What We Just Open Sourced
The agent skills and MCP tools behind Credible's agents now live in Malloy Publisher, the open source server for Malloy models. Retrieval tools that let an agent look up what your model actually defines, and 24 skills encoding the discipline to use them well: the query patterns, the gotchas that trip up frontier models, and the analysis rigor that separates a real answer from a plausible one. Everything runs wherever you already work. No Credible account required.
This is the opening post of a series. The deep dives -- what we shipped, the principles behind it, and a tutorial that gets an agent answering real data questions in minutes -- land over the coming weeks.
The reaction to the news is predictable, because I'd have the same one: those skills are the product. That's the expertise. Why would you give that away?
Because of an uncomfortable conclusion we reached about our own work, and about this entire market.
Excellence Is No Longer a Moat
In 1884, the Washington Monument was capped with a small pyramid of aluminum -- then a precious metal, priced alongside silver. Napoleon III served his most honored guests on aluminum; gold was for everybody else. Two years after the capstone went on, a new smelting process made aluminum cheap. Within a decade it was a commodity.
The metal did not get worse. It's as light and strong and useful as it ever was. What died wasn't the value. It was the scarcity. Value and defensibility came apart.
Data analysis tools and skills are having their aluminum moment.
Here's the recipe for a high-quality set of agent skills in any specialized domain: take a few of the best textbooks on the discipline you're trying to encode, and a couple of experts who've lived it. Hand Claude a tool and skill design philosophy and a style guide. Iterate.
That's exactly how ours were built, and I'd put them against anything on the market. That's the point: the skills are now aluminum -- valuable, but now anyone can make them. Excellence stopped being a moat the moment it became reproducible.
The Premium Is Gone
Once everyone knows a competent team with Claude, a few textbooks, and a couple of experts can rebuild this layer, nobody pays a premium subscription for it. Not for anyone else's version, and not for ours. The specialized SaaS pricing model no longer applies here - open sourcing does not destroy that premium, it simply acknowledges that it is already gone.
The obvious alternative is to keep the skills proprietary, wrap them in “trusted AI” language, and spend the next few years convincing customers, investors, and ourselves that the moat still exists. Plenty of companies are making that bet, and some may succeed for a while.
We chose a different direction while it was still a choice. This layer is not the business. Give it away and build on what is actually hard.
The market is already moving this way. Many “data analysis agents” are little more than skills and MCP tools behind a proprietary wrapper - or bespoke agent harnesses, prompts, and patterns that felt novel nine months ago. These are bets against the emerging standard.
MCP and agent skills are where the ecosystem is concentrating its engineering. Each model release makes standards-based systems more capable and leaves proprietary abstractions further behind. The craft still matters, but the experts who thrive will build their advantage in the open, not by guarding the wrapper.
Trust but Verify
Coding agents handed the industry the thing it spent a decade dreaming about -- and now anyone can build the dream, chat your data & build a dashboard demo. Natural language in, answer out, chart on top. Everyone can build it, so everyone claims it, and every product page in the AI data space reads like every other one. When the demos converge, the claims converge, and marketing stops being a differentiator.
Which explains the one word you'll find everywhere: trust. Trusted analytics. Trusted data for AI agents. Trustworthy AI architecture. When you can't differentiate on what the product does, you reach for the one claim nobody can falsify. Underneath the word sits the same positioning every time: a black box you can't see inside and can't take with you -- trust us and lock in. It's the same tired SaaS model -- proprietary product, premium subscription, walled garden -- but based on “trust”.
We're taking the opposite approach. There is no magic. Here are our skills -- every one of them, readable, in the open -- shaped by a design philosophy we also published and distilled from a community whose expertise runs deeper than any proprietary bench. Read them, fork them, extend them with your own institutional knowledge, or tell us where they're wrong. Publicly.
Trust is earned with candor; it can't be demanded in marketing copy. Every vendor asking you to trust their black box is asking you to skip the only step that creates trust: looking inside.
Open Architecture Compounds
If marketing can't break the tie, product quality will. Most of the market is looking for quality in the wrong place -- the prompts, the wrapper, the surface. Quality lives in the architecture underneath, and over time, architecture will dictate everything above it.
AI has disrupted software architecture as much as the software itself. Coding agents have flooded the market with systems that demo well and fall apart in production -- AI slop, shipped as architecture. Most of it will crumble under the next wave of AI adoption. The architectures that scale and extend will survive. And each wave thins the field: the survivors get scarcer, the problems get harder, the wins get bigger and more valuable, and somewhere along the way they become durable businesses.
This is why we bet on Malloy - a strong foundation for the architecture: built and stress tested by a community, extended in directions we didn't plan, broken by workloads we didn't imagine, and fixed by people whose expertise runs deeper than any one company's bench. This ensures the end result is more flexible and more robust than anything we could harden alone.
A closed architecture ages at the speed of one team's payroll, whereas the benefits of an open one compound.
Meet Every User Where They Work
No company has one kind of data user. But every SaaS product is built as if it does: one surface, one login, one way of working, and everyone gets marched through the same door.
We built the opposite. Because the stack is open and conforms to the standards, you can use the same tools and skills in whatever environment best fits your workflow:
Self-host the open source. Download the server, the tools, and the skills and run the stack yourself. No account, no sales call. The software you get is the core of the architecture 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 the bundled app. The familiar turnkey package -- hosted, governed, with a world-class in-app experience for people who want to open one thing and start building and asking questions. The difference from every other SaaS app you've bought: what's underneath isn't a secret.
And whichever you choose, extend it. Start from our tools and skills and capture your organization's unique insights -- the edge cases, the business rules, the way your team works data. Your experts encode institutional knowledge as skills on the same architecture we build ours on, and bundle them into your plugin or app.
However you connect, it's the same governed model underneath. Your data engineer works the open source stack in her IDE. Your analyst lives in the app. Your ops lead asks questions through a plugin in the agent he already had open. Different personas, different packaging, one model -- all collaborating on the same meaning. Same components, same tools, same skills, remixed to meet each user where they are.
That's the promise to customers.
The promise to the community is the flywheel: every team that extends the skills and contributes back grows the commons -- expertise, encoded as skills, compounding into the package everyone builds from. No closed vendor can match it. They sell you one way in; we support them all -- switch anytime, or run several at once, and nothing underneath changes.
The Premium for Trust, Durability, and Distribution
We made one decision -- stop defending a layer that can't be defended -- and it bought us the three things no premium could: trust, durability, and distribution.
What we gave up was the premium. It was already gone. We will make it up in volume and scale: when the doors are open, far more people walk through them, and the architecture is built to carry them all. That's the “actually hard” thing we committed to building -- not the layer we gave away, but the platform underneath that has to carry everyone who shows up.
And this is the part that we are most excited about.
AI has dramatically reshaped the playing field, and open source came out of it with more leverage than it's ever had -- every model release amplifies shared code, shared tools, and shared expertise faster than any proprietary shop can match. That's the future we're betting on, and our mission is to realize it.
The piece this post hasn't covered is our business model itself: what we charge for, and why the same open strategy aligns our incentives with our customers' instead of against them. That's a later post.
In the meantime, we're proving the candor is real. The series starts with the engineering: what we shipped and the principles behind it, then a hands-on tutorial that gets an agent answering questions against real data in minutes. From there we publish the expertise itself, one theme at a time, following the life of a governed model in the agent era: how agents analyze without hallucinating, how they build the model, how they ship data apps, how the model scales, how you prove it works with evals, and how you govern what agents see. The open-sourcing continues on the same arc -- skills and the contributor template today, more of the stack landing in Publisher as the series runs.
The tools and skills are free now. The expensive thing was admitting why they had to be.
Start now: clone github.com/malloydata/publisher and open Claude in the folder, ask it "let’s get started! ". Ask data questions in plain English and it uses the skills and tools in the repo to find the right data and answers, no black box in between. Want it against your own data, or help turning your team's expertise into skills? Get in touch at credibledata.com.
