What is MCP Token Cost?

For the model to know what it can call, the client serialises each server's tools/list output into the prompt. A single verbose tool — long description, large JSON schema — can cost hundreds of tokens; a server exposing dozens of tools can cost thousands; a typical multi-server setup can consume a five-figure token count before the user types anything. That overhead recurs on every request in the session.

The costs compound in three ways. Money: those tokens are billed as input on every call. Latency: larger prompts take longer to process. Quality: tool definitions compete with actual work for context window space, and models pick tools less reliably as the catalogue grows — the tool sprawl problem measured in tokens.

Mitigation options, roughly in order of effort:

• Connect fewer servers — disconnect servers a project does not use; the cheapest token is one never injected.
• Tool filtering — expose only the subset of a server's tools you actually call, where the client or an intermediary supports it.
• Virtual servers — aggregate many upstreams behind one curated facade exposing a minimal tool set (see MCP virtual server).
• Measurement — count the serialised definition size per server so trimming decisions are data-driven rather than guesswork.

PolicyLayer publishes per-server token-cost data at policylayer.com/token-cost, measured from each scanned server's actual tool definitions, so you can see what a server costs your context window before connecting it. The gateway's virtual server support lets teams expose a filtered tool subset from registered upstreams, cutting injected definitions to the tools policy actually permits.