whobuildsagents.com is live. It answers one question: which companies are actually building for AI agents? Today that is 16,990 companies publishing 21,739 MCP servers, ranked by how many each one ships, refreshed every day from the Official MCP Registry.
Why build another directory
It started with a scraped LinkedIn profile. A people directory had copied a teammate's profile onto a page about our company, and the only way to fix it was to sign up with a company email it would not accept. Directories of people are built on data the people never gave them.
A directory of companies does not have that problem if the data is already public and published by the company itself. The MCP Registry is exactly that: every server in it was put there by someone who proved control of a domain. So the rule for Who Builds Agents was set before any code: companies only, never people, and every fact on a page comes from the registry or the company's own homepage.
A registry name is a verified domain
Registry names are reverse DNS. com.example/mcp can only be published by
someone who proved they control example.com. That makes the namespace a
free, already-verified company identity, so the directory rolls every server
up to its registrable domain: com.example.docs/mcp lands on example.com
too.
Two kinds of name are skipped on purpose. io.github.<user> is usually one
person, and listing people is the thing we do not do. Platform subdomains
(the kind the Public Suffix List marks as private, like x.workers.dev) are
somebody's side project, not a company.
Claim and remove without emailing anyone
A company can claim its page or remove it outright. Both work the same way:
enter an address at your own domain, get a code, done. Owning a mailbox at
example.com is cheap to check and expensive to fake at scale, which is
also why it is the identity we plan to build votes and submissions on.
Removal is permanent: the daily ingest never re-adds a suppressed domain.
Written for agents as much as people
Most of the traffic a directory like this earns will come through answer engines, so the machine surfaces are first class. There is a live llms.txt with the current numbers, every page carries JSON-LD (Organization, ItemList, FAQPage, Dataset), the data is CC0 like the registry it comes from, and pages are submitted to search engines through IndexNow. Crawlers, AI ones included, are allowed everywhere.
22 categories, no model call
Every company is filed into up to three of 22 categories, from payments to developer tools, using keyword rules over its server descriptions and homepage. It is free, deterministic and reruns in about three seconds after each daily ingest. A model would be more accurate, and it would also spend shared credit every night on a job that is already roughly 80% right. The rules are tested against the false positives we found by hand, like "stock footage" landing in finance.
Counting is the hard part
Ranking by servers published is objective, and it is also easy to game. Two things surfaced in the first week.
Hosts are not builders. Smithery and usefulapi publish hundreds of servers that wrap other companies' products. They keep their place in the ranked list, but their pages say "hosts" rather than "publishes", and they are left out of categories and of any "who builds the most" answer.
Listed is not shipped. Anyone can publish a hundred concept entries to the registry for free. Rather than keep a hand-made list of noisy publishers, which would make the ranking an editorial opinion, the next step is to count only servers that are real: a remote endpoint that answers an MCP handshake, a package that resolves on npm or PyPI, or a repository that exists. Pages will show both numbers. Paid placement, when it comes, sits in its own labelled slot and never changes rank.
If your company publishes MCP servers, it is probably already listed. Look it up and claim the page.