A key challenge in building the MCP server was the number of Policy Checks available. With over 50 predefined checks, agents needed a way to easily select and configure the ones relevant to the user's query. Loading the descriptions of all the checks and their parameters into the context of the LLM would have been inefficient and costly. Therefore, implementing each check as a separate MCP tool was not a viable option.

To solve this, Futurescale proposed a progressive discovery mechanism. The server instructions make the agent aware that a workflow of tools exists, and how to use them. A list of available checks can be requested, and when the agent selects a check to configure, it can request the detailed description and parameters for that specific check. This approach minimizes the amount of information loaded into the LLM context at any given time, improving performance and reducing costs. From there, the agent can proceed to configure the selected check as needed. It can also run the configured checks against database schemas to identify any policy violations.

Written in TypeScript, the Policy Checks server has been packaged as a Claude extension, but is also configurable for use with Cursor, VSCode, Jetbrains IDEs, Goose, and practically any other AI agent.