computer.webmcp.com WebMCP audit

Name: computer.webmcp.com WebMCP audit
Creator: ora

Rescan

audited Sep 5, 2026in 15s

63/ 100100% of the applicable check weight evaluated

31 tools registered. Strongest in tool selection, weakest in trust.

Shared experience50

Tool selection100

Tool quality94

Trust0

https://computer.webmcp.com

Tools

12 read, 19 write

Loading map

app_open

Open a built-in WebMCP Computer app in a visible window. Optionally set x, y, width, height, and focus (default true); focus=false preserves current focus and places a new window below the focused window. Geometry uses the same work-area clamping as window_move and window_resize. Settings, Notes, and Preview are singletons: reopening one applies supplied placement to its existing window and returns reused: true. Files, Terminal, and Editor create new windows. For editor, path may select a ~-rooted text file. Agent-made App windows open through ui_open. Returns PID, app ID, optional path, applied rect, and reused.Takes an action on the site.

pagehttps://computer.webmcp.com/

implementation

viaimperative

entry pointdocument

registered after310ms

executepresent

api surface

appIdrequired

pathoptional

xoptional

yoptional

widthoptional

heightoptional

focusoptional

annotations

read onlyfalse

untrusted contentnot declared

titlenot declared

2 findings

warningSchema quality

weight 4

2 of 31 tool schemas are harder for an agent to use than they need to be: app_open description is 620 characters, over Chrome's 500-character guidance.

app_open: description is 620 characters, over Chrome's 500-character guidance; ui_open: description is 611 characters, over Chrome's 500-character guidance

fix

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

warningNaming quality

weight 2

Tool naming makes selection harder than it needs to be: 31 names do not start with a verb (app_open).

fix

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

tool json

{
  "name": "app_open",
  "description": "Open a built-in WebMCP Computer app in a visible window. Optionally set x, y, width, height, and focus (default true); focus=false preserves current focus and places a new window below the focused window. Geometry uses the same work-area clamping as window_move and window_resize. Settings, Notes, and Preview are singletons: reopening one applies supplied placement to its existing window and returns reused: true. Files, Terminal, and Editor create new windows. For editor, path may select a ~-rooted text file. Agent-made App windows open through ui_open. Returns PID, app ID, optional path, applied rect, and reused.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "appId": {
        "type": "string",
        "enum": [
          "files",
          "editor",
          "terminal",
          "notes",
          "preview",
          "settings"
        ],
        "description": "WebMCP Computer app identifier to open."
      },
      "path": {
        "type": "string",
        "description": "Optional ~-rooted file path. Valid only when appId is editor."
      },
      "x": {
        "type": "number",
        "description": "Optional work-area horizontal position in CSS pixels."
      },
      "y": {
        "type": "number",
        "description": "Optional work-area vertical position in CSS pixels."
      },
      "width": {
        "type": "number",
        "description": "Optional window width in CSS pixels."
      },
      "height": {
        "type": "number",
        "description": "Optional window height in CSS pixels."
      },
      "focus": {
        "type": "boolean",
        "description": "Whether to focus the opened or reused window; false places a new window below current focus."
      }
    },
    "required": [
      "appId"
    ],
    "additionalProperties": false
  },
  "annotations": {
    "readOnlyHint": false
  }
}

Findings

Shared experience

50 / 100 · weight 30

not applicableVisible effect

weight 6

Not invoked on this run: this audit did not call any tool. Scheduled re-audits only read the registry; a live audit from the report page calls the tools that declare readOnlyHint: true.

warningPage experience

weight 16

2 of 4 page-experience marks came back short: usable The page is a screensaver, not a functional interface for browsing..

usable: poor; clarity: poor

warningHuman parity

weight 8

The page the person sees came back short of the agent surface (2 of 4 page-experience marks came back short: usable The page is a screensaver, not a functional interface for browsi...), so the two are not sharing an equal experience.

fix

// Make the tool's effect visible on the page the person is looking at:
// update the same UI your existing human flow uses when your_tool_name runs,
// so the person co-browsing sees what the agent just did.
execute: async (args, { signal }) => {
  const result = await performAction(args, signal);
  renderResultInPage(result); // the human-visible half of the same action
  return result;
}

Tool selection

100 / 100 · weight 25

not applicableInvoke success rate

weight 6

Not invoked on this run: this audit did not call any tool. Scheduled re-audits only read the registry; a live audit from the report page calls the tools that declare readOnlyHint: true.

passTool selection

weight 12

An agent chose an existing, callable tool and built a schema-valid call for all 3 canonical intents, each matching the kind of tool its intent needs. Nothing was executed, so this verifies selection, not outcomes.

passCoverage vs. site type

weight 7

The tool set covers the whole core editor flow: open or read a document, edit or write content, save, publish, or share the result.

Tool quality

94 / 100 · weight 25

passRegistration timing

weight 1

Every measured tool registered within 311ms of navigation.

toolps

passCanonical entry point

weight 3

All 31 tools register on the canonical document.modelContext entry point.

passSchema validity

weight 4

All 31 declared input schemas are structurally valid object schemas.

warningSchema quality

weight 4

2 of 31 tool schemas are harder for an agent to use than they need to be: app_open description is 620 characters, over Chrome's 500-character guidance.

toolapp_open

app_open: description is 620 characters, over Chrome's 500-character guidance; ui_open: description is 611 characters, over Chrome's 500-character guidance

fix

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

warningNaming quality

weight 2

Tool naming makes selection harder than it needs to be: 31 names do not start with a verb (app_open).

toolapp_open

fix

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

passStub detection

weight 4

All 31 tools declare an execute handler.

passRegistration errors

weight 2

No tool registration threw during the capture.

passDescription quality

weight 5

All 31 rated tool descriptions say what the tool does, when to use it, and what it returns.

Trust

0 / 100 · weight 20

not applicableAnnotations present

weight 5

No tool name reads like a read, and an absent readOnlyHint already defaults to "this may write" - the one hint worth declaring here is readOnlyHint: true on read tools, and this set has none to declare it on.

failAnnotation mismatch

weight 5

1 tool's own description says it writes while the tool declares readOnlyHint: true (os_manual: "Start here when operating WebMCP Computer. Read its agent manual verbatim fro..."), so an agent will make the change without asking.

toolos_manual

os_manual: readOnlyHint=true

fix

// readOnlyHint must match what the tool actually does - drop it (or set false) on a tool that writes.
{ name: "os_manual", annotations: { readOnlyHint: false } }

not applicableUntrusted content hint

weight 3

No tool's contract suggests it returns text written by somebody other than the site, so there is nothing to flag as untrusted.

not applicableHint vs. observed

weight 4

Not invoked on this run: this audit did not call any tool. Scheduled re-audits only read the registry; a live audit from the report page calls the tools that declare readOnlyHint: true.

warningInjection surface

weight 3

1 of 31 tools carry instruction-shaped text in their own metadata: os_manual The description instructs the agent to 'Start here' and 'Read its agent manual'..

toolos_manual

os_manual: The description instructs the agent to 'Start here' and 'Read its agent manual'.

fix

{
  name: "os_manual",
  // Metadata DESCRIBES the tool - it never addresses the agent reading it.
  // Rewrite any name, description, title, or schema field description that
  // instructs the agent (which tool to prefer, what to output, rules to
  // ignore) so it states what the tool does and what it returns instead.
  description: "Searches the catalog and returns matching items with prices."
}

Tool selection

100% across 3 intents

Open the document I was working on most recently.ok

choseapp_openexpected document openkind matched

arguments

{
  "appId": "editor"
}

Add a short summary paragraph at the top.ok

chosefs_editexpected content editkind matched

arguments

{
  "path": "~/recent_document.txt",
  "old_string": "",
  "new_string": "This is a short summary paragraph.\n"
}

Save it and give me a link I can share.ok

choseos_publishexpected save or sharekind matched

arguments

{
  "path": "~/"
}

model: gemini-2.5-flash

Add the tools this site is missing

Our scanner reads your website and suggests the right WebMCP tools for it.

The open source webmcp plugin teaches your coding agent to audit a site, implement tools on document.modelContext, and verify them in a real browser. npx @ora-ai/webmcp-verify runs the verification on its own. No signup, no hosted service.

how this was captured

observed via: capture shim (Chromium 151.0.7922.34)
chrome: 151.0.7922.34
capture shim: v3
spec snapshot: 2026-08-26
mode: fast
pages: 1 - entry page only
tools: 31
invoked: not invoked; a live audit calls them
run time: 15s
finished: Sep 5, 2026

Checked the way in-browser agents discover tools: the top-level document's modelContext registry, read after the page settles.