MCP Servers Can Fill Your Context Window Before You Type a Word

An article compares the cost of MCP servers and CLI tools, explaining that MCP servers send every tool's name, description, and JSON Schema to the model before any prompt is entered. A developer reported a first request consuming 28,400 input tokens from seven MCP servers, with no files or history attached. The author measured four common MCP servers adding about 11,923 tokens of tool definitions, compared with 112 tokens for an Apidog CLI skill. Costs grow with tool count, schema complexity, and repeated sends across each agent loop turn. The piece notes that prompt caching reduces monetary cost but not the context space or attention these definitions consume.
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