By Gustavo Paixão
Building a marketplace where people decide how AI works
How Ronne AI Marketplace went from copying AI skills between projects by hand to two open-source releases in its first week, with items flowing between the registry and Claude Code, Codex and Cursor.
The idea
Every time I started a new project, I copied my AI skills, commands and agents into it by hand, then adapted them to fit. Each project ended up with a slightly different version of the same thing, and keeping them in step by hand wasn't working.
AI coding tools are only as good as what you give them: skills, agents, rules, hooks and connections to other tools. So I wanted one place that keeps those pieces together and decides what goes in. That became Ronne AI Marketplace. It's open source, and whoever uses it installs it on their own servers. Anyone can propose an item, and nothing reaches the AI tools until it's been reviewed.
The line I kept coming back to was simple: people should control how their AI works, not the other way around.
Why "Ronne"
When I started preparing the project, I needed a name. I wanted a person's name, because the product could act as an AI assistant. I spent a few hours going back and forth with ChatGPT, Claude and Gemini: male and female names, in different languages. Every name I liked was already in use, often by another AI product, or had no domain left to register.
Just before bed, two names came to mind: Rainier and Ronne, two of my uncles. Rainier passed away in 2009. Ronne is my mother's youngest brother, and he lives in the countryside of Pernambuco, in Brazil. I asked ChatGPT about the two, and it pointed to Ronne: it's easy to say in English, Portuguese and French.
That settled it. The name means something to me, and it's my way of celebrating Ronne, Rainier and my family.
Shaping it
One format, many tools. Claude Code, Codex and Cursor each keep these items in different files. So each item is written once, and a command-line tool, rmk, writes the right files based on each harness's current specifications.
Review before release. Every new item, and every change to one, needs an approval from someone other than the author. Releasing is a separate step, and a released version never changes.
A rule for its own code. Everything the marketplace depends on has to be free to redistribute, on its latest stable version, and free of known vulnerabilities. The build checks the rule on every change, and it has already ruled out two libraries.
Then I asked a question that widened the plan: why only three tools? The AI coding tools are converging on the same formats, and they change fast. So each tool became a small module that translates items into its files.
How I build it
I build the marketplace with an AI coding agent. It writes a lot of the code, but the decisions are mine.
Early on, the agent started to assume it could push changes on its own. That's not how I want to work, so I set rules for it, the same way the marketplace lets people decide what their AI tools can do. Before any commit, it runs the lint, the type checks, the tests and the build. If something fails, it tells me what's wrong and proposes what to do, and I decide. It never pushes code or opens a pull request unless I ask.
Later, I let it build whole features on its own, one task and one commit at a time. Then I review the pull request, try the result, and decide whether it goes in.
From an install to a catalogue
The first code made the marketplace installable with one command, on SQLite, MySQL, MariaDB or PostgreSQL. Every change is tested on all four, because each behaves a little differently. To check that this was worth it, I broke the search on purpose. Three of the databases still passed. Only PostgreSQL caught it.
Then came items and review. A test that submitted two drafts with the same name at the same moment found that MySQL let both through, and I fixed it before it could reach anyone. Whoever reviews a submission sees what the item can do on a developer's machine, such as running a command or reaching the web.
Using it as a new user would, the first thing I asked was "what is a scope?". So I added short answers inside the app, where people tend to get stuck.
Into the tools
Then came the reason the marketplace exists: putting items into the tools. rmk records every file it writes, so it can update or remove exactly that, and never touches a file someone made by hand.
Before writing each tool's part, I checked its documentation again. Three things had changed in Claude Code since I first looked. And Cursor turned out to read Claude Code's skills and run its hooks, so installing for both would have run every hook twice. Now Cursor leaves those to Claude Code.
The MCP server lets people install items without leaving the conversation with their AI tool. It plans first and writes second: the assistant shows every file an install would change and what each item can do, and only then applies exactly that plan. To keep the terminal and the AI tools from drifting apart, both use the same install code, which changed the architecture I had planned.
Two releases in a week
On 29 September I released version 0.1.0: rmk and the MCP server on npm, and the web app as a Docker image. The release found a bug in the publishing step, and 0.1.1 followed the same day.
Until then, items only went from the registry into the AI tools. But most skills and agents start in the tool, while someone is working. So I made them go both ways: people can send what they write in Claude Code, Codex or Cursor back to the marketplace, as a draft or a proposed change, through the same review as everything else. That meant putting more AI tools on hold, to test and validate the marketplace first with the tools I use most. Version 0.2.0 came out on 1 October, with export and a visual composer to build agents and bundles on a canvas.
What's next
For now I'm its first user, on purpose: I use it across my own projects to see which features hold up and which don't. Soon a team will use it on a single product, the case it was built for: many people sharing one curated set of items. What I learn from both decides what changes before I add the other tools.
Open source
Ronne AI Marketplace is open source under the MIT license, and the repository has been public since the first day. ronne.ai explains what it does and how to run it. The code is on GitHub, the web app is on Docker Hub, and rmk is on npm.
Contributions are welcome. The work is organised as features, each with a spec and a plan in the repository, and every pull request runs the same checks I run, including the license and security checks. Two good places to start:
- Another AI tool. Copilot, Gemini, Antigravity and Devin are on hold, and each tool is a small module. If you use one of them every day, you know its files better than I do.
- Issues. If something doesn't install, a step isn't clear, or an item behaves differently in your tool, open an issue.
Let's contribute
If you try it, or you manage your AI tools a different way, I'd like to hear about it.