An MCP server used 42 times the tokens of a CLI on one task

OpenRouter tested 3 types of tool for an AI agent on one GitHub task. They are an MCP server (a standard connection to external tools), a command-line tool (CLI) and a skill (a set of instructions). On Claude Sonnet 5.5, all 3 were correct, but MCP used 59,417 input tokens (the text units that you pay for) and the CLI used 1,421.

How it works

Each tool sends its definition to the model each time. The cost grows with each tool and changes with the model. The 46 default GitHub tools add 19,436 tokens on Claude Haiku 5.5.

The numbers

• MCP cost 17 times the CLI: $127 against $7.50 for 1,000 tasks.
• OpenRouter states that Tool Search cut the start cost of 46 tools to less than 600 tokens.

What is not known

The test used 1 task and the median of 3 runs. Tool Search halved the total cost on Sonnet but raised it on Haiku. OpenRouter also states that failures came with Tool Search.

Insight: MCP costs tokens, not time

A paper states that MCP adds 0.39 ms or less on 1 machine. OpenRouter found that it adds more than 10,000 tokens. BeyondTrust states that 85% of the coding agent users it saw had an MCP server. MCP costs most in the context window. OpenRouter sells a solution, which reads as a move to control tool loading. Measure tokens for each tool before you add it.