6 MCP Servers, One Agent: What the Token Bill Really Showed

An engineer wired six MCP servers — filesystem, Postgres, GitHub, CRM, browser, and search index — into a single production agent and logged 41,200 tool calls over a week, tracking tokens and latency per server. The finding: tool schemas alone consumed 41% of the available context window, and prompt cache hit rates collapsed from 92% to 48% as more servers were added, before a user typed anything.

Streamable HTTP proved slower and less reliable than stdio for locally hosted servers, adding roughly 87% to average latency. Rewriting vendor-shipped tool descriptions turned out to be the cheapest fix available, lifting first-try tool-selection accuracy from 64% to 86%. Meanwhile, the new OAuth 2.1 requirement introduced undocumented pain: token refresh storms, friction between Dynamic Client Registration and enterprise SSO, and the absence of tool-level scoping in the auth model itself.

Four fixes — capping tools loaded per session, reverting local servers to stdio, auditing every description, and adding an internal authorization layer on top of OAuth — cut per-session cost by 53%. For teams scaling MCP past a single tool, the takeaway is that schema overhead, transport choice, and authorization granularity are cheap to get right early and expensive to fix later.