16 303 servers counted · 2026-10-15

How many context tokens does an MCP server cost?

The median server in the official MCP registry costs 2 737 tokens of context per session in Claude Opus 5.5. This is the cost of its tool definitions only. The 90th percentile is 13 050 tokens. The 99th percentile is 49 714 tokens. The heaviest server costs 944 524 tokens. That is more than 4 full 200k windows. In Claude Code the median is 2 832 tokens, because each tool name gets an mcp__ prefix. We counted 16 303 servers with the Anthropic count_tokens endpoint on 2026-10-15.

Why does an MCP server cost tokens before I send a prompt?

A client sends the model the name, the description and the input schema of each tool. It does this for each connected server at the start of a session. It does this again in each turn. These definitions are part of the prompt. You pay for them in each request. This is true also when the model calls no tool.

How do we count the tokens?

We ask the Anthropic count_tokens endpoint. It returns the exact number of input tokens for a request, as the model will see it. We send one request with the server's tools and one request without them. The difference is what the tools cost. The endpoint is free and calls no model. The number is exact for Claude Opus 5.5. Other models count in a different way. We also show an o200k_base estimate over the compact JSON of each tool. The exact count is 1.76 times the estimate at the median (1.63 to 1.85 for 80 % of servers).

What does Claude Code add?

Claude Code names each tool mcp__<server>__<tool>. The longer name costs more tokens. We count that variant too. The median server costs 2 832 tokens in Claude Code against 2 737 with the published names. Claude Code also adds a tool-use system prompt one time per request. For Claude Opus 5.5 it is 286 tokens. When MCP tools take more than 10 % of the context, Claude Code can defer them and load them on demand through tool search.

How many tokens does one specific MCP server cost?

Each remote server in the registry, and each npm or PyPI package server, has a page at https://mcpcost.ulehla.dev/s/<registry name>. The page shows the exact cost in Claude Opus 5.5, the cost in Claude Code, the estimate, the tool count and the status. For a server that is not in the registry, https://mcpcost.ulehla.dev/check probes it now and counts it the same way. In a terminal, npx mcpcost measures the servers in your own Claude Code, Claude Desktop, Cursor, VS Code or Windsurf config. With ANTHROPIC_API_KEY set and --claude, mcpcost asks the same endpoint.

What is a good budget?

Half of all servers cost less than 2 737 tokens. A config with a total of a few thousand tokens is cheap. If one server costs more than 20 000 tokens, ask if you need all of its tools. 20 of the 20 heaviest servers are above that limit. The command npx mcpcost --budget 30000 fails a CI run when a config grows past the budget.

Does a server that needs auth also cost tokens?

Yes. After you authenticate, it sends its tool definitions like any other server. The board cannot list the tools of a server that refuses clients without credentials. These servers show “needs auth” and no count. mcpcost measures them with the headers from your config.

The heaviest servers today

All servers by token cost → · Statistics → · Method →