Cartograph is a proposed Model Context Protocol proxy that helps AI agents find tools without loading an entire catalog into their context. In a test deployment spanning 22 servers and 374 tools, it exposed three proxy functions and revealed specific tool descriptions only as an agent needed them.
The system first ranks servers and then tools. Its capability cards are generated and cryptographically signed under the deploying operator’s control, reducing reliance on promotional descriptions supplied by tool publishers. A companion analysis identifies clusters of tools with confusingly similar descriptions and records which description provenance influenced each retrieval decision.
On an author-built set of 49 queries, Cartograph placed the relevant result in its top five 81.6% of the time, versus 59.2% for a Jaccard keyword baseline. The measured top-five exchange consumed 475 tokens, compared with 42,450 under the paper’s full-catalog accounting, while the proxy added about 5 milliseconds of mean latency in ten trials. These results are from a small, constructed environment rather than a broad production benchmark. Still, progressive disclosure offers a practical way to keep expanding tool catalogs from crowding out the task itself in an agent’s limited context.