Novas WebMCP Proving Ground

• 

  Nova

Inspiration

Most WebMCP demos prove that a tool call works. We wanted to answer the more useful question: what does WebMCP actually buy an agent?

What it does

WebMCP Proving Ground is a live, inspectable challenge with three tasks. Every task can be solved through the visible interface or through the page's nine WebMCP tools. Both paths use the same state and mutations, while the page measures interactions, rejected attempts, and completion.

The tasks expose failures that real browser agents hit:

• An appointment slot that looks available until conflicts are checked
• An order where one item has already shipped
• A report where selected filters are not yet applied

The WebMCP tools return the hidden state before Codex acts. Codex can therefore choose a valid appointment, cancel only eligible items, and export the intended report without guessing. The live mission monitor records every tool call and ends at 3/3 objectives complete.

How we built it

The project is a dependency-free static website using semantic HTML, CSS, and JavaScript. It registers nine task-level tools through document.modelContext.registerTool, with navigator.modelContext and provideContext fallbacks. The interface and tool handlers share one state model, so the comparison is inspectable rather than simulated.

For the recorded run, Nova Browser gives Codex control of a real browser tab through MCP. Because Nova does not yet discover native page WebMCP tools directly, the demo openly uses a main-world compatibility bridge to the exact same definitions and handlers. The page labels this boundary instead of hiding it.

Why Nova Browser

Nova is not just a click driver. It provides more than 800 MCP tools in 25 bundles for browser control, target ownership, safety, evidence, local memory, and unattended work. Verified website behavior can become local, evidence-bound knowledge for later agents, with drift checks to guard against stale rules.

Challenges we ran into

The hardest part was making the comparison fair. UI clicks and tool calls had to mutate the same data, start from identical state, and be counted consistently. We also had to show why task-level tools matter without turning WebMCP into renamed UI controls.

What we learned

The largest benefit is not fewer clicks by itself. It is that an agent receives task-relevant state before making a consequential decision. WebMCP turns hidden interface state into an explicit contract.

What's next

We want to run the same patterns against larger authenticated workflows and connect native WebMCP discovery directly to Nova's browser MCP layer.

Built With

• ai
• css3
• html5
• javascript
• webmcp