{
  "mcpServers": {
    "hci-quality": {
      "command": "npx",
      "args": [
        "--yes",
        "--package=https://github.com/priyankark/interface-observatory/releases/download/v0.1.0/hci-quality-mcp-0.1.0.tgz",
        "hci-quality-mcp"
      ]
    }
  }
}

Use this server entry in Claude Desktop or Cursor's MCP configuration. For VS Code, use the same entry under the top-level servers key and add "type": "stdio". Other clients can run the same command with stdio transport. Node.js 22.12+ is required; the SaaS requires Node 24+. Chrome for Testing is installed by Puppeteer. AUDIT_ALLOW_LOOPBACK=false enables public-only page inspection. Keep Chrome's sandbox enabled.

When npm publication is confirmed, the command can be shortened to npx --yes hci-quality-mcp. The GitHub release works independently of npm publishing credentials.

Tools

Tool	Input	Output
`inspect_interface`	`url`, optional `device` (`desktop`/`mobile`)	Rendered control sizes and label candidates, navigation groups, overflow, review findings, limitations
`model_task`	`goal`, `methods` with `name`, `selectionRule`, `operators`; optional `timings`	GOMS method comparison using explicit KLM operator sums
`model_pointing`	`distance`, `width`; optional `intercept`, `slope`	Fitts Shannon difficulty in bits; seconds only when both coefficients supplied
`model_choice`	`choices`; optional `probabilities`, `intercept`, `slope`	Hick–Hyman entropy; seconds only when both coefficients supplied
`cognitive_walkthrough`	`goal`, `persona`, `steps` with action, four answers and evidence	Analyst concerns and explicitly unassessed questions

Example task model

{
  "goal": "Open account settings",
  "methods": [{
    "name": "Menu navigation",
    "selectionRule": "User knows the menu location and uses a pointer",
    "operators": [
      {"operator": "M", "note": "Prepare to navigate"},
      {"operator": "P"},
      {"operator": "K"},
      {"operator": "R", "seconds": 0.5}
    ]
  }]
}

Operators: K keystroke/button press, P pointing, H hand movement between devices, M mental preparation, R explicit system response wait. count defaults to 1. seconds can override an individual operator; timings customizes K/P/H/M defaults. Defaults are illustrative K=0.2, P=1.1, H=0.4, M=1.35 seconds. Operator placement is an analyst responsibility. This predicts routine skilled, error-free execution, not novice learning or satisfaction.

Cognitive walkthrough

Each step supplies action, evidence, and goalVisible, actionDiscoverable, mappingClear, feedbackClear as true, false, or null. null stays unassessed. The tool organizes supplied judgments; it does not pretend to have interviewed users.

The inspector is a top-level DOM screen, not a complete accessibility-name computation. It does not enter credentials or perform clicks. Small-target observations do not evaluate WCAG spacing exceptions. Use the companion a11y MCP and manual review for accessibility.

Primary sources: Card, Moran & Newell (1980), MacKenzie (1992), Hick (1952), NN/g cognitive walkthrough, W3C target size guidance.