The Model Context Protocol (MCP) lets AI agents discover and invoke tools, yet loading every definition scales as $O(n)$ when catalogs grow. Cartograph introduces a federated MCP proxy that reduces agent‑visible discovery to $O(k)$ progressive disclosure.\ \ Cartograph integrates three mechanisms:\
- Operator‑attested capability cards – Ed25519‑signed tool descriptions generated under the deploying operator’s control, replacing ranked publisher copies;\
- Rift – a three‑layer confusable‑cluster analysis comprising density clustering, query‑margin analysis, and token diagnosis;\
- Two‑stage retrieval – ranking servers before ranking tools.\ \ In a deployment of 22 servers and 374 tools, Cartograph exposed only three proxy tools instead of all 374 definitions. On a 49‑query author‑crafted benchmark, Cartograph achieved R@5 = 0.816 versus 0.592 for a Jaccard keyword baseline. A measured top‑5 discovery exchange consumed 475 tokens, dramatically lower than the 42,450 tokens required by full‑catalog accounting. Rift identified 49 confusable clusters, including four high‑risk clusters originating from bootstrap‑generated cards. An exploratory comparison of 119 LLM‑generated descriptions eliminated the zero‑distance cluster but showed that mixing card‑generation regimes can reduce R@5. Ten gateway trials added an average latency of 5 ms (≈0.8%) relative to direct stdio MCP calls.\ \ Cartograph complements code‑execution approaches: it controls which tool descriptions are surfaced and records the provenance of the descriptions used for ranking each query, enabling traceable retrieval.\ \ Review