What 45,156 MCP servers tell you, and what they don’t.

Ask how many MCP servers exist and you will get a confident number. Most of those numbers are counts of something else: registry listings, package versions, repositories that mention the protocol. Marklode wanted the number underneath, so we counted it, and we want to be just as clear about what the count can carry.

How the count is made

The Model Context Protocol (MCP) lets AI agents call out to tools and data through servers. Public registries list those servers. Marklode reads the public registry listings, then deduplicates them into the distinct servers behind them: the same server published under several versions, names or entries counts once.

It is metadata only. No MCP server is ever contacted, probed or pinged. We read what was published about each server, and nothing else. That is a deliberate limit, and it shapes everything below.

What the count says

The October 2026 recount read 156,111 registry entries. Deduplicated, they come to 45,156 distinct servers. The first census, on 2026-09-12, found 31,309 servers behind 101,219 entries.

Three things follow. First, entries are not servers: a registry count overstates the field several times over, and any headline built on raw listings is counting versions and duplicates. Second, the field is moving fast: the distinct-server count rose from 31,309 to 45,156 between the first count and the recount. Third, a large share of servers can’t be inspected at all: 13,383 of them, 29.6%, have no public source repository.

That last figure is the one we would put in front of anyone wiring an agent to a server they found in a registry. Nearly three in ten servers offer no public code to read. That does not make them unsafe. It means you can’t check, and checking is the habit we are trying to build.

What the count does not say

It does not say any server is safe, or unsafe. We never ran one, so we can’t tell you what any of them does.

It does not say how many servers are maintained, used, or used by anyone at all. A listing is a claim that something exists, not evidence that anyone depends on it.

It does not rank servers or vendors, and money never buys a place in it. It does not say a server without a public repository is hiding anything; plenty of honest software is closed source.

And it is a snapshot. Registries change daily, which is why the census is recounted monthly and every figure on this site carries the date it was checked.

Why a diligence company counts registries

Marklode’s day job is startup due diligence before the pitch: who else is doing this, what got there first, the three strongest attacks on the idea, and a verdict on what survives. The census is the same discipline pointed at a public question. Count the thing itself, not its shadow. Say what the count supports. Say, just as plainly, what it doesn’t.

That is our third Flight Rule, If it matters, make it provable. Every figure in this note is an entry in the Record, with its source and the date we last checked it. If a number changes, the change is logged there, not quietly edited here.

If you build with MCP

Before you connect an agent to a server you found in a registry, ask three questions the census can’t answer for you. Is there code I can read? Who maintains it, and when did they last change it? What can it reach once my agent calls it? If the answer to the first is no, the other two get harder, and the guardrails around your agent matter more. That is where the rest of our fleet comes in.