New Fix Makes AI Assistants Cheaper and Faster at Using Your Apps

AI assistants like ChatGPT and Claude can plug into outside tools — your calendar, a database, a payment system — using a shared standard called MCP, think of it as a universal plug socket for AI apps. The problem: as companies connect hundreds of tools, the assistant has to read every tool's instruction manual before it answers anything. That is slow, costs money (you pay by the word), and makes the AI more likely to grab the wrong tool.

Cartograph, from a team of researchers in Ghana, fixes that with three tricks. First, it shows only a handful of options instead of the whole catalog — 3 entries instead of 374 in their test. Second, it hunts for near-duplicate tools that confuse the AI (it found 49 confusing clusters, 4 flagged as high-risk). Third, each tool description is cryptographically signed by whoever runs it, so a random publisher cannot slip in a misleading description. In their tests, a typical lookup dropped from about 42,450 words of text to 475, roughly 89 times less. Accuracy at finding the right tool rose from 0.59 to 0.82.

For you, that means AI helpers that respond faster, cost less to run, and make fewer wrong moves — booking the wrong thing, querying the wrong system, or leaking data because a fake tool description tricked the AI. The added delay was just 5 milliseconds, about a tenth of a blink, so you would never notice it.

The catch: this is a research paper, not a product you can buy. It was tested on a modest setup of 22 servers, not a giant corporate catalog. When descriptions were written by an AI instead of a human operator, accuracy sometimes dropped — so someone still has to keep those descriptions clean. Expect this kind of plumbing to appear first inside big companies' AI platforms, then filter down to everyday apps.