What is discomcp?

DiscoMCP helps AI agents work with your tools more effectively by teaching them how you actually use those tools, not just what they can technically do. MCP, short for Model Context Protocol, is a way for AI agents to connect to external systems and services. The problem this project addresses is that when an agent connects to a tool with dozens or even a hundred possible actions, it often does not know which ones matter for your specific work, how your data is organized, or what is safe to touch. As a result, the agent either guesses or gets stuck. DiscoMCP solves this by pointing itself at an MCP server and exploring it the way a careful, cautious user would. It learns which handful of actions you actually rely on, in what order you tend to use them, and how the results of one action lead naturally into the next. It builds this understanding by observing your real workspace and the data that is genuinely there, rather than relying on a generic list of what the tool offers. Safety is a central design principle. The exploration process runs behind what the project calls a default deny gate: it will only perform an action if it can prove that action only reads data and cannot write, change, or delete anything, even if the tool itself claims to be harmless. Any sensitive information encountered during exploration is stripped out before being saved. According to benchmarks in the README, agents equipped with a generated skill from DiscoMCP complete tasks in noticeably fewer steps than agents working cold, with the biggest improvements showing up on harder, more complex tasks. The tool is a single binary requiring no separate installation, and getting started involves running it directly with a simple configuration file pointing to the target MCP server. Under the hood, DiscoMCP is built in Rust, with a small deterministic core enforcing every safety check while a language model handles the actual thinking and exploration. Every fact recorded in a generated skill is labeled by how it was learned, whether declared outright, documented, directly observed, or inferred, so the agent using it does not mistake a guess for something confirmed. The project is dual licensed under the Apache License 2.0 and the MIT License.