TLDR Tech / Practical
How much MCP token overhead you pay depends on the harness
An analysis of Model Context Protocol (MCP) token overhead reveals that whether tool definitions consume context on every request depends entirely on whether the agent harness supports lazy loading or tool search. Testing four real servers with an OpenAI tokenizer showed that default tool definitions range from thousands to tens of thousands of tokens, dominated primarily by JSON schemas rather than descriptions. The author argues that MCP overhead is comparable to CLI help output per operation, but differs fundamentally in eager versus lazy loading mechanics.
Read the full story