Skip to the content

DocsConcepts

The agent directory

How the directory of MCP servers, systems and skills is made: its sources, what first-party means, and what nobody has checked.

The agent directory lists what an agent can connect to a system with: MCP servers, command-line tools and the vendor's own skills, by system. This page says where each fact comes from and what it does not tell you.

Where the data comes from

FactSourceRead
Servers: name, description, version, address, packagesThe official MCP Registry, as its publishers wrote themDaily, whole
Systems, and which server is a system's ownA list we keep by handWith each release
Weekly downloadsnpm's downloads API and pypistats.orgDaily; rounded to two figures
Known vulnerabilitiesOSV, by package and versionDaily
Stars, last push, licenceGitHubDaily, where read
SkillsThe vendors' own repositories on GitHubDaily

Each page shows the date a fact was read. A date can be older than today: the pages are rebuilt when something in the catalog changes, not every day.

Which servers are listed

The registry lists tens of thousands of servers. This directory keeps the ones that are for a system it covers: a server is listed for a system when its registry name or title names the system as whole words, when it is published under the system's own domain, or when we have named it as the system's official server. A server that only mentions a system in its description is not listed for it.

A server that is its system owner's own, or that names a repository or a package, gets a page. One by another publisher that names neither is shown by name on its system's page and has no page of its own, because there is nothing else to say about it. A server the registry marks deprecated keeps its page, marked; it is left out of the index of servers and comes last on its system's page.

What the publisher labels mean

LabelMeans
First-partyPublished by the system's owner. A server gets this label in one of two ways: we named it as the system's official server after reading its registry record and the repository it names, or it is published under the owner's own domain, which the registry verifies. Most official servers are published under a GitHub organisation's name, which shows who can log in, not who owns a product, so a GitHub name alone never earns the label.
SourcedNames a public repository or a package. Anyone can publish one. We have not read its code.
ListedIn the registry with no repository and no package named. Nothing else is known.

These are not grades. A first-party server can have a known vulnerability, and a sourced one can be excellent.

The order of a list

There is no score. Servers are listed live before deprecated, then first-party first, then those with no known vulnerability before those with one, then by weekly downloads, then by the most recent push, then by name. Every one of those facts is in the row, so you can see why a server is where it is. Skills are listed by vendor, most recently changed first.

What is not checked

  • Vulnerabilities are what OSV lists for the exact npm or PyPI version a server names. "None known" means OSV named nothing on that date. A container image, a bundle or a package with no version is not checked at all, and the page says so.
  • Tools are not read. This directory does not connect to servers, so it cannot say which tools one offers or how much room their definitions take in a context window.
  • Auth is what a server's registry record declares. A server that declares no header may still ask you to sign in.
  • Names, descriptions and addresses are the publisher's own. They are shown as plain text, cut to a fixed length.
  • Skills are listed only from a vendor's own repository. A skill's weight is its SKILL.md counted with OpenAI's o200k_base tokenizer; other tokenizers count differently. A skill that ships scripts carries code an agent may run.

Something is wrong

If a server is listed for the wrong system, an official server is missing, or a page says something untrue, tell us through the contact page. To use Rhumb itself from an agent, see Use from an agent.