Prism
Evaluate the contract, not just the URL

Tell us what the WebMCP is for.Then test whether it delivers.

Generic scanners reward surface area. Prism scores the tool contract against a declared job, its real side effects, and the evidence a person can verify on the page.

Purpose-specific coverage Mutation read-back Human approval boundaries

Evaluation setup

A declared contract keeps the score honest.

Profile + evidence

Help a shopper find, compare, and safely purchase a product.

Evidence source
search_productsget_productmanage_cartbegin_checkout
Example evaluation

A score that explains what “good” means

Contract profile: Commerce & checkout

Purpose fit

88

Needs proof

The tool set covers the declared journey. At least one action still needs stronger behavioral evidence.

See scoring evidence

Journey coverage

Expected job → discovered tool → observable outcome

Discover

search_products

Understand

get_product

Change page state

manage_cart

Hand off safely

begin_checkout

Evidence, not vibes

Checks are tied to the declared contract.

Live tools were discovered on the top-level page

4 WebMCP tools in the runner snapshot

manage_cart does not prove visible state changed

Return a state revision and verify the same change through the page or a read-back tool.

Read and write tools are distinguishable

2 read · 2 write

Purpose coverage

100

Does each declared user-journey step have a suitable tool?

Contract quality

100

Are schemas narrow and side effects honestly annotated?

Observable proof

50

Do mutations change the same visible state and survive read-back?

Runtime hygiene

100

Are tools imperative, top-level, and lifecycle-owned?

Agent activity

A privacy-preserving trace of WebMCP calls made on this page. Inputs and outputs are never logged here.

Last 0 of 5
No agent calls yet. Invoke one of Prism's four WebMCP tools to see an auditable read/write trace.
Honest execution boundary

The browser observes. The dashboard judges.

WebMCP tools belong to the page that registered them. A normal hosted scanner cannot discover or execute another origin's tools through an iframe or an HTTP fetch. Prism therefore separates collection from evaluation.

1 · Declare

Choose the product type, job, expected tools, and human-only boundary.

2 · Collect

A compatible browser opens the target page, discovers tools, executes test cases, and records state evidence.

3 · Evaluate

Prism maps observed behavior to the declared journey and produces an explainable repair list.

Declared context

Choose a profile or define the exact user journey and expected capabilities.

Behavioral checks

Exercise tools and compare their results with the same live page state.

Actionable repairs

Get a contract-level fix with the evidence required to re-run the check.