(1 Dec 2025) AI search tools like Elicit, Consensus, and Scite.ai have spent years racing to build centralised indexes of academic content— first by indexing the open content that is available and then trying to get publisher partnerships. But unlike their predecessors of the 2010s – the Web Scale Discovery systems (now called “Discovery layers” or “Discovery Services”) like Summon, Ebsco Discovery Service, Primo, publishers are much more wary of handing over content for indexing amid intense concerns about AI training on copyrighted material. The Model Context Protocol(MCP) offers an alternative path. MCP allows AI models to query publisher content directly through standardised servers, keeping full text on publisher infrastructure while enabling real-time retrieval.
Interestingly, this isn’t just an architectural workaround; it changes what’s possible. Using MCP, Claude can now pilot-test Boolean searches against PubMed, check MeSH headings, assess recall, and iteratively refine strategies—capabilities that were impossible when LLMs generated search strings blindly. For evidence synthesis and library discovery, this may be the most consequential development since federated search gave way to discovery layers.
The Model Context Protocol may fundamentally change how AI tools access academic content. Rather than building massive centralised indexes (the approach taken by Elicit, Consensus, and others), MCP allows AI models to connect directly to publisher content through standardised “servers.”
Aaron Tay test two MCP servers: the Wiley AI Gateway and a PubMed MCP server .
