Registry search matches a substring of the server name and returns matches sorted by that name. So the useful question about a candidate name is not whether it is taken, it is how many rows already match the word you want and where your namespace would land among them. Both are one API call.
curl -s 'https://registry.modelcontextprotocol.io/v0/servers?search=<token>&limit=100&version=latest' \
| python3 -c 'import json,sys; d=json.load(sys.stdin); print(d["metadata"]["count"])'version=latest matters. Without it the endpoint returns one row per published version,
not one per server, and the count is meaningless. One publisher in this registry holds 739 version
rows for a single server.
"Servers" is the count at version=latest; a plus sign means the page of 100 was full
and the true count is higher. The last three columns are the position of the first row in each
namespace family, which is what your own rank would be bounded by.
| Token | Servers | First ai. | First com. | First io.github. |
|---|---|---|---|---|
| github | 100+ | 1 | 4 | 10 |
| search | 100+ | 1 | 46 | none on page 1 |
| memory | 100+ | 1 | 13 | 40 |
| agent | 100+ | 1 | none on page 1 | none on page 1 |
| crypto | 100+ | 1 | 9 | 29 |
| 100+ | 1 | 4 | 28 | |
| 93 | 1 | 5 | 27 | |
| late | 84 | 1 | 4 | 22 |
| weather | 79 | 1 | 3 | 15 |
| invoice | 73 | none | 32 | 46 |
| browser | 57 | 1 | 5 | 14 |
| aws | 49 | 1 | 5 | 10 |
| calendar | 38 | 1 | 7 | 12 |
| postgres | 34 | 1 | none | 15 |
| database | 25 | 1 | 5 | 12 |
| figma | 22 | none | 1 | 3 |
| schedule | 22 | none | 1 | 8 |
| docker | 21 | 1 | none | 3 |
| slack | 20 | 1 | 3 | 5 |
| delivery | 18 | none | 1 | 4 |
| jira | 16 | 1 | 2 | 3 |
| stripe | 13 | none | 1 | 3 |
| notion | 12 | 1 | 2 | 4 |
| filesystem | 11 | none | 1 | 2 |
| excel | 11 | none | 2 | 4 |
| sqlite | 9 | 1 | none | 3 |
| linear | 8 | none | none | 2 |
| milestone | 3 | none | 1 | 2 |
| kubernetes | 2 | none | none | 1 |
The spread is two orders of magnitude. kubernetes returns 2 servers
and github fills a page of 100 with the cursor still set. A word that feels equally
obvious to you can be uncontested or hopeless, and there is no way to guess which.
An ai. namespace holds first place on 19 of the 29 tokens. That is
not a quality signal, it is the alphabet. Anything published under ai.something sorts
before every com., dev. and io. row that matches the same
word.
The crowded tokens are crowded at the top. On search the first
com. row is 46th and no io.github row reaches page one at all. Choosing a
narrow word is the only move available to a publisher who cannot change namespace, and it works: on
milestone, 3 servers match and everybody is visible.
delivery and
milestone appears in an 18-row list and a 3-row list.What no naming choice can fix is the namespace prefix, which is compared before your first character. That is measured separately in how MCP registry search actually works.
Source: the live registry API at registry.modelcontextprotocol.io/v0/servers, one
call per token at limit=100&version=latest on 2026-09-09. Every result set was
checked against its own byte-order sort and every one matched. This project publishes its own catalogue of
MCP servers and chose their names before measuring any of this, which is how the gaps in the table
were found.
That is the maximum page size. When metadata carries a next cursor the real count is higher, and you have to paginate to learn it. For naming purposes a full page already answers the question: the token is crowded.
It means few names contain it. Nothing stops someone else publishing the same word tomorrow, and nothing reserves it for you. The count is a measurement of a moment, which is why it is worth rerunning before you commit.
Both are near-universal, so they add nothing to a search that is already restricted to MCP servers, and they consume characters. The one case for them is a package name that will also appear on npm or PyPI, where the context is missing.
Not through this endpoint. Search matches the name only. A description is what a person or a model reads once you are on the list, which makes it decisive for selection and irrelevant for retrieval.
The registry is growing, so counts drift upward and ranks drift downward. A measurement filed on the registry's issue tracker on 2026-09-07 put the whole table at 28,139 entries. Rerun the call rather than quoting this table back at yourself in six months.
All MCP servers and prices · All guides · Buy the bundle $39