What We Open-Sourced
The agent skills and MCP tools behind Credible's agents now live in Malloy Publisher, the open source server for Malloy models. There are two parts. The tools let an agent look up what your data model actually defines. The skills, more than 30 of them, teach the agent to use those tools well: how to write the queries, which mistakes trip up even the best models, and how to check an answer before anyone relies on it. It all runs wherever you already work. You don't need a Credible account.
This post started a series. The posts that followed cover what we shipped, the ideas behind it, and a hands-on tutorial. The links are at the end.
The obvious reaction is the one I'd have: the skills are the product. They're the expertise. Why give them away?
Because we came to an uncomfortable conclusion about our own work, and about this whole market.
Excellence Is No Longer a Moat
In 1884, the Washington Monument was topped with a small pyramid of aluminum. Aluminum was a precious metal then, priced like silver. Napoleon III served his most honored guests on aluminum plates and everyone else on gold. Two years after the capstone went up, a new smelting process made aluminum cheap. Within ten years it was everywhere.
The metal didn't get worse. It was just as light, strong, and useful as before. It just stopped being rare. It was still valuable, but it no longer gave anyone an edge.
Data analysis tools and skills are having their aluminum moment.
Here's how to build a strong set of agent skills in almost any field. Take a few of the best textbooks on the subject and a couple of experts who have done the work. Give Claude a design approach for tools and skills, and a style guide. Then iterate.
That's how we built ours, and I'd put them up against anything on the market. But that's the point. Anyone can make them now. Good skills stopped being an edge as soon as anyone could reproduce them.
The Premium Is Gone
If a capable team with Claude, a few textbooks, and a couple of experts can rebuild these skills, nobody will pay a premium for them. Not for someone else's version, and not for ours. Open-sourcing them doesn't take the premium away. It admits the premium is already gone.
The other option is to keep the skills closed, wrap them in "trusted AI" language, and spend years convincing customers, investors, and ourselves that they're still worth paying for. Plenty of companies are making that bet, and some may get away with it for a while.
We chose differently while we still had the choice. The skills aren't the business. We gave them away so we could put our effort into the parts that are actually hard.
The market is already heading this way. Many "data analysis agents" are just skills and MCP tools inside a closed product, or custom prompts and agent code that felt new nine months ago. Those are bets against where the industry is going.
MCP and agent skills are the standards the industry is building on. Every new model release makes standard tools work better and leaves closed, custom ones further behind. Expertise still matters, but the experts who do well will share their work in the open, not hide it inside a product.
Trust but Verify
Coding agents made the industry's long-time dream easy to build: ask a question about your data in plain English, get an answer and a chart. Now everyone can build that demo, so everyone shows it, and every AI data product page sounds the same. When the demos look alike and the claims sound alike, marketing stops setting anyone apart.
That's why you see one word everywhere: trust. Trusted analytics. Trusted data for AI agents. Trustworthy AI. When you can't stand out on what your product does, you fall back on a claim nobody can check. Behind the word is usually the same offer: a closed product you can't look inside and can't take with you. Trust us, and stay locked in. It's the old SaaS model, closed product, premium price, walled garden, with "trust" printed on the box.
We're doing the opposite. There's no magic. Here are our skills, all of them, readable, in the open. They follow a design approach we also published, and they draw on a community whose expertise goes deeper than any one company's team. Read them, fork them, add your own team's knowledge, or tell us where they're wrong, in public.
You earn trust by being open. You can't get it by claiming it. A vendor that asks you to trust a black box is asking you to skip the one step that builds trust: looking inside.
Open Architecture Keeps Getting Better
If marketing can't set products apart, quality will. But most of the market looks for quality in the wrong place: the prompts, the wrapper, the screens. Quality comes from the architecture underneath, and over time the architecture decides everything built on top of it.
AI has changed how software gets built as much as what it does. Coding agents have flooded the market with systems that demo well and break in production. Most of them won't survive the next wave of AI adoption. The ones built to scale and grow will. Each wave leaves fewer survivors, harder problems, and bigger wins, and the survivors turn into lasting businesses.
That's why we built on Malloy. A community builds it and tests it. People extend it in ways we didn't plan, push it with workloads we didn't expect, and fix what breaks, and many of them know more than any one company's team. The result is more flexible and more solid than anything we could build alone.
A closed architecture improves only as fast as one team can work. An open one gets better every time someone new uses it.
Meet Every User Where They Work
No company has just one kind of data user. Yet most SaaS products are built as if it did: one screen, one login, one way of working, and everyone has to use it the same way.
We built the opposite. Because the stack is open and built on standards, you can use the same tools and skills wherever you work best:
Run the open source yourself. Download the server, the tools, and the skills and run them on your own machine. No account, no sales call. It's the same core we run.
Plug into your agent. Keep working in Claude, ChatGPT, Gemini, or your IDE. Install the skills and point the MCP tools at any Publisher server, yours or ours.
Use Credible. The hosted version, for people who want one place to open and start working: connect your data, ask questions, and build apps and agents, with access control built in. Start free. Unlike most SaaS products, nothing underneath it is a secret.
Whichever you choose, you can extend it. Start from our tools and skills and add what's specific to your organization: the edge cases, the business rules, how your team works with data. Your experts turn what they know into skills on the same foundation we use, and package them into your own plugin or app.
However you connect, it's the same data model underneath. Your engineers build agents and apps on the open source stack in their IDE. Your analysts work in the app. Your ops lead asks questions through the agent they already have open. Different people, different tools, one data model, so every answer agrees.
That's our promise to customers.
Our promise to the community is that it adds up. Every team that extends the skills and shares them back makes them better for everyone. No closed vendor can match that. They give you one way in. We support all of them. Switch any time, or use several at once, and nothing underneath changes.
What We Get in Return
We made one decision: stop defending something that can't be defended. In return we got three things a premium price never could buy: trust, an architecture built to last, and a way to reach far more people.
What we gave up was the premium, and it was already gone. We'll make it up in volume. When the doors are open, many more people walk through them, and the architecture is built to handle all of them. That's the hard part we chose to work on. Not the skills we gave away, but the engine that has to serve everyone who shows up.
This is the part we're most excited about.
AI has changed the playing field, and open source has come out of it stronger than ever. Every model release makes shared code, tools, and expertise more useful, faster than any closed company can keep up with. That's the future we're betting on.
One thing this post doesn't cover is our business model: what we charge for, and why being open puts us on the same side as our customers. That's a later post.
What's Shipped Since
Since this post went up in July, the series has walked through the engineering and the practice:
- The skills and tools that make an agent a data expert: what we shipped, and the ideas behind it.
- Tutorial: Use Malloy Publisher to model and analyze data with an agent: an agent answering real questions about your data in minutes.
- How an agent builds a data model and shows its work
- How an agent turns a question into a trustworthy answer
- A governed dataset end to end: Claude Code and Malloy on real CVE data
- Curate What Your Agents See: deciding what each agent, and each person, is allowed to see.
- Your Vibe-Coded App Just Found Its Serving Layer: taking an agent-built data app to production.
More of the stack has landed in Publisher too. The skills now go past modeling and analysis to evals (checking an agent's answers against a set of known-good ones and finding out why it missed), reports, HTML data apps, and materialization.
The tools and skills are free now. The expensive part was admitting why they had to be.
Start now: clone github.com/malloydata/publisher, open Claude in the folder, and say "let's get started!" Ask questions about your data in plain English. The agent uses the skills and tools in the repo to find the right data and the answer, with no black box in between. Want it on your own data, or help turning your team's expertise into skills? Get in touch at credibledata.com.
