There Are 40,000 MCP Servers. How Does an Agent Find One That Works and Takes Payment?

MCP, the Model Context Protocol, is the standard way for AI apps to plug into outside tools. Its official directory, the MCP Registry, is growing very fast. In November 2025 it held "close to two thousand" servers. On October 6, 2026, the team at host0 counted 40,047, about 20 times more in under a year.

That's great news for AI agents. More tools means more things they can do. But a bigger list isn't the same as a better one. In the first week of October, four separate teams checked the registry. Together they show three problems an agent runs into: finding the right tool, finding one that actually works, and paying for it.

What the registry is (and isn't)

The registry is a list of names, descriptions, and addresses. It doesn't store the tools themselves, and it doesn't test them. As Few Servers points out, the registry's own docs call its data "deliberately unopinionated" and say it's meant to be used mainly by other directories that add their own curation.

That's a reasonable design. A neutral phone book is useful. It just means someone else has to answer "is this any good?"

What four crawls found

Who checked	When	What they counted	Key finding
Few Servers	Oct 1	38,212 servers	62% are remote (you connect by URL). 42% of current versions were published in September alone.
openforallofus	Oct 5	39,734 listed; 23,119 remote endpoints tested	61% answered properly, 28% asked for sign-in, 63 asked for payment, 2,381 failed.
host0	Oct 6	40,047 servers	The 10 biggest publishers make up 18.7% of entries. One account alone has 2,365.
CrawlCheck	Oct 7	25,983 remote endpoints (registry at 40,721)	59.9% worked, 24.7% wanted credentials, about 1 in 7 were broken or weren't MCP at all.

The counts differ a little because each team crawled on a different day and counted slightly differently. The pattern is the same in all of them.

Problem 1: finding the right one

Forty thousand entries sounds like plenty of choice. But a lot of that list is noise. Few Servers found 158 entries with the same one-line description and 22 that still say "Description of my MCP server," the default text from a template. A small number of accounts publish thousands of near-copies. After filtering out empty and mass-produced entries, Few Servers kept about 29,700.

For a person, scrolling past junk is annoying. For an agent, it's worse. It may pick the first match that looks right.

Problem 2: finding one that works

Only about six in ten remote servers answered a basic "hello, what tools do you have?" check, according to both openforallofus and CrawlCheck. About a quarter wanted a login first. And roughly one in seven gave an error, timed out, redirected to a marketing page, or wasn't an MCP server at all.

A server can be down for an hour and fine the next, so these are snapshots. Still, for an agent working on a task, a broken tool is wasted time and wasted tokens.

Problem 3: paying

This is the biggest gap. Out of 23,119 remote endpoints openforallofus tested, only 63 asked for payment. Almost every server is either free or locked behind an account that a person has to set up.

That's not because nobody wants to charge. It's because MCP itself doesn't include a way to pay. The big July 2026 update to the standard, the 2026-07-28 spec, made servers simpler to run and added headers so that "gateways can route and authorize on headers directly." It didn't add payments. That job is left to tools around MCP, like HTTP 402 payment standards (L402, x402 and MPP) and products such as Cloudflare's new Monetization Gateway.

The sign-up wall is the real cost here. If a useful tool needs an account, an API key and a credit card, an agent can't use it without a person stepping in.

A hidden cost: every tool you add takes up space

There's one more catch. Each tool an AI app connects to comes with a description the model has to read before it does anything. That uses tokens, the small chunks of text AI models charge by. host0 cites measurements where the GitHub MCP server's old default set of 101 tools used about 64,600 tokens before a single question was asked.

So "just connect lots of servers" isn't free. Every extra server makes every task a bit more expensive.

What a storefront does differently

This is why we built Amazap as a storefront rather than another list. An agent connects to one MCP server and gets a whole catalog of paid tools behind it: about 1,200 listings, including about 140 L402 listings an agent can pay for over Lightning.

Here's what that changes for each problem:

• Finding: the agent searches the catalog by what it needs ("web page to markdown," "Lightning node data") and gets back matching listings, each with one all-in price.
• Working: every purchase starts with a free quote. If a paid call fails, the price goes back to the agent's balance as credit, automatically when the failure is caught, or by claim within 72 hours. Listings that keep failing after payment, or draw buyer reports, are paused and hidden from search.
• Paying: the human tops up one balance once. The agent buys single calls from it, inside per-call and per-day limits, and every buy has a hard price cap. No account per provider, no API keys.
• Space: in our own test on October 11, all 19 of Amazap's MCP tool descriptions together came to about 1,100 tokens (measured with OpenAI's o200k tokenizer).

A real example: our own Page to Markdown tool costs 13 sats, about a cent. In our 20-page benchmark, it turned a typical web page from about 107,500 tokens of raw code into about 1,600 tokens of clean text.

You can see how the whole buyer flow works in Agentic payments, explained.

To be fair about it

A storefront isn't the only answer. Curated directories like PulseMCP pick servers by hand. Cloudflare says it plans to help sellers get found by agents. CrawlCheck and Few Servers both publish free checks you can run before you connect. And if you only need one free tool from a company you trust, the registry plus that company's own server is often the simplest path. Few Servers' first tip is exactly that: start with the vendor's own server.

What a storefront adds is the part none of these do alone: a checked, priced, paid-per-call catalog an agent can buy from without a person filling in forms.

A quick checklist for picking an MCP server

Whether you use a storefront or not, these habits help (adapted from Few Servers and CrawlCheck):

1. Prefer the vendor's own server for a known product.
2. Check there's a code repository. About a quarter of registry entries have none.
3. Look at what it asks for before you hand over a login.
4. Treat version 0.x as early. Four in ten entries are still there.
5. Watch the token cost of the tools you load.
6. Know the price before the call. If a tool costs money, the price should be clear up front.

If you sell a tool

List it in the registry. That's how many agents and directories discover servers. Then, if you want to get paid per call, put it behind a 402 payment step and list it where agents shop. Our Sell to agents guide shows how to do that in three steps.