Mitate: AI-built tools with human-checked permissions

Mitate is aimed at a specific gap: AI can now generate working tools from a sentence of request, but the people using those tools often cannot read code to verify what the AI decided. The same request can produce different designs on different days — in the author's environment, the label for a reminder notice changed from "schedule response reminder" to "reminder to non-respondents." The project's answer is to separate roles: the design AI writes no code and only builds inspectable definitions, while a verified execution platform enforces permissions and records operations. Changes are classified mechanically, and those touching permissions are shown on a confirmation card.

In a demo scenario with a fictional club, an organizer asks for a schedule poll and reminders, and later asks that a vice-organizer also be able to send reminders. The platform flags that the change also grants read access to the schedule record, and both lines appear on the card before anything is applied. When a general member tries to send a reminder, the system rejects it and the rejection is logged. After the organizer lets the treasurer read the operation log, the guarantee list's line stating that nobody can read the record disappears.

The project reports 716 tests passing, 10 kinds of meaning units, and 70 design decision records as of October 10, 2026. It also states its limits: the execution platform currently runs as a single instance, filtering and aggregation of records remain basic, and the system refuses definitions it does not support rather than letting the AI implement them on the fly.