Intel / article

Cartograph: Federated Tool Discovery with Operator-Attested Retrieval for AI Agents

Source: arxiv.org
Author: Justice Owusu Agyemang, Michael Agyare, Kwame Opuni-Boachie Obour Agyekum, Kwame Agyeman-Prempeh Agyekum, Francisca Adoma Acheampong, Jerry John Kponyo
Date: Published Sep 28, 2026

Context

Cartograph helps an find relevant tools without first loading every tool's instructions. It places a discovery service in front of connected tool servers, ranks servers and then tools, and returns a shortlist. The descriptions used for ranking are signed under the deploying operator's control, recording their origin.

In the authors' 374-tool deployment, the discovery interface exposed three proxy operations. Retrieving five tools used 475 , compared with 38,250 for loading all tool definitions. The paper's larger 42,450-token baseline also includes an estimated reasoning allowance. On 49 author-written queries, the intended tool appeared among the first five results 81.6% of the time. The result demonstrates a smaller discovery exchange with a retrieval tradeoff, rather than guaranteed tool selection.

Key takeaways · AI-distilled

Cartograph is a federated proxy that swaps full tool-catalog loading for progressive disclosure: on a 22-server, 374-tool deployment the agent sees three proxy tools instead of 374 definitions.
A top-5 discovery exchange used 475 tokens versus 42,450 under full-catalog accounting, and on a 49-query author-built recall at 5 was 0.816 against 0.592 for a keyword baseline.
Tool descriptions are operator-attested capability cards signed with Ed25519 under the deploying operator's control, so ranking does not rely on publisher-written copy.
Its Rift analysis flagged 49 clusters of easily confused tools, four rated high risk, and the gateway added about 5ms mean latency (0.8%) over direct stdio MCP calls.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source arxiv.org