An MCP server costs you tokens before you ask it anything. The client sends the model every tool definition, name, description and schema, and those definitions take space in every request of the session, used or not.

Last month Sarvar Nadaf wrote a good post about this from the inside: he cut one server from 15 tools to 4 and measured what changed. I happen to have the other half of that picture. Not one server measured closely, but 35,633 of them measured the same way.

The short version: the typical server is cheap, and the average is a lie. The median is 2,518 tokens. The average is 8,059, more than three times that, because of a long tail.

How I counted

I run a probe on the official MCP registry, mcpcost. Once a day it connects to every remote server and asks for tools/list. Servers that ship as a package get started in a container once a week. It keeps every answer.

For the count I did not want an estimate, so I use Anthropic's count_tokens endpoint. It returns the number of input tokens the model will see and runs no model. I send one request with the server's tools plus a tiny marker tool, and one request with the marker only. The difference is what the server's tools cost. The model is Claude Opus 5.5.

In the run of 14 November 2026, 35,633 servers listed their tools and got a count.

The middle is fine

The median server costs 2,518 tokens. That is 1.3 % of a 200k window. Five median servers together take about 12,600 tokens, around 6 %. You can live with that.

Half of all servers have 8 tools or fewer. More than one in five costs under 1,000 tokens.

If the story ended here, "MCP eats your context" would be a myth.

The tail is not

It does not end here.

Tokens
Median 2,518
90th percentile 17,043
99th percentile 54,779
Largest 1,633,624

One server in ten costs more than 17,000 tokens. Put five of those in a config and over 40 % of the window is gone before the first prompt.

3,167 servers cost more than 20,000 tokens. 133 cost more than 100,000. And 45 cost more than 200,000, so their tool list alone does not fit in the window they are supposed to work in.

The largest one has 73 tools and costs 1.6 million tokens. That is eight full context windows of tool definitions.

The top 1 % of servers, 356 of them, hold 16 % of all the tokens in the registry.

It is the number of tools. Almost always.

A server can be heavy for two reasons: it has many tools, or its tools are fat. I checked which one it is. It is nearly always the first.

Tools on the server Servers Average cost Per tool
1 to 3 7,236 872 425
4 to 10 13,932 2,394 375
11 to 25 7,090 6,577 402
26 to 50 5,941 21,314 548
51 to 100 979 33,160 520
over 100 455 91,804 428

Look at the last column. It hardly moves. The median server pays about 320 tokens per tool, and nine servers in ten pay under 855, whether they have three tools or three hundred.

So there is a rule of thumb that works surprisingly well: count the tools, multiply by 320.

It also puts a number next to something from Sarvar's post. The AWS guidance he quotes says a tool schema runs 550 to 1,400 tokens. In the registry most servers pay less per tool than the bottom of that range.

There are exceptions, and they are extreme. 293 servers average more than 2,000 tokens per tool. The 1.6-million server is one of them: 22,000 tokens per tool. I have not read all 73 of its schemas. I am not sure anybody has.

Your estimate is probably low by 1.75x

Before I had exact counts, I estimated. I serialized each tool as compact JSON and counted it with the o200k_base tokenizer, which is what a lot of scripts do.

Then I ran both on every server. The exact count is 1.75 times the estimate at the median. Not a few percent off. Almost double. And it is consistent: for eight servers in ten the ratio sits between 1.62 and 1.87.

So if you measured your tool list with tiktoken and felt fine, multiply by 1.75 and feel again.

One smaller thing. Claude Code renames every tool to mcp__<server>__<tool>. I count that variant too. It adds about 3.5 % at the median, 89 tokens. Not nothing, not a problem.

What these numbers do not tell you

This is the part I would want to read first if somebody else had written this.

Only servers that answer without credentials are in it. 13,140 servers in the registry ask for auth, and I cannot list their tools. That group includes many of the hosted servers of big vendors. The GitHub server from Sarvar's post, 55,000 tokens across 93 tools, is one I cannot check. It works out to about 590 per tool, which fits what I see elsewhere, but I did not measure it.

One publisher bends the curve. There is a bump between 10,000 and 20,000 tokens that looked like a real pattern. It is 2,343 registry entries from a single publisher on a single host, each between 12,600 and 13,200 tokens. With them, 22 % of servers cost over 10,000. Without them, 16 %. Both numbers are true. Pick the one that matches your question.

It is the worst case, not always the bill. Some clients now load tool definitions on demand instead of all at the start. For those, these numbers are what you would pay if everything loaded. And it is definitions only. What a tool returns can cost far more, and I do not measure that at all. Sarvar's post is the better read for that side.

One model. These are counts for one Claude model. Another model has another tokenizer and wraps tools in its own way. Do not carry the numbers over.

And I got it wrong for a week. In October the counting endpoint started to refuse my API key. My code stored each refusal as "this server's schema was rejected". So for seven days the board blamed servers for a fault that was mine, and it took me days to see it. The lesson was boring and useful: a refusal of the key and a refusal of the schema need different names in the code.

What I take from it

Three things.

Count tools before you count tokens. Eight tools is normal. Forty is a decision that somebody should make on purpose.

Do not trust an average in this ecosystem. Median and 90th percentile, side by side, or nothing.

And the fix sits with the publisher more often than with the protocol. Sarvar's rebuild from 15 tools to 4 is the proof for one server. The table above says the same thing for 35,000.

I am curious about the other side of this. If you publish a server with more than 50 tools: do your users need all of them loaded at once, or is that just how the SDK made it easy? And if you have measured your own setup, how far off was your first guess?