Cartograph cuts the token bill for AI agents hunting for tools, without dumping every tool definition into the model's context.
The Model Context Protocol, or MCP, lets AI agents discover and call outside tools, but loading every connected tool's definition gets expensive as catalogs grow. A new proxy called Cartograph tackles that by combining three pieces: cryptographically signed capability cards written under the deploying operator's control rather than self-promoted copy from tool publishers; a clustering system called Rift that flags tools with confusingly similar descriptions; and a two-stage search that picks the right server first, then the right tool within it. In a test deployment spanning 22 servers and 374 tools, Cartograph exposed just three proxy tools instead of all 374 definitions. On a 49-query benchmark it retrieved the correct tool in the top five results 81.6% of the time, versus 59.2% for a basic keyword-matching baseline.
The token math is the real headline. A typical top-five discovery exchange dropped from about 42,450 tokens under full-catalog loading to 475 tokens, a cut of nearly 99%, while adding only 5 milliseconds of latency. That is the difference between an agent that can realistically browse dozens of connected tool servers and one that burns its context window before doing anything useful. Rift's clustering also caught 49 sets of confusingly similar tool descriptions, four of them flagged high-risk, a sign that today's sprawling, unvetted tool registries are as much a trust problem as a cost problem.
The catch is that Cartograph still depends on the operator writing honest capability cards, and the researchers found that mixing hand-written and AI-generated descriptions can actually make retrieval worse. This counts as progress on cost, not yet a fix for trust.