Why
The claim in one line
Give your own frontend app high-accuracy, visible and controllable AI operations.
The user says one sentence to the app ("file a personal leave for tomorrow", "filter expenses by travel"), and the AI completes the corresponding multi-step operation — switching pages, filling forms, clicking, submitting — all of it visible.
The core insight
The semantics of a modern frontend app already exist; they are just scattered around and never collected:
- Which pages exist — written in the route config.
- Which fields each form has, their types, whether required — written in
<Form.Item>,register, zod/yup schemas. - Which clickable actions exist — written on
<button>/<a>with their labels.
This information is already the authoritative definition of "what the app can do". The problem is not missing semantics, but that nothing collects them into a capability manifest the AI can read.
So the core action of this framework is: use static AST analysis at build time to extract a "capability manifest" from the structure already in the code, instead of asking developers to annotate everything again for the AI.
How we differ
The one-line distinction from competitors: others let the AI "look at the screen and guess how to operate any website", we let the AI "read the structure and operate your own app precisely".
- vs browser agents (browser-use / Operator / Computer Use): they operate any website by visual guessing — general but accuracy-limited; we only serve apps whose source you control, trading a compile-time capability boundary for high accuracy.
- The key qualifier: only for apps whose source the developer controls (internal tools, SaaS, admin systems), not automation of arbitrary third-party websites.
An honest boundary
- Auto-inference accuracy < manual annotation — that's physics. We don't claim "fully automatic and most accurate".
- Embrace MCP, don't reinvent the transport protocol; no pure visual recognition.
See RFC 0001 for the full positioning and differentiation.