything they use. People who would rather run their business from the root level, in source code, than through someone else’s interface.

It’s also for people who understand the value of software tailored to their own business. Why bother finding off-the-shelf software and adapting your business to fit it, when an agent can build your own apps from the ground up, shaped to the unique blueprint of your business? That’s the idea behind agent services for businesses, and this skill is the foundation that work runs on: an agent that remembers your business, from files you own.

Why the source code matters

Most people aren’t aware of how much software runs their life and their business: the quoting tool, the CRM, the booking system, the spreadsheets, the inbox rules. Because it all arrives finished, it feels fixed.

A coding agent changes that. An AI agent like Claude Code reads and writes the source code of software, so once you have the source, you can change anything. Instead of bolting on generic off-the-shelf software and then adapting the business to it, you can personalise the software around the way you naturally run your business. This skill gives your agent the memory of how your business runs, so every change it makes fits.

More than a business operating system

Building your own software with an AI agent, around how you actually live and work, is the ultimate productivity tool. It’s almost more than a business agent operating system. It’s a life agent operating system.

I run more than the business on mine. Health research, recipes, the dog’s training, the garden and the house each have their own notes, and the agent builds small tools for them the same way it does for clients. The skill’s projects/ folder is there for exactly that: anything in your life that would run better with its own memory and its own software.

Where it runs

It runs on any Mac. It’s designed for an always-on machine, such as a Mac mini at home or in the office, that you reach over SSH from your phone, tablet or laptop. tmux keeps each client’s session running while you’re away, Tailscale gives you a private route in from anywhere, and ShellDrop is the iPhone and iPad app I use to connect. The full setup is in how I run Claude Code via tmux from an iPhone.

Install

Install it with one line, then ask your agent to set it up:

npx skills add danielbilsborough/agent-skills --skill business-agent-os

Need Claude Code first? How to install Claude Code.

The source is on GitHub, it’s listed on skills.sh, and there’s a read-only MCP server for agents that prefer tools. It’s MIT licensed.

What it sets up

• A business CLAUDE.md from five questions: what you do, who your customers are, what you sell, how the agent should talk to you, and what it must never do without asking.
• A folder per client and per project, each with five plain text files: state, changelog, to-do, inbox and the client’s own rules.
• A knowledge base following Andrej Karpathy’s LLM wiki idea. You save the raw material (articles, call transcripts, PDFs), and the agent writes and maintains linked topic, people and source pages from it, with an index and a log. Open the folder in Obsidian to see it as a map.
• Session commands. /pickup loads a client and gives you a short brief, blockers first. /wrap saves the session back into the files. /status shows every client on one screen, /review tells you what you did on any day or month, /trace shows how a decision evolved, and /ingest and /lint run the knowledge base.
• Save points, so any change the agent makes can be seen or undone.
• A tmux “go” menu: one session per client, one keypress to jump between them.

Why a system like this, when Grok Bot and the desktop apps exist

I’ve used Grok Bot with Grok 4.7, and it’s a very good app. The Claude and ChatGPT desktop apps are good too. They’re built to be products, and the memory, the tools and the way they work belong to the product.

This works one level down, at the coding agent level, from the root folder of your business. The memory is plain Markdown text files in a folder you choose. A year of notes for a busy business is a few megabytes. You can open every file in a text editor, search it, back it up, move it to another computer, and read exactly what your agent knows about each client.

Because the memory is just text files, it isn’t tied to one agent or one app. The skill is optimised for Claude Code in the terminal, and the same folder works with Codex and Grok. It also pairs with tools like MCP servers and Claude Cowork. Switch agents and the memory comes with you.

What it gives you

• No re-explaining. /pickup acme-plumbing reads that client’s files and briefs you in seconds: where things stand, what’s blocked, what happened recently, what’s next.
• A complete record. Every /wrap leaves a dated entry in the client’s changelog, and old entries are never changed. At the end of a month you can ask “what did I do on 14 March?” or “when did we agree the new price, and why?” and get an answer from the files.
• Small and fast. Older changelog months move to a monthly archive, so nothing is lost and pickups stay quick.
• Safe by design. The agent asks before every step, installs nothing for you, and keeps everything inside the one folder.

It’s free to use as it is. If you’d like to set it up live with someone, that’s what the Claude Code course covers in half a day. If you’d like help shaping it around how your business actually runs, that’s what I do as a business technology partner.