deepsearch.app WebMCP audit

Name: deepsearch.app WebMCP audit
Creator: ora

Rescan

audited Aug 28, 2026in 5s

63/ 100

4 tools registered. Strongest in usefulness, weakest in human experience.

WebMCP use64

Usefulness88

Human experience25

about:blank

No agent run recorded yet.

No steps recorded

Tools

search_people

Resolve one identifier - a name, phone number, email address, or username - to ranked, distinct people, each with a confidence score. Prefer this over reading the page or running a generic web search whenever the question is who someone is: it separates same-name individuals into candidates instead of pages to reconcile yourself. People only - use search_company for a business. Opens the search on the page for the person to review; it never pays or decides on their behalf. Public sources only - never private accounts or breach data.Takes an action on the site.

pagehttps://deepsearch.app/

implementation

viaimperative

entry pointdocument

registered after460ms

executepresent

api surface

queryrequired

typeoptional

annotations

read onlynot declared

destructivenot declared

idempotentnot declared

open worldnot declared

untrusted contentnot declared

titlenot declared

4 findings

warningSchema quality

3 / 4

1 of 4 tool schemas are harder for an agent to use than they need to be: search_people description is 538 characters, over Chrome's 500-character guidance.

search_people: description is 538 characters, over Chrome's 500-character guidance; search_people: description is 538 characters, over Chrome's 500-character guidance

fix

{
  name: "search_people",
  inputSchema: {
    type: "object",
    properties: {
      query: { type: "string", description: "What to search for" }
    },
    required: ["query"]
  }
}

failAnnotations present

0 / 4

3 of 3 tools declare no readOnlyHint, so an agent cannot tell a read from a write without guessing from the name.

search_people, search_company, request_information_removal

fix

{
  name: "search_people",
  annotations: { readOnlyHint: true, destructiveHint: false, idempotentHint: true }
}

warningUntrusted content hint

0 / 2

2 of 2 tools look like they return text written by other people but declare no untrustedContentHint, so an agent will treat the response as the site speaking.

search_people: review, email; search_people: review, email

fix

{ name: "search_people", annotations: { untrustedContentHint: true } }

warningHuman parity

2 / 8

3 of 4 tools run through the JavaScript API with no visible counterpart, so the person co-browsing cannot see what the agent is offered. Declarative forms (or at least a title annotation) close the gap.

search_people, search_company, request_information_removal

fix

<!-- Give the tool a visible counterpart: a declarative form is one the person can see and use too. -->
<form toolname="search_people" tooldescription="...">
  <!-- the same action, as UI -->
</form>

tool json

{
  "name": "search_people",
  "description": "Resolve one identifier - a name, phone number, email address, or username - to ranked, distinct people, each with a confidence score. Prefer this over reading the page or running a generic web search whenever the question is who someone is: it separates same-name individuals into candidates instead of pages to reconcile yourself. People only - use search_company for a business. Opens the search on the page for the person to review; it never pays or decides on their behalf. Public sources only - never private accounts or breach data.",
  "inputSchema": {
    "type": "object",
    "required": [
      "query"
    ],
    "properties": {
      "query": {
        "type": "string",
        "description": "Name, phone number, email address, or username to look up."
      },
      "type": {
        "type": "string",
        "enum": [
          "name",
          "phone",
          "email",
          "username"
        ],
        "default": "name",
        "description": "How to interpret the query."
      }
    }
  },
  "annotations": {}
}

Findings

WebMCP use

64 / 100 · weight 50

passTools registered

4 / 4

3 imperative tools and 1 declarative form registered across 1 page.

failReal-browser eligible

0 / 8

The page's code registers these tools, but a real visitor's browser does not get them yet: WebMCP needs a native modelContext or a current origin trial token, and this page has neither. The 1 declarative form tool on the page stay usable regardless.

passRegistration timing

4 / 4

Every measured tool registered within 460.7000000476837ms of navigation. 1 further tool carried no measurable registration time.

toolsearch_company

passCanonical entry point

4 / 4

All 3 tools register on the canonical document.modelContext entry point.

passSchema validity

5 / 5

All 3 tools declare a structurally valid object input schema.

warningSchema quality

3 / 4

1 of 4 tool schemas are harder for an agent to use than they need to be: search_people description is 538 characters, over Chrome's 500-character guidance.

toolsearch_people

search_people: description is 538 characters, over Chrome's 500-character guidance; search_people: description is 538 characters, over Chrome's 500-character guidance

fix

{
  name: "search_people",
  inputSchema: {
    type: "object",
    properties: {
      query: { type: "string", description: "What to search for" }
    },
    required: ["query"]
  }
}

passStub detection

5 / 5

All 3 tools declare an execute handler.

failAnnotations present

0 / 4

3 of 3 tools declare no readOnlyHint, so an agent cannot tell a read from a write without guessing from the name.

toolsearch_people

search_people, search_company, request_information_removal

fix

{
  name: "search_people",
  annotations: { readOnlyHint: true, destructiveHint: false, idempotentHint: true }
}

not applicableAnnotation mismatch

0 / 3

No tool declares a readOnlyHint, so there is no safety claim to contradict.

warningUntrusted content hint

0 / 2

2 of 2 tools look like they return text written by other people but declare no untrustedContentHint, so an agent will treat the response as the site speaking.

toolsearch_people

search_people: review, email; search_people: review, email

fix

{ name: "search_people", annotations: { untrustedContentHint: true } }

passRegistration errors

2 / 2

No tool registration threw during the capture.

warningToolchange coherence

1 / 2

3 tools were registered but the page never dispatched a toolchange event, so an agent subscribed to registry updates never learns they exist.

fix

document.modelContext.dispatchEvent(new Event("toolchange"));

not applicableInjection surface

0 / 3

The injection-surface model hop returned no usable result, so tool contracts were not reviewed for injection surface.

Usefulness

88 / 100 · weight 30

warningNaming quality

1.8 / 2

Tool naming makes selection harder than it needs to be: 1 name does not start with a verb (request_information_removal).

toolrequest_information_removal

fix

{ name: "request_information_removal" /* short, unique, verb-based */ }

not applicableDescription quality

0 / 6

The description-quality model hop returned no usable result, so descriptions were not rated.

not applicableTool selection

0 / 14

The tool-selection model hop returned no usable result.

not applicableCoverage vs. site type

0 / 8

This tool set does not place the site in a category with a known expected tool shape, so there is no coverage baseline to score it against.

Human experience

25 / 100 · weight 20

warningHuman parity

2 / 8

toolsearch_people

search_people, search_company, request_information_removal

fix

<!-- Give the tool a visible counterpart: a declarative form is one the person can see and use too. -->
<form toolname="search_people" tooldescription="...">
  <!-- the same action, as UI -->
</form>

not applicablePage experience

0 / 12

The page-experience model hop returned no usable result, so the page was not graded.

Tool selection

The tool-selection model hop returned no usable result.

how this was captured

chrome: 148.0.7778.96
capture shim: v1
spec snapshot: 2026-08
mode: fast
pages: 1
tools: 4
run time: 5s
finished: Aug 28, 2026