Humcp turns AI tool descriptions into accessible forms

Samuel Proulx wanted the information behind a website without fighting its interface. In an account published on 21 October, he describes building humcp: an application that turns tools intended for AI assistants into forms he can operate with a screen reader.[1]

The connection is Model Context Protocol, or MCP. Its tool definitions describe a function, the inputs it expects and, optionally, the structure of its results. A client can discover those definitions instead of having a separate interface written for each service.[3]

Although MCP tools are designed for model-controlled use, the specification leaves the interaction pattern open. A human can choose the operation and supply the inputs directly. The protocol supplies the description of the task; it does not require a chatbot to stand between the person and the service.[3]

An accessible form can still ask the wrong question

Proulx’s weather example captures the remaining friction. The forecast tool wanted coordinates, so he first had to use geocoding. He then had to select weather variables before it returned the forecast he wanted. His baseball example required a player lookup before a statistics request could use the resulting identifier.[1]

He reports using humcp daily and finding some lookups quicker than websites or chatbots. This is its creator’s experience, not a comparative user study. The account describes version 1.1 and identifies saved queries and easier linking between tools as useful next steps.[1]

The repository inspected on 29 October describes a Windows 10/11 client with labelled fields, validation, keyboard access and structured result displays. Tables become grids with headers; nested results become trees. It supports local servers and web connections, with authenticated web sign-in still unavailable in the documented version.[2]

The project says it is tested with the NVDA screen reader, including checks of the operating system’s accessibility tree and listening through release workflows. AI News inspected the documentation and source repository; it did not run those accessibility tests.[2]

Humcp offers an inspectable experiment in reusing AI infrastructure for direct human control. Its current platform and connection limits make it a specific tool to evaluate, with a concrete design problem for others to improve.[2]

Report an error or suggest a correction