Customex

Visit - https://customex.am1.tech

Inspiration

As a serial web browser (as in a person that browses the web), I wondered whether with the power of AI, especially when baked into the browser itself, could be used to make websites appear and function as per a user's specific needs. However, we don't want users to inject arbitrary code, and web developers should have the final say over the possible ways to modify a website. Thus, I built "Customex", a protocol that bakes this customisability into a contract that is safe and easy to interface. This hackathon gave me the perfect excuse to finally start building it.

What it does

Customex is an open protocol that lets a website publish a machine-readable visual contract (manifest). AI can then translate natural-language requests into reviewable, reversible view changes.

• Sites declare meaningful components, capabilities, allowed values, protected areas, and named views.
• Users can say “make this calmer” or “show only "design"- and "operations"-related posts.” The model proposes a patch, which the user has the authority to either accept or reject.
• The extension validates every action against the live manifest. Apply is explicit, changes are atomic, and Undo/clicking on a previous state restores the previous state. ## How I built it I used Codex with GPT-5.6 to build the core protocol, SDK, Chrome extension, demo pages, and documentation. Sites publish a constrained manifest through the SDK. The extension reads it, uses Chrome’s Prompt API to turn a request into a proposed patch, validates it again, and applies it only after approval. ## Challenges I ran into Chrome extension content scripts run in an isolated world, so they could not safely read the page’s JavaScript manifest directly. I fixed this by publishing the manifest as JSON in a shared DOM script element. I also had to prevent the model from treating broad style requests as permission to hide arbitrary feed items (sometimes the AI would misunderstand, but that is not a protocol-level flaw per se). ## What I learned That AI is useful here as a mapper between user intent and allowed changes (it turns “make this calmer” into values exposed by the site). I also learned that a clear metadata and constraints make AI behaviour far more predictable than merely relying on the built-in model itself. ## What's next for Customex So much is possible with such a protocol! At the risk of over-exaggeration, it can reshape the way we browse the web. Customex can add publisher-declared grouping, ordering, collapsing, accessibility modes, responsive views, richer read-only data projections, and custom design tokens. It can also support direct named views with no AI, Chrome's on-device model for local requests, and an opt-in cloud model for larger semantic tasks such as interpreting a long design brief or finding posts relevant to a design-system migration. Every client would still be bound by the publisher manifest. I also plan a manifest-authoring assistant for developers, and a community library of safe view recipes.

Built With

• codex
• gpt
• javascript