Paper finds five patterns in 15 MCP servers
On one machine MCP adds 0.39 ms or less, and this is not the slow part of an agent.
Researchers made a paper on the Model Context Protocol (MCP). Anthropic made MCP in November 2024 to connect AI models to external tools and data. The group looked at 15 servers: five from the ANSYR voice AI platform and ten from the open registry. They name five patterns that occur again and again, and four anti-patterns, which are patterns to prevent.
The five patterns
- Resource Gateway
- Tool Orchestrator
- Stateful Session Server
- Proxy Aggregator
- Domain-Specific Adapter
The numbers
Two AI raters put 54 more servers into these patterns. Their Cohen kappa score was 0.76. This score shows how much two raters agree. On one machine the median delay was 0.01 ms with stdio and 0.39 ms with streamable-http. For hosts that are apart, the authors model about 30 ms of network delay plus the protocol cost.
What is not known
The sample is small. The authors removed the tool-count test of version 1, because its numbers could not be found in production records.
Insight: The protocol is not where cost sits
The protocol adds less than 1 ms on one machine, and the network adds about 30 ms. Other work cuts cost in the model. In one test, llm-d found 97% of cached work to use again. Cohere says that Rerank 5 sends less text to a model. This reads as a sign that speed and cost are set by the network and the model, not by MCP. Measure those two first.
Sources
Posted
