Cartograph
An arXiv paper proposing a federated agent tool discovery scheme for the Model Context Protocol (MCP). It employs progressive disclosure, capability card signing, and obfuscation cluster analysis to substantially reduce token overhead and improve retrieval accuracy in AI agent tool discovery.
Timeline (last 90 days)
Cartograph is a federated proxy solution for the Model Context Protocol (MCP), designed to restructure the tool discovery mechanism of AI agents.
In a deployment environment containing 22 servers and 374 tools, Cartograph exposes only 3 proxy tools to agents instead of the full 374 definitions.
A single top-5 discovery interaction measured in the paper uses only 475 tokens, whereas the same operation under a full-catalog accounting approach requires approximately 42,450 tokens.
Cartograph reduces the tool discovery process from O(n) full-catalog traversal to O(k) progressive disclosure.
Cartograph's implementation is built on three mechanisms: operator-endorsed capability cards, Rift confusability cluster analysis, and two-stage retrieval.
Cartograph's capability cards use Ed25519 signatures and are generated by deployment operators rather than directly adopting official descriptions from tool publishers.
Rift is a three-layer confusability cluster analysis comprising density clustering, query margin analysis, and token diagnostics.
Cartograph employs two-stage retrieval: first ranking servers, then ranking specific tools.
On an evaluation benchmark of 49 queries constructed by the authors, Cartograph achieves an R@5 of 0.816, while the Jaccard keyword-matching baseline achieves only 0.592.
Compared to direct stdio MCP calls, the Cartograph gateway adds an average of only 5 ms (approximately 0.8%) of latency across ten trials.
3 more timeline events
All Facts (13)
Cartograph is a federated proxy solution for the Model Context Protocol (MCP), designed to restructure the tool discovery mechanism of AI agents.
50%UnverifiedIn a deployment environment containing 22 servers and 374 tools, Cartograph exposes only 3 proxy tools to agents instead of the full 374 definitions.
50%UnverifiedA single top-5 discovery interaction measured in the paper uses only 475 tokens, whereas the same operation under a full-catalog accounting approach requires approximately 42,450 tokens.
50%UnverifiedCartograph reduces the tool discovery process from O(n) full-catalog traversal to O(k) progressive disclosure.
50%UnverifiedCartograph's implementation is built on three mechanisms: operator-endorsed capability cards, Rift confusability cluster analysis, and two-stage retrieval.
50%UnverifiedCartograph's capability cards use Ed25519 signatures and are generated by deployment operators rather than directly adopting official descriptions from tool publishers.
50%UnverifiedRift is a three-layer confusability cluster analysis comprising density clustering, query margin analysis, and token diagnostics.
50%UnverifiedCartograph employs two-stage retrieval: first ranking servers, then ranking specific tools.
50%UnverifiedOn an evaluation benchmark of 49 queries constructed by the authors, Cartograph achieves an R@5 of 0.816, while the Jaccard keyword-matching baseline achieves only 0.592.
50%UnverifiedCompared to direct stdio MCP calls, the Cartograph gateway adds an average of only 5 ms (approximately 0.8%) of latency across ten trials.
50%UnverifiedExperiments found that using 119 LLM-generated descriptions can eliminate the observed zero-distance clusters, but mixing different card generation mechanisms may actually reduce R@5.
50%UnverifiedThe paper emphasizes that Cartograph is complementary to, rather than a replacement for, code-execution-based approaches, with Cartograph focusing on the discovery and trusted provenance layer.
50%UnverifiedCartograph's benchmark is still author-constructed and limited in scale, and its real-world effectiveness still requires broader validation.
50%