Description

This policy brief maps how Anthropic’s Model Context Protocol (MCP) is reshaping the AI developer ecosystem and assesses emerging risks of concentration and gatekeeping. Using a dataset of 2,874 MCP servers scraped from Pulse MCP and enriched with GitHub star counts, we categorize servers into 15 functional types and hand-review the top projects by usage. We find MCP activity is highly concentrated: the top 10 servers account for 45.7% of all stars (and the top decile for 88.3%), suggesting many agents rely on a small set of APIs and services. Three categories dominate—Computer & Web Automation, Software Engineering, and Database & Search—capturing 72.6% of observed usage. Popular servers frequently enable both read (context access) and write (agentic) operations, with agentic browsing tools especially prominent; several leading integrations are unofficial, raising auditing and trust concerns. We argue protocols like MCP are “rules of the road” that can decentralize markets but, without safeguards, may simply relocate bottlenecks. We recommend: expanding structured API access, prioritizing local processing and multi-provider gateways, and instituting transparency requirements (usage stats, access denials, rate limits, dependency disclosures) for widely used servers. These measures would keep the MCP network contestable and resilient while enabling safe, equitable innovation.

Citation

Rosenblat, Sruly, Strauss, Ilan, O’Reilly, Tim, and Isobel Moure. “MCP in Practice.” Policy Brief.Social Science Research Council, September 2025. https://www.ssrc.org/publications/mcp-in-practice-mapping-power-concentration-and-usage-in-the-emerging-ai-developer-ecosystem/