# We Connected to the MCP Servers in the Official Registry. Here Is What Answered

> Source: <https://www.stork.ai/blog/mcp-registry-endpoints-checked-november-2026>
> Published: 2026-11-16 23:34:15+00:00

## Why we did it[#](#why-we-did-it)

The registry tells you a server exists. It does not tell you whether the address still answers, how many tools it loads into your model, or whether those tools changed last week. It checks who published a name. It does not test what the server does. Every directory built on it, ours included, inherits the gap. So we stopped listing addresses we had never tried.

[Read the docs](https://www.stork.ai/developers/talking-head-api?utm_source=house_ad&utm_medium=inline&utm_campaign=api)

[No code? Make a photo talk](https://www.stork.ai/tools/talking-photo?utm_source=house_ad&utm_medium=inline&utm_campaign=api-second)

## What we counted in the registry[#](#what-we-counted-in-the-registry)

Reading the registry's public API on November 16, asking for the latest version of each server, we counted **65,580 entries**. **68%** of them carry a remote endpoint, an address you can connect to over HTTP. The rest are packages you run on your own machine, which we cannot test from outside.

## Result 1: who answered[#](#result-1-who-answered)

| One initialize request and one tools/list request per endpoint, sent from Stork's servers with no account, November 16, 2026. Rounded, so the column adds to 101. |  | 
|---|---|
| What happened when we connected | Share of endpoints | 
|---|---|
| Completed the handshake with no account | 54% | 
| Answered and asked for a sign-in | 36% | 
| Did not answer | 10% | 
| Opened an older event-stream connection only | 1% | 

Two readings of that table matter. First, **nine in ten addresses are alive**, which is better than the "graveyard" people describe. Second, a third of the catalog is closed to anyone without an account. That is normal for a company's own server, and it means nobody outside can tell you what tools it offers until you sign in.

## Result 2: what a connected server costs you[#](#result-2-what-a-connected-server-costs-you)

When a client connects to an MCP server, it loads the definition of every tool into the model's context, before you ask anything, unless the client loads tools on demand. We measured the size of those definitions for every server that listed its tools.

| Estimated from the length of the tool definitions at 4 characters a token. On four servers we compared against a real tokenizer the true figure was 4.1 to 4.65 characters a token, so these run slightly high. |  | 
|---|---|
| Point in the range | Tool definitions, estimated tokens | 
|---|---|
| Lightest tenth, at most | 360 | 
| Median | 1,400 | 
| Heaviest tenth, at least | 8,100 | 
| Heaviest hundredth, at least | 40,900 | 
| Largest we saw | 576,000 | 

- **39%** of servers cost under 1,000 tokens. Most servers are cheap.
- **8%** cost more than 10,000 tokens each.
- The median server lists **7 tools** . The busiest tenth list 28 or more.
- One server in a hundred takes about **41,000 tokens** , a fifth of a 200,000-token context, for its definitions alone.

The spread is the finding. Two servers that do the same job can differ a hundredfold in what they cost to keep connected, and until now no listing told you which was which.

Enjoying this? Get one like it in your inbox each morning.

one email a day · unsubscribe in two clicks · no third-party tracking

## Result 3: addresses that look alive and are not[#](#result-3-addresses-that-look-alive-and-are-not)

Smithery [switched off its hosted server addresses](https://smithery.ai/future) on October 31, 2026. More than two hundred registry entries still point at those addresses. They are easy to mistake for working servers, because the old host still answers with a sign-in challenge. We removed them from our search results. If one of them is in your config, [this guide](https://stork.ai/blog/smithery-shutdown-mcp-server-urls) shows how to find the server's own address.

## What this does not tell you[#](#what-this-does-not-tell-you)

- **Whether a tool works.** A check is a handshake and a tool listing. We never call a tool.
- **Whether a server is safe.** Nothing here inspects what a tool does with your data.
- **Anything about local servers.** Packages you run yourself were not tested.
- **A trend.** This is one pass. We repeat it weekly for servers that answer, and the next report will say how many changed their tool definitions without changing their version number.

## How to use the results[#](#how-to-use-the-results)

- **On the web:** every listing in[Stork's MCP directory](https://stork.ai/mcp) shows the result of our last check, the date, the number of tools and the estimated context cost.
- **From an agent:** connect to`https://mcp.stork.ai/mcp?tools=directory` . It is a free MCP server with five tools and no key. Ask it for a server by task and each result says how to connect, whether the address answered, and what it costs. It can also filter: only servers that answered, or only those under a token budget you set.

That endpoint's own tool definitions come to about 940 tokens by the same measure. It seemed wrong to report everyone else's number and hide ours.

## Frequently asked questions[#](#frequently-asked-questions)

### How many MCP servers actually work?[#](#how-many-mcp-servers-actually-work)

Of the remote endpoints we connected to on November 16, 2026, 54% completed an MCP handshake with no account and 36% answered and asked for a sign-in. Ten percent did not answer. That measures whether the address responds, not whether its tools do their job.

### How many tokens does an MCP server use?[#](#how-many-tokens-does-an-mcp-server-use)

The median server's tool definitions are about 1,400 tokens. The heaviest tenth are above 8,100 and the heaviest hundredth above 40,900. Clients that load tools on demand avoid most of that cost.

### How did you estimate tokens?[#](#how-did-you-estimate-tokens)

We took the name, description and input schema of every tool a server listed, and divided the character count by four. Against a real tokenizer on four servers that estimate ran 3% to 16% high.
