Thin MCP design cut AI token use by about 75% in Cyclr test
Cyclr benchmarked MCP setups for HubSpot, NetSuite and QuickBooks, finding narrower tool exposure reduced token consumption without hurting first-answer accuracy.
By Hana Yoshida · Markets Reporter
3 min read
Cyclr has released benchmark findings that put a hard cost number on a design choice facing software teams building AI integrations. A task-scoped “Thin MCP” setup used about 75% fewer tokens than a broad “Thick MCP” approach in controlled tests, according to Cyclr, while matching clean first-answer accuracy.
The research covers Model Context Protocol server design, a growing concern for business software vendors exposing application data and actions to large language models. Token use affects AI operating costs directly, so the way tools and schemas are presented to a model can change the economics of production AI features.
Thin MCP limits the model to the tools needed for a specific task. Thick MCP exposes the model, through MCP, to all tools available within an API without narrowing the surface area first.
The report, MCP Server Design and Token Efficiency, evaluated 30 controlled configurations across HubSpot, Oracle NetSuite and QuickBooks. The tests used Claude Haiku 4.5 and GPT-5-mini.
Cyclr said the narrower design delivered the same clean first-answer accuracy as the broader approach while using far fewer tokens. The result points to server design as a cost-control lever for teams building AI features on top of SaaS systems.
The benchmark also compared Thin MCP with raw Direct API access. Direct API access had no tool-definition overhead, but consumed 58% more tokens per task than Thin MCP and produced the lowest clean first-answer accuracy in the study.
That finding runs against a common assumption in AI integration work: removing middleware does not automatically reduce model cost. In these tests, the typed MCP layer appeared to help by giving the model a more structured interface for the job at hand.
“These findings show that MCP server design is not a secondary consideration — it is a major driver of cost, speed and reliability,” Fraser Davidson, Cyclr’s CEO, said. He said SaaS companies distributing MCPs to customers will have to consider how exposed actions and data affect the usability and economics of those experiences.
The main design lesson from the Thin MCP server design benchmark is to expose the fewest tools that still cover real user tasks, keep the typed MCP layer, and shape endpoints around the questions users ask.
Nic Butler, Cyclr’s CPTO, said structure reduced waste in the benchmark. “When a model gets a narrow, typed, well-scoped interface, it performs more efficiently and more reliably,” Butler said. “When you expose too many tools or force trial-and-error parameter discovery, token usage rises and task completion becomes less dependable.”
The report found output tokens made up only a small share of total usage. Most token cost came from context loaded before useful work began, including tool definitions, schemas and response payloads.
For software teams putting MCP-based AI features into production, that shifts attention from prompt length alone to interface design. The benchmark frames MCP design as part of the operating-cost stack for AI products connected to business applications.