Researchers have mapped the emerging architecture of the Model Context Protocol ecosystem — and found it already has recurring shapes worth naming.
The Model Context Protocol, which Anthropic shipped in November 2024, gives large language models a standardized way to connect to external tools and data sources. Within months, hundreds of community-built servers appeared on GitHub. A new industry experience paper analyzed fifteen independently developed servers — five from the ANSYR voice AI platform and ten from the official MCP registry — and identified five recurring patterns: Resource Gateway, Tool Orchestrator, Stateful Session Server, Proxy Aggregator, and Domain-Specific Adapter. The authors also documented four anti-patterns and flagged authentication, versioning, and observability as cross-cutting concerns that keep showing up regardless of which pattern a team picks.
The most concrete finding is a tool-count cliff. The study found that tool-selection accuracy for Claude Haiku 4.5 drops below 90% when a context window holds more than 10 to 15 tools; for Sonnet 4 the threshold is 20 to 30 tools. That is a practical ceiling, not a theoretical one, and it gives teams building MCP integrations a concrete reason to decompose large servers rather than piling tools into a single context. Inter-rater reliability across two independent LLM raters on 54 held-out servers came in at a Cohen's kappa of 0.76, which is solid but leaves three pattern-boundary cases the taxonomy could not cleanly resolve.
MCP is less than two years old, and formalized architectural guidance usually lags ecosystems by far longer — the Gang of Four design patterns arrived years after object-oriented code was already everywhere. Getting a structured catalog this early, with a replication package attached, is useful; whether the five patterns hold as the ecosystem matures is a question the next wave of servers will answer.