Cyclr Benchmark Finds MCP Server Design Choice Cuts Token Costs by 75%

Cyclr study shows MCP server design can swing AI operating cost by more than 4x

London, LONDON, July 06, 2026 (GLOBE NEWSWIRE) -- Cyclr today released benchmark findings showing that task-scoped “Thin MCP” server design can cut token consumption by roughly 75% compared with a broad-surface “Thick MCP” approach, without reducing clean first-answer accuracy.

A “Thin MCP” limits the tools available to an LLM to only those necessary to perform the task, a “Thick MCP” is where no discretion is applied and the LLM is exposed via MCP to all the tools available within an API.

The report, MCP Server Design and Token Efficiency, evaluated 30 controlled configurations across HubSpot, Oracle NetSuite and QuickBooks using Claude Haiku 4.5 and GPT-5-mini. Cyclr found that Thin MCP designs achieved the same clean first-answer accuracy as Thick MCP while consuming far fewer tokens.

The study also found that raw Direct API access was not the cheapest route. Despite having no tool-definition overhead, Direct API consumed 58% more tokens per task than Thin MCP and delivered the lowest clean first-answer accuracy.

“These findings show that MCP server design is not a secondary consideration — it is a major driver of cost, speed and reliability,” said Fraser Davidson, CEO of Cyclr. “As SaaS companies move toward distributing MCPs to their clients, AI Interoperability and more headless product experiences, the way they expose actions and data will directly affect the economics and usability of those experiences.”

Cyclr said the key lesson is to expose the fewest tools that cover real user tasks, keep the typed MCP layer in place and design endpoints around the questions users actually ask.

“One of the clearest findings in the benchmark is that structure reduces waste,” said Nic Butler, CPTO of Cyclr. “When a model gets a narrow, typed, well-scoped interface, it performs more efficiently and more reliably. 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 that output tokens accounted for only a small share of overall usage, with most cost driven by context loaded before useful work began, including tool definitions, schemas and response payloads. Therefore MCP design is critical in overall token efficiency as MCP provides a lot of the relevant context in the initial load.

Cyclr said the findings are especially relevant for software teams building production-grade MCP experiences, where cost control, context efficiency and reliability are becoming core requirements.

Learn more about Embedded MCP and AI Interoperability.

About Cyclr

Cyclr is an embedded iPaaS and MCP platform helping SaaS and services companies deliver integrations, automation and AI-ready connectivity natively within their products and customer environments. Cyclr provides the secure, multi-tenant infrastructure, connectors, orchestration tools, embedded marketplaces and governance capabilities needed to scale integration ecosystems, expose headless services and launch MCP-enabled experiences faster to enable the agent economy. Cyclr is a UK company with offices in London and Canada.

Press Inquiries

Daniel Twigg
daniel [at] cyclr.com
https://cyclr.com
CO.5C, Corinthian House, 17 Lansdowne Road, Croydon, CR0 2BX

A video accompanying this announcement is available here: https://youtube.com/watch?v=A64QrW7qdig

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.