locator-mcp

MCP server that scans live browser pages and builds structured locator registries for test automation. Connects to Chrome via CDP or Playwright WebSocket, discovers elements, generates ranked XPath/CSS/testId locators (including relational XPaths for duplicate data-testid values), validates uniqueness on the page, and saves per-page JSON registries that AI clients can read incrementally.

Built for use alongside Playwright MCP — navigate and interact with the browser using Playwright MCP, then call scan_page to capture locators into a registry.

Table of contents

• Features

• How it works

• Prerequisites

• Quick start

• Run locally

• MCP configuration

• MCP tools reference

• Scan modes

• Registry format

• Token-efficient workflow

• Project structure

• Development

• Testing

• Git conventions

• Troubleshooting

Related MCP server: paparazzi

Features

• Live page scanning — connects to an already-open browser tab (no headless relaunch)

• Per-page registries — each scan saves to registry/{scanName}.json for targeted reads

• Relational XPath — ancestor-scoped and label-sibling XPaths disambiguate duplicate generic testIds (box, flex, etc.)

• Multi-strategy locators — testId, id, aria, form attributes, text, class, positional fallbacks

• Uniqueness validation — each locator is checked on the live page (matchCount, confidence)

• Dynamic testId templates — detects unstable suffixes and emits contains() + template XPaths

• Three scan modes — interactive (default), testId, full to control noise vs coverage

• Token-efficient read tools — get_registry_keys and get_locator vs full get_registry

• Portable MCP config — committed .cursor/mcp.json with ${workspaceFolder} support

How it works

flowchart LR
  subgraph browser [Browser]
    page[Open tab via CDP/WS]
  end
  subgraph scan [scan_page pipeline]
    discover[discover-elements]
    extract[extract-element-context]
    generate[generate-xpath-variants]
    rank[rank-xpath-variants]
    validate[validate-locator]
    save[save-registry]
  end
  subgraph registry [registry/]
    json["{scanName}.json"]
  end
  subgraph read [Read tools]
    keys[get_registry_keys]
    one[get_locator]
    search[search_registry]
  end
  page --> discover --> extract --> generate --> rank --> validate --> save --> json
  json --> keys
  json --> one
  json --> search

1. Discover — query the DOM using mode-specific selectors (interactive, testId, or full)

2. Extract context — per element: attributes, direct text, ancestor chain (up to 6 levels), label siblings

3. Generate variants — XPath candidates across 9 tiers (testId → relational → text → class → positional)

4. Rank & validate — score variants, verify matchCount on the live page, pick recommended + fallbacks

5. Save — write registry/{scanName}.json with keys, metadata, and locator bundles

6. Read — AI client fetches keys first, then individual entries as needed

Prerequisites

• Node.js 18+ (tested on Node 24)

• npm

• Google Chrome (for CDP mode) or a Playwright browser server (for WebSocket mode)

• Cursor or Claude Desktop for MCP integration

Quick start

git clone https://github.com/Eswarr11/locator-mcp.git
cd locator-mcp          # folder name may differ on your machine
npm install

1. Open the cloned folder as your Cursor workspace

2. .cursor/mcp.json is preconfigured — reload MCP: Cmd+Shift+J → MCP

3. Launch Chrome with remote debugging (see With Playwright MCP)

4. Navigate to your target page using Playwright MCP

5. Call scan_page with a scanName and cdpEndpoint

Run locally

# MCP server (stdio transport — used by Cursor / Claude)
npm run mcp

# Production
npm run build && npm start

MCP configuration

Run npm install in the repo before connecting.

Why not cwd + relative paths? Global ~/.cursor/mcp.json ignores cwd, so relative paths like src/server.ts resolve from your home directory and fail with ERR_MODULE_NOT_FOUND. Use the committed project config or npm run mcp --prefix <abs-path>.

Cursor (project) — recommended

.cursor/mcp.json is committed. Open this repo as your workspace — no manual edits needed:

{
  "mcpServers": {
    "locator-mcp": {
      "command": "npm",
      "args": ["run", "mcp", "--prefix", "${workspaceFolder}"]
    }
  }
}

${workspaceFolder} resolves to the project root automatically. Reload MCP after clone: Cmd+Shift+J → MCP.

If locator-mcp is also defined in global ~/.cursor/mcp.json, remove one entry to avoid duplicate servers.

Cursor (global)

For use across workspaces, add to ~/.cursor/mcp.json:

"locator-mcp": {
  "command": "npm",
  "args": ["run", "mcp", "--prefix", "<path-to-repo>"]
}

Example for this machine:

"locator-mcp": {
  "command": "npm",
  "args": ["run", "mcp", "--prefix", "/Users/eswar/Desktop/locator-collector"]
}

Claude Desktop

Claude does not support ${workspaceFolder}. Edit:

~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "locator-mcp": {
      "command": "npm",
      "args": ["run", "mcp", "--prefix", "<path-to-repo>"]
    }
  }
}

Restart Claude Desktop after saving.

Production (compiled)

Run npm run build first, then point MCP at the compiled output:

"locator-mcp": {
  "command": "node",
  "args": ["<path-to-repo>/dist/server.js"]
}

With Playwright MCP (for scan_page)

scan_page does not launch a browser — it connects to one that is already open. Pair with Playwright MCP on the same CDP endpoint (typically in global ~/.cursor/mcp.json):

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest",
        "--cdp-endpoint",
        "http://localhost:9222"
      ]
    },
    "locator-mcp": {
      "command": "npm",
      "args": ["run", "mcp", "--prefix", "<path-to-repo>"]
    }
  }
}

Step 1 — Launch Chrome with remote debugging:

open -a "Google Chrome" --args --remote-debugging-port=9222 --user-data-dir=/tmp/pw-chrome

Step 2 — Navigate to your target page using Playwright MCP tools.

Step 3 — Scan with locator-mcp:

{
  "cdpEndpoint": "http://localhost:9222",
  "scanName": "goal-side-panel-locators",
  "scanMode": "interactive"
}

Optional: pass pageUrl (substring) to target a specific tab when multiple are open.

MCP tools reference

scan_page

Scan an open browser page and save locators to registry/{scanName}.json.

Example response (abbreviated):

{
  "scanId": "goal-side-panel-locators",
  "registryFile": "registry/goal-side-panel-locators.json",
  "pageUrl": "https://app.example.com/goals",
  "scannedAt": "2026-06-23T12:00:00.000Z",
  "totalElements": 42,
  "registrySaved": true,
  "stats": {
    "total": 42,
    "unique": 38,
    "duplicate": 4,
    "lowConfidence": 2,
    "interactive": 30
  },
  "warnings": [
    "5 elements share generic testId 'box' — relational XPath applied"
  ],
  "sample": [ "...first 5 elements..." ]
}

list_registries

List all available registry files in registry/.

{ "registries": ["goal-side-panel-locators", "goal-create-side-panel-v2"] }

get_registry_keys

Return a lightweight list of element keys — use this before fetching individual locators.

Returns per key: key, tagName, testId, confidence, matchCount, strategy

Much smaller than get_registry — typically 90%+ token savings on large pages.

get_locator

Fetch a single element entry by key.

Returns: key, tagName, text, attributes, locators (recommended, fallbacks, xpath, confidence, matchCount, strategy)

get_registry

Return the full registry JSON. High token cost — prefer get_registry_keys + get_locator.

search_registry

Filter registry entries by attributes.

All provided filters are combined with AND logic.

Scan modes

Recommendation: start with interactive. Switch to testId when the app has good testId coverage. Use full only when you need exhaustive discovery.

Registry format

Each scan writes registry/{scanName}.json — a map of element keys to metadata:

{
  "saveButton": {
    "key": "saveButton",
    "tagName": "button",
    "text": "Save",
    "attributes": {
      "testId": "goal-form_button_save",
      "id": null,
      "role": null,
      "ariaLabel": null,
      "placeholder": null
    },
    "locators": {
      "recommended": {
        "xpath": "//button[@data-testid='goal-form_button_save']",
        "tier": 1,
        "strategy": "testId",
        "matchCount": 1,
        "confidenceScore": 90
      },
      "fallbacks": [ "..." ],
      "xpath": "//button[@data-testid='goal-form_button_save']",
      "confidence": "high",
      "matchCount": 1,
      "strategy": "testId"
    }
  }
}

Locator confidence:

Registry JSON files are gitignored — generated locally via scan_page. Only registry/.gitkeep is committed to preserve the folder.

Token-efficient workflow

For AI clients reading registries, follow this order to minimize token usage:

1. list_registries          → see what's available
2. get_registry_keys        → lightweight key list (~500–4k tokens)
3. get_locator (per key)    → single entry (~100–200 tokens each)
   OR search_registry       → filtered subset

Avoid: get_registry         → full file (10k–80k+ tokens)

Estimates use characters / 4 — good for relative comparison, not exact billing.

Project structure

locator-mcp/
├── .cursor/
│   ├── mcp.json              # Cursor MCP config (committed, portable)
│   └── rules/                # Cursor AI rules for this repo
├── registry/
│   └── .gitkeep              # Scan output dir (JSON files gitignored)
├── src/
│   ├── server.ts             # MCP tool definitions
│   ├── index.ts              # Package entry
│   ├── scanner/
│   │   ├── scanner.service.ts        # Scan orchestration
│   │   ├── discover-elements.ts      # DOM element discovery by scan mode
│   │   ├── extract-element-context.ts # Attributes, ancestors, labels
│   │   ├── generate-xpath-variants.ts
│   │   ├── generate-relational-xpath.ts
│   │   ├── rank-xpath-variants.ts
│   │   ├── validate-locator.ts
│   │   ├── generate-key.ts
│   │   ├── detect-dynamic-value.ts
│   │   ├── generate-locators.ts
│   │   ├── save-registry.ts
│   │   └── scanner.types.ts
│   └── shared/
│       ├── constants.ts      # Selectors, scan modes, deny lists
│       ├── registry.ts       # Registry file I/O
│       └── utils.ts
├── tests/
│   └── scanner.spec.ts       # Unit tests
├── package.json
├── tsconfig.json
└── README.md

Development

# Install dependencies
npm install

# Run MCP server in terminal (stdio)
npm run dev

# Run tests
npm test

# Compile TypeScript
npm run build

Key conventions:

• ESM imports with .js extension: import { x } from './foo.js'

• XPath locators prefer data-testid: //tag[@data-testid="..."]

• Relational XPaths disambiguate duplicate generic testIds

• Registry path: registry/{scanName}.json

See .cursor/rules/ for full AI coding standards.

Testing

npm test

Uses Node built-in test runner (node:test) with tsx for TypeScript. Tests cover:

• XPath variant generation and scoring

• Generic testId disambiguation and relational XPath

• Dynamic value detection and templates

• Key generation with ancestor context

• Constants and denylist behavior

Git conventions

Committed:

• Source (src/), tests, config, .cursor/mcp.json, .cursor/rules/

• registry/.gitkeep (empty folder placeholder)

Not committed (.gitignore):

Troubleshooting

Cannot find module '/Users/you/src/server.ts'

Global ~/.cursor/mcp.json ignored cwd. Fix: use project .cursor/mcp.json or npm run mcp --prefix <abs-path>.

Failed to connect to browser

• Chrome must be running with --remote-debugging-port=9222

• Playwright MCP must use the same --cdp-endpoint

• Relaunch Chrome:

open -a "Google Chrome" --args --remote-debugging-port=9222 --user-data-dir=/tmp/pw-chrome

Connected to browser but no open page was found

Open at least one tab in the debug Chrome instance before calling scan_page.

registry '{scanName}.json' not found

Run scan_page first, or call list_registries to see available files.

Duplicate / low-confidence locators

• Use scanMode: "interactive" to reduce noise

• Check warnings in the scan_page response

• Prefer entries with confidence: "high" and matchCount: 1

• For duplicate generic testIds, relational XPaths are generated automatically

MCP server shows twice in Cursor

Remove locator-mcp from either global ~/.cursor/mcp.json or project .cursor/mcp.json — keep only one.