MCP servers by category

Every MCP directory sorts by something. Most sort by ad-hoc headings that differ from directory to directory, so the same server lands under "AI Tools" in one and "Productivity" in another, and you cannot compare anything.

This is a snapshot of 15,473 published MCP servers, taken 2026-08-25, sorted into a closed 22-term taxonomy. One term per server, chosen by what its tools do rather than by who published it. The counts and the alive column are both from the same build: 5,850 are alive.

Read the third column against the second before you draw any conclusion from the first. browser-automation has 249 servers and 184 of them are alive, a 74 % survival rate. data-analytics has more than twice as many servers and fewer alive. Size of a category and health of a category are not the same measurement.

The biggest category is the one we could not classify

other is 21.5 % of the corpus and 3,330 servers, and only 146 of them are alive. That second number is a warning about the evidence behind many other rows.

About 48 % of the public MCP registry exposes a name and nothing else. No tool list, no repository topics, no package coordinate, frequently no repository. There is nothing to read, so there is nothing to classify. We forbid the classifier from labelling those rows, and it says other rather than guessing.

Earlier builds of this dataset reported 7.5 %, then 8.6 %, then 21 %. Every one of those numbers looked healthier than the one after it and every one was less true. The rate rose because we stopped letting a model fill in blanks, not because the taxonomy got worse.

Two attempts to push the number back down both produced plausible-but-wrong data. The second attempt corrupted descriptions as well as categories: an MCP server that wraps Hacker News came back described as agent-based AI workflows. Pressure to empty one column manufactures false facts in the next one.

If a directory shows you a clean category on every single row, that is worth a moment's thought about where the labels came from.

The second biggest is doing too much work

ai-ml is 2,374 servers, and it is carrying two different things.

The first is what the name suggests: model gateways, embeddings, RAG plumbing, prompt tooling. The second is agent infrastructure, meaning MCP registries, proxies, routers, memory and context stores, skill catalogues. Those are not machine learning. They are the plumbing of the protocol itself, and there was nowhere else to put them.

Splitting agent-tooling out is the obvious next revision and it is deferred rather than forgotten, because doing it honestly needs a full re-classification of every row. A term added halfway through a run is only ever offered to the rows still being classified, and it lands under-populated and misleading. So it waits.

What the shape of this actually says

Four things stand out once you sit with the table.

Developer tooling is not the story anymore. dev-tools, version-control, ci-cd, databases, cloud-infra and observability together are about 14 % of the ecosystem. The protocol started as a way to give a model access to your development environment, and the published ecosystem has already moved past that.

Real-world domain wrappers are the growth. finance-payments, commerce-retail, legal-compliance, geo-weather and health-science are 3,471 servers between them, nearly a quarter of the corpus. These are thin agent-facing shells over an existing domain API: a statute database, a parcel tracker, a geocoder, a clinical reference. That category did not exist in our first taxonomy at all, and having to invent four terms for it was the clearest signal of where this ecosystem is going.

legal-compliance is inflated by one publisher shipping dozens of per-country statute servers, which is a good reminder that a category count is a count of publications and not of distinct ideas.

And the smallest categories are the healthiest. browser-automation, cloud-infra, security and version-control all sit at or near a 50 % alive rate, well ahead of the corpus average of 35.8 %. Narrow, well-defined problems attract people who finish things.

How the categories are assigned

A local model reads the evidence we hold for each server and answers against the closed 22-term list. The answer is then validated by script against that list. An answer outside the enum is rejected and re-asked, never coerced into the nearest term, because a coerced label is a guess wearing the costume of a fact.

It is one of only two fields in the whole dataset a model touches. Everything else is computed by script from collected evidence, and the rules are published.

The data

Category, alive status, licence verdict, maintenance date, auth shape and install method for all 15,473 servers: MCP Server Landscape 2026, €4 snapshot or €10/month API. Free 50-row sample, every field, no email gate.

Filter it yourself.