FED-GRAPH-MCP: Scaling LLM Tool Discovery via Sparsified Functional Graphs and Incremental Synchronization

Assumes 1–2 courses of background. Domain terms may appear without definition.

Large language models (LLMs) often need to use external tools—like APIs or database queries—to complete tasks. These tools are listed in a growing number of servers using the Model Context Protocol (MCP), which can now exceed 10,000 servers. When an LLM needs a specific tool, it must quickly find the right one among thousands. Current methods treat each tool as an isolated text document and search by matching keywords, but they miss important connections between tools and struggle when the list changes frequently. To solve this, researchers developed FED-GRAPH-MCP, a new system that builds a map (or “graph”) showing how tools are related, such as which tools often work together. It then combines this relational map with standard text-based search to rank tools more accurately. The system also updates its map gradually whenever a tool is added, removed, or changed, rather than rebuilding everything from scratch. In tests, this approach dramatically reduced memory usage by 95.6% and update times by up to 99.9% compared to older methods. It also slightly reduced the number of words the LLM needed to process per task. Overall, FED-GRAPH-MCP makes it feasible for LLMs to quickly and efficiently discover tools in a massive, ever-changing ecosystem.

AI-generated (deepseek-v4-flash) · created 2026-06-09