How MCP Wastes 4-32x More Tokens Than CLI (and How to Fix It)

Here are the key facts from the story about MCP wasting more tokens than CLI, and how to fix it:

- 71,929 tokens are used by MCP to load 255 tools from 50 servers as raw JSON schemas into context window, compared to just 123 tokens when using a CLI.

- This represents a massive 300-page book worth of token usage vs. a sticky note, before the agent even answers a single question.

- When an agent connects to an MCP server, the full tool catalog is sent, including descriptions, input schemas, required parameters, etc. This schema information is only needed twice per session.

- Most of the time, the 71K-token schema just occupies space in the context window while the actual code, conversation and diffs fight for limited tokens.

- Benchmark studies have found that MCP can cost 4-32 times more tokens than CLI for the same work. On a 128K context window, tool definitions alone consume ~56% of the space.

- This context overhead makes it impossible to fit the agent's codebase, conversation history and reasoning chains on a 64K window. The agent will degrade, forget earlier instructions, or truncate context sooner.

- The wasted tokens represent real money spent on redundancy like punctuation and curly braces. It adds up to significant costs, especially on cheaper models.

- The problem of schema redundancy was acknowledged by the official MCP proposal and confirmed by multiple independent benchmarks.

- To fix it, the solution is to only pass a names-only listing of tools to the model, and fetch the full schemas only when needed. A CLI called mcptoon implements this approach by compressing the schemas out of the context window.

- This dial approach keeps the tool listing to just 123 tokens for 255 tools, instead of the massive 71K+ tokens currently sent. The model can then retrieve schema details selectively.

- Independent experts from Anthropic, Firecrawl, Scalekit, Scale, Microsoft, and ProMCP all agree that eager schema injection is inefficient and the current approach doesn't scale. The fix is to pay for the names, not the full schemas.