LinkedIn 销售导航器 MCP 服务器

MCP LinkedIn Sales Navigator

MCP server for LinkedIn Sales Navigator — Browser automation via Playwright.

LinkedIn Sales Navigator doesn't provide a public API. This MCP server uses browser automation (Playwright) to control Sales Navigator through an authenticated browser session. It enables AI assistants to search leads, view profiles, manage lists, send InMails, and export data — all through the standard Model Context Protocol.

⚠️ Important Disclaimers

• This tool requires an active LinkedIn Sales Navigator subscription
• Browser automation: This operates by controlling a real browser — no API hacking
• Session required: You must be logged into LinkedIn in the browser instance
• Rate limiting: Use responsibly. LinkedIn may restrict accounts that perform excessive automation
• Terms of Service: Review LinkedIn's ToS before using automation tools
• No credentials stored: This tool never stores or handles LinkedIn passwords

Features

Quick Start

Prerequisites

• Node.js ≥ 20
• LinkedIn Sales Navigator subscription with active session
• A browser with an active LinkedIn login (Chrome recommended)

Install

npm install globodai-mcp-linkedin-sales-navigator

Or clone and build from source:

git clone https://github.com/globodai-group/mcp-linkedin-sales-navigator.git
cd mcp-linkedin-sales-navigator
npm install
npm run build

Configuration

The server is configured via environment variables (read in src/index.ts):

Rate limits and account safety

This server throttles your own Sales Navigator activity so a Claude session does not fire actions back-to-back. It is not anti-detection: there is no fingerprint spoofing, user-agent rotation, or proxy support.

0 on a daily budget means unlimited. Raise a cap by setting the matching LSN_DAILY_* variable; check usageToday on linkedin_session_status for used / cap / remaining. Counters reset at the next local midnight and live in LSN_USAGE_FILE (or ~/.mcp-linkedin-sales-navigator/usage.json). That file stores only dates, category names, and counts.

LinkedIn does not publish official automation limits and restricts automated activity in its User Agreement. No tool can guarantee an account will not be restricted. Start with the defaults, spread activity over working hours, and stay within your Sales Navigator InMail credits.

Authentication Methods

1. CDP (Chrome DevTools Protocol) — Recommended

Connect to an already-running Chrome browser with an active LinkedIn session.

Start Chrome with remote debugging:

# macOS
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222

# Linux
google-chrome --remote-debugging-port=9222

# Windows
& "C:\Program Files\Google\Chrome\Application\chrome.exe" --remote-debugging-port=9222

# If Chrome is installed elsewhere, quote the full path. Node.js 20+ is required.

Use PowerShell or Command Prompt. Keep the debugging port free (default 9222) and match LSN_CDP_ENDPOINT.

Then log into LinkedIn Sales Navigator manually. The MCP server connects to this browser.

Shutdown behavior (CDP): when the MCP server exits, it detaches from your Chrome instance only. It does not close your tabs or quit the browser.

LSN_AUTH_METHOD=cdp LSN_CDP_ENDPOINT=http://localhost:9222 npx globodai-mcp-linkedin-sales-navigator

2. User Data Directory

Use an existing Chrome profile that's already logged into LinkedIn:

LSN_AUTH_METHOD=session LSN_USER_DATA_DIR=/path/to/chrome/profile npx globodai-mcp-linkedin-sales-navigator

3. Cookie-Based

Export your LinkedIn cookies and provide them as a JSON file:

LSN_AUTH_METHOD=cookies LSN_COOKIES_PATH=/path/to/linkedin-cookies.json npx globodai-mcp-linkedin-sales-navigator

MCP client setup

Use the same command, args, and env block in any MCP client. Replace paths and endpoints for your machine.

Claude Desktop

Edit the config file:

Example entry:

{
  "mcpServers": {
    "linkedin-sales-navigator": {
      "command": "npx",
      "args": ["-y", "globodai-mcp-linkedin-sales-navigator"],
      "env": {
        "LSN_AUTH_METHOD": "cdp",
        "LSN_CDP_ENDPOINT": "http://localhost:9222"
      }
    }
  }
}

Restart Claude Desktop after saving.

Claude Code

From a shell (adjust env vars as needed):

claude mcp add linkedin-sales-navigator -- \
  env LSN_AUTH_METHOD=cdp LSN_CDP_ENDPOINT=http://localhost:9222 \
  npx -y globodai-mcp-linkedin-sales-navigator

Generic MCP client

Point the client at the package binary via npx or node /path/to/mcp-linkedin-sales-navigator/dist/index.js, stdio transport, with the environment variables from the table above.

Clawdbot / CORTX Integration

This MCP server is designed to work seamlessly with Clawdbot browser relay:

{
  "mcpServers": {
    "linkedin-sales-navigator": {
      "command": "npx",
      "args": ["globodai-mcp-linkedin-sales-navigator"],
      "env": {
        "LSN_AUTH_METHOD": "cdp",
        "LSN_CDP_ENDPOINT": "http://localhost:9222"
      }
    }
  }
}

When using Clawdbot's browser relay, the AI assistant can directly control a browser tab where you're logged into Sales Navigator — no additional setup needed.

Tool Examples

Search for leads

{
  "tool": "linkedin_search_leads",
  "arguments": {
    "title": "VP of Engineering",
    "location": "San Francisco Bay Area",
    "companySize": "201-500",
    "industry": "Computer Software"
  }
}

Get a lead's profile

{
  "tool": "linkedin_get_lead_profile",
  "arguments": {
    "profileUrl": "https://www.linkedin.com/sales/lead/ACwAAAxxxxxx"
  }
}

Send an InMail (with dry run)

{
  "tool": "linkedin_send_inmail",
  "arguments": {
    "profileUrl": "https://www.linkedin.com/sales/lead/ACwAAAxxxxxx",
    "subject": "Quick question about your team",
    "body": "Hi, I noticed your team is growing...",
    "dryRun": true
  }
}

Export leads to CSV

{
  "tool": "linkedin_export_leads",
  "arguments": {
    "source": "current_search",
    "format": "csv",
    "limit": 50
  }
}

Troubleshooting

Start with the linkedin_session_status tool. It connects lazily (same as other tools), reports authMethod, browserConnected, authenticated, usageToday, and a short hint for the next step.

See CONTRIBUTING.md for reporting selector breakage without sharing secrets.

Architecture

┌─────────────────────────────────────────────────┐
│  AI Assistant (Claude, GPT, etc.)               │
│  ↕ MCP Protocol (stdio)                         │
├─────────────────────────────────────────────────┤
│  MCP Server (this package)                      │
│  ├── Tools (search, leads, inmails, lists, etc.)│
│  ├── Browser Controller (Playwright)            │
│  └── Selectors & Auth                           │
├─────────────────────────────────────────────────┤
│  Playwright → Chromium / Chrome                 │
│  ↕ CDP or Direct                                │
├─────────────────────────────────────────────────┤
│  LinkedIn Sales Navigator (web app)             │
└─────────────────────────────────────────────────┘

Development

# Install dependencies
npm install

# Install Playwright browsers
npx playwright install chromium

# Development mode (watch)
npm run dev

# Build
npm run build

# Lint
npm run lint

# Type check
npm run typecheck

# Unit tests
npm test

Selector Maintenance

LinkedIn rotates hashed CSS module classes on every deploy. Selectors in this project follow a fixed priority (see PR #3 and src/browser/selectors.ts):

1. data-anonymize="..." and similar LinkedIn field markers (person-name, headline, title, company-name, …)
2. data-x--..., data-sn-view-name, data-control-name — product hooks tied to behavior
3. Semantic HTML / ARIA and shared artdeco-* design-system classes
4. Playwright text matchers (:has-text(), :text-matches()) only when nothing else is stable

Field locators are prioritised arrays, resolved in order by src/browser/query.ts (not comma-separated CSS lists, which follow document order).

The lead profile topcard headline and location often lack stable hooks. src/browser/dom-extract.ts walks structure from stable anchors (name h1, Save button) and reads text by position, with selector fallbacks afterward.

When tools break after a LinkedIn UI change:

1. Reproduce on the affected page and tool
2. Inspect DOM hooks (data-anonymize, data-control-name) before class names
3. Update selectors, query arrays, or dom-extract heuristics as needed
4. Open a PR with a redacted HTML snippet (see CONTRIBUTING.md)

License

MIT — see LICENSE

Disclaimer

This project is not affiliated with, endorsed by, or sponsored by LinkedIn Corporation. LinkedIn and Sales Navigator are trademarks of LinkedIn Corporation. Use this tool responsibly and in accordance with LinkedIn's Terms of Service.