MCP servers consume context window space through tool definitions, unlike on-demand CLIs

A developer found that a single prompt sent 28,400 input tokens, largely due to seven MCP servers whose tool definitions were loaded before any work began. MCP servers announce their tools via tools/list, and clients include each tool's name, description and input schema in every turn, so costs scale with tool count, schema depth and conversation steps. In one measurement, four common MCP servers added about 11,923 tokens of definitions, while a Apidog CLI skill added 112 tokens. Anthropic's documentation reports that a multi-server setup can consume around 55,000 tokens before any work and that tool selection accuracy degrades beyond roughly 30 to 50 tools. The author's figures used the OpenAI o200k_base tokenizer, so exact counts for other models will differ, though the ratios between configurations are the focus.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in