Dominaitrix

Inspiration

The web was built for people clicking buttons, not agents trying to complete tasks. When agents rely on screenshots, coordinates, and brittle selectors, even simple workflows become slow and unreliable.

WebMCP offers a better interface—but most existing websites do not support it yet. Dominaitrix was inspired by a practical question: what if we could teach the web we already have to speak WebMCP?

What it does

Dominaitrix is a sideloadable Chrome extension that installs curated WebMCP adapters on supported websites. Each adapter translates a site’s live interface into structured tools that agents can discover and invoke directly.

Our adapters currently support:

• Managing a live ESPN Fantasy Football draft
• Exploring Domino’s menus and building a cart
• Reading Hacker News stories and discussions
• Searching and navigating Wikipedia articles

The difference is dramatic. In two ESPN drafts using the same prompt, ordinary browser control successfully drafted only 3 of 16 intended players and followed 1 of 5 roster instructions. With Dominaitrix, the agent drafted 16 of 16 players and followed 5 of 5 instructions.

Dominaitrix does not provide the agent. It gives WebMCP-capable agents a faster, safer, and more reliable way to use existing websites.

How we built it

Dominaitrix consists of three main pieces:

1. A Chrome extension that installs adapters through chrome.userScripts
2. A shared runtime that registers tools through document.modelContext.registerTool()
3. A static adapter registry containing versioned metadata, compatibility information, tool descriptions, and SHA-256 source digests

The extension periodically checks the registry, verifies each adapter’s integrity, and installs or updates it on matching pages. Adapters run in an isolated user-script world by default and expose narrowly scoped tools with structured inputs, outputs, and errors.

Every adapter is reviewed and tested in the repository before publication. The project includes an adapter template, automated registry generation, integrity validation, and a browser-like test harness for exercising tools without depending entirely on live sites.

We also built privacy-conscious local diagnostics. Dominaitrix records minimized health events—such as the adapter version, tool name, outcome, and broad error category—but never records tool arguments, page content, cookies, tokens, or response bodies. Nothing is uploaded automatically.

Challenges we ran into

The hardest problem was translating dynamic, human-oriented interfaces into stable agent tools.

ESPN’s draft room uses virtualized tables, changing position filters, a live clock, and controls whose meaning changes when the user is on the clock. A delayed instruction could accidentally draft a player during the wrong pick. We addressed this by using stable player IDs, requiring the expected pick number for consequential actions, detecting stale state, and verifying the result after every selection.

We faced similar issues across sites: hidden application state, asynchronous UI updates, undocumented APIs, route changes, and actions that can return success before the visible interface has actually updated.

Distribution created another challenge. Downloadable adapters make rapid compatibility fixes possible, but they also create a trust boundary. We designed a curated registry with no public upload path, exact source digests, restrictive match patterns, explicit execution worlds, and automated consistency checks.

Finally, WebMCP and Chrome’s user-script support are still experimental. We had to handle missing APIs, disabled browser settings, unsupported pages, and partial initialization without leaving users with mysterious failures.

Accomplishments that we're proud of

We are most proud that Dominaitrix demonstrates a measurable improvement rather than merely registering demo tools.

The ESPN experiment showed that structured tools can turn a fragile, time-sensitive workflow into one an agent completes accurately and with time to spare. The agent could inspect the draft, manage autopick, search available players, maintain a queue, and make validated selections while following a multi-part strategy.

We are also proud that the adapter system is reusable. The same runtime supports experiences as different as drafting a fantasy team, configuring a pizza, exploring a discussion tree, and navigating an encyclopedia.

What we learned

We learned that agents do not need unrestricted access to a website’s internals. They need a small set of well-designed capabilities with stable identifiers, precise descriptions, bounded results, and actionable errors.

We also learned that successful execution must be verified. A click or HTTP response is not proof that the intended state changed—especially during live, asynchronous workflows.

Most importantly, WebMCP is valuable even before websites adopt it natively. An adapter layer can provide an immediate bridge while showing site owners what an agent-native interface for their product could look like.

What's next for Dominaitrix

Next, we want to expand the curated adapter library and make adapters easier to develop, review, test, and maintain.

We plan to add stronger compatibility monitoring, richer local diagnostics, safer confirmation patterns for consequential actions, and tooling that detects when a site change breaks an adapter.

Longer term, we hope Dominaitrix can serve two purposes: make today’s web useful to agents now, and provide working prototypes that help websites design native WebMCP support of their own.

Built With

• chrome
• html
• javascript
• node.js
• pnpm
• s3
• webmcp