English · 中文
You added one MCP server, then another, then ten. Now the assistant forgets what you said five minutes ago. The cause is not the model — it is tool discovery: the full JSON schema of every tool, shipped to the model before you type a single word.
The same 255 MCP tools. Full schemas on the left; a compact name index on the right.
Measured with tiktoken cl100k_base. That is −99.2% of the discovery cost, with
every tool still installed.
MCP is stateless by design. A client cannot know which tool you will need, so it sends the model the complete input schema of every connected tool, on every request. A single well-documented tool is often 300–600 tokens of JSON. Multiply by a hundred tools across a handful of servers and the tool list alone can swallow tens of thousands of tokens before the conversation even starts.
That cost is paid every turn. It crowds out your actual context window — the code, the docs, the earlier messages — and it grows silently every time you add a server.
Get your own number in 30 seconds with the free MCP token calculator — it runs entirely in your browser.
| Approach | Effect | Trade-off |
|---|---|---|
| Delete tools you "don't use" | small, temporary | you lose capability; you re-add them next week |
| Disable servers per project | manual, fragile | you must remember to re-enable; breaks flows |
| Index tools by name, load schema on demand | ~90–99% cut | one extra lookup step, handled by the tool |
mcptoon keeps every tool installed but sends the model only a compact name index. When the model actually needs a tool, it asks for that one tool's real schema and calls it. You keep all your tools; the model stops paying for the ones it is not using.
pip install mcptoon mcptoon manifest # the compact name index the model sees mcptoon inspect <server> <tool> # full schema, only when needed mcptoon call <server> <tool> '{...}' mcptoon usage # how much it actually saved
Zero dependencies, standard-library Python, works across 103+ desktop and CLI agents, and it also manages your Agent skills through the same index. Nothing leaves your machine.
Get mcptoon on GitHub Calculate your token cost
Is there a hard limit on how many MCP tools I can have?
No fixed number — the limit is your context window. Each tool's schema is billed every turn, so the practical ceiling is "however many schemas fit without crowding out your real work."
Do I have to delete tools to reduce MCP token usage?
No. That is the point of an index: you keep every tool and only pay for the schema of the one you actually call. Deleting tools is the temporary fix that forces you to re-add them later.
Does this work with Claude Code, Cursor, Codex, opencode?
Yes — mcptoon targets 103+ desktop and CLI agents through a zero-config install path. See the README for the current list.