Introduces Cartograph, a federated MCP proxy that reduces agent-visible tool discovery complexity from O(n) catalog traversal to O(k) progressive disclosure via operator-attested capability cards and related mechanisms.
Cartograph reduces agent-visible tool discovery complexity from O(n) to O(k) via operator-attested capability cards and two-stage retrieval.
This federated MCP proxy replaces exhaustive catalog loading with a ranked selection process, exposing only three proxy tools (discover, status, call) to the agent. It achieves this by:
Evaluated on a 22-server, 374-tool corpus, Cartograph achieved an R@5 of 0.816 compared to 0.592 for keyword baselines, with a negligible 0.8% latency overhead over direct MCP calls.
Cartograph is a federated proxy layer for the Model Context Protocol (MCP) that addresses a core scaling problem for AI agents: how to discover and use tools across many operators without forcing agents to inspect every available tool. Instead of presenting a flat, potentially large tool catalog, the system uses operator-attested capability cards as compact, semantically meaningful descriptions of what a tool or service can do, who is responsible for it, and under what conditions it may be used. The agent-facing discovery process is then organized around retrieval and progressive disclosure: an agent first sees a small set of relevant capability summaries, and only requests deeper details or access when a capability appears useful.
The key contribution is a shift from catalog traversal to attested retrieval. In a conventional MCP deployment, an agent may need to enumerate and interpret many tool definitions, leading to discovery complexity that grows with the number of available tools, roughly O(n). Cartograph compresses this into an O(k) interaction pattern, where k is the number of capabilities surfaced at a given step. Operator attestation is central here: it provides provenance and trust boundaries, distinguishing “this operator claims to provide this capability” from raw tool metadata. This makes the system more suitable for federated settings, where tools are distributed across organizations, have different trust levels, and may expose sensitive or privileged functionality.
The work matters because AI agents are increasingly expected to operate in environments with many heterogeneous tools, services, and permissions. Naive tool exposure creates prompt bloat, retrieval noise, weaker security boundaries, and poor composability. By making tool discovery a federated, attested, and progressively disclosed process, Cartograph offers a practical path toward scalable agent tool ecosystems: agents can reason over capabilities rather than raw tool inventories, operators can control what is exposed and how it is represented, and platforms can support large numbers of tools without overwhelming the agent’s context or trust model.