Universal LinkedIn MCP Server

A production-grade Model Context Protocol (MCP) server that connects your AI assistant (Claude, Grok, Cursor, Antigravity) to LinkedIn with human-like stealth browser automation, mathematical account boundary guardrails, and zero official API restrictions.

💡 Why This Exists: The Problem & The Solution

❌ The Problem

1. The Closed API Wall: LinkedIn's official Developer APIs are gated behind enterprise partner programs (LinkedIn Marketing Developer Platform / LinkedIn Talent Solutions). Solo builders, freelancers, and AI agent developers cannot obtain write access to publish posts, send connection invites, or update profiles programmatically.

2. The Anti-Bot Ban Trap: Standard automation tools (Puppeteer, Selenium, raw Chromium binaries) trigger LinkedIn's bot detection checkpoints within minutes. They broadcast navigator.webdriver = true, lack persistent cookie management, type at robotic speeds, and fail when presented with 2FA or CAPTCHAs, resulting in instant account restrictions.

3. The LLM Safety & Prompt-Injection Hazard: Giving an autonomous LLM browser access is dangerous. Without rigid guardrails, prompt injections can trick an AI into editing unintended accounts, scraping unauthorized targets, or blasting spam across your professional network.

4. Desktop vs. Cloud AI Fragmentation: Desktop AI tools (Claude Desktop, Cursor) communicate over local stdio, whereas web-based cloud AI (like Grok.com) run on remote servers and cannot access your local browser or session without secure tunneling and CORS support.

✅ The Solution

1. Full 29-Tool Platform Without API Keys: Provides your AI with complete human capabilities—profile editing, rich posting (with images/PDFs), interactive polls, inbox history, connection management, analytics, and executive intelligence briefings.

2. Native Google Chrome Stealth Engine: Uses your computer's real, native Google Chrome installation rather than generic Chromium binaries. Strips automation markers (navigator.webdriver), introduces human typing delay jitter, and executes natural Bezier-curve mouse movements.

3. Passwordless, 2FA-Friendly Session Management: Never asks for or stores your plaintext LinkedIn password. You log in interactively once through your real browser, solve any 2FA challenge, and the encrypted session state is saved locally to ~/.linkedin_mcp. Built-in keep-alive telemetry extends LinkedIn's 30-day sliding activity window automatically.

4. Mathematical Account Boundaries: Profile mutation tools (update_my_headline, update_my_about, add_experience, etc.) do not accept a target profile parameter. They are hardcoded to /in/me. It is mathematically impossible for an AI to modify an external profile.

5. Universal Multi-Transport Architecture: Runs locally via stdio using zero-clone uvx, or remotely via streamable-http / sse with full CORS support and a 1-click Cloudflare Tunnel script for Grok.com.

Related MCP server: kaushik-linkedin-mcp

📋 Prerequisites

Before setting up, make sure your computer has:

1. Google Chrome: Installed and functioning normally.

2. Python 3.10 or higher: python.org/downloads

3. uv (Fast Python Package Runner):

   • Windows (PowerShell):

     powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

   • macOS / Linux:

     curl -LsSf https://astral.sh/uv/install.sh | sh

⚡ Step 1: One-Time Interactive Authentication

You only need to log in once. The server saves your session state to ~/.linkedin_mcp so your AI assistants remain authenticated across restarts.

Open your terminal and run:

uvx --from git+https://github.com/ChimbuezeDavid/linkedin-mcp python -c "import asyncio; from linkedin_mcp.tools.auth import linkedin_start_login; asyncio.run(linkedin_start_login())"

What happens:

1. A real Google Chrome browser window will open automatically.

2. Enter your LinkedIn credentials and complete 2FA / verification if prompted.

3. Once your LinkedIn home feed loads, the tool automatically verifies your identity, saves your encrypted cookies, and closes the browser.

4. Your terminal will display: Active session verified for <Your Name>.

🔌 Step 2: Connect to Your AI Assistant

Choose your setup below based on whether you are using a Desktop AI Client (Claude, Cursor, Antigravity) or a Cloud Web AI (Grok.com).

Option A: Desktop AI Clients (Zero-Clone via uvx)

You do not need to download or clone this repository. Your AI client will execute it directly using uvx.

1. Claude Desktop

Add this snippet to your claude_desktop_config.json:

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "linkedin": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/ChimbuezeDavid/linkedin-mcp",
        "linkedin-mcp"
      ]
    }
  }
}

2. Antigravity

Open your Antigravity MCP settings (mcp_config.json or Settings > MCP):

{
  "mcpServers": {
    "linkedin": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/ChimbuezeDavid/linkedin-mcp",
        "linkedin-mcp"
      ]
    }
  }
}

3. Cursor & Windsurf

Add to .cursor/mcp.json or your global Cursor settings:

{
  "mcpServers": {
    "linkedin": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/ChimbuezeDavid/linkedin-mcp",
        "linkedin-mcp"
      ]
    }
  }
}

4. Claude Code CLI

Run directly from your terminal:

claude mcp add linkedin uvx --from git+https://github.com/ChimbuezeDavid/linkedin-mcp linkedin-mcp

Option B: Cloud Web-Based AI (Grok.com)

Cloud-based AI assistants (like grok.com/connectors) cannot connect to localhost. They need a public, encrypted tunnel with CORS and Streamable HTTP support.

We include a pre-configured 1-click startup script that starts the MCP server in Streamable HTTP mode on port 8765 and launches a Cloudflare Tunnel.

Step 1: Run the Grok Connector Script

Clone the repository and run the script:

git clone https://github.com/ChimbuezeDavid/linkedin-mcp.git
cd linkedin-mcp

• On Windows (PowerShell):

  powershell -ExecutionPolicy Bypass -File .\run_for_grok.ps1

• On macOS / Linux:

  # Terminal 1: Launch MCP server in Streamable HTTP mode
  uv run linkedin-mcp --transport streamable-http --port 8765
  
  # Terminal 2: Expose via Cloudflare Tunnel
  cloudflared tunnel --url http://127.0.0.1:8765

Step 2: Configure Grok.com

1. Copy the public tunnel URL displayed in your terminal (e.g. https://example-subdomain.trycloudflare.com).

2. Go to grok.com/connectors in your browser.

3. Click Add Custom MCP Server:

   • Name: LinkedIn MCP

   • URL: https://example-subdomain.trycloudflare.com/mcp (⚠️ Important: you must append /mcp to the end of the URL)

4. Click Save and start chatting with Grok!

Option C: Local Development Setup

If you want to contribute or modify the codebase locally:

git clone https://github.com/ChimbuezeDavid/linkedin-mcp.git
cd linkedin-mcp
uv sync

To configure your AI client to point to your local code:

{
  "mcpServers": {
    "linkedin": {
      "command": "uv",
      "args": [
        "--directory",
        "<ABSOLUTE_PATH_TO_LINKEDIN_MCP>",
        "run",
        "linkedin-mcp"
      ]
    }
  }
}

🛠️ Available MCP Tools (29 Tools)

All tools operate strictly within the authenticated account boundary and enforce @require_auth.

🧪 Automated Testing & Verification

The repository includes a comprehensive unit test suite verifying boundary constraints, security invariants, parameter validation, and session health calculation:

uv run python -m unittest discover tests

Expected output:

Ran 12 tests in 0.015s

OK

🔒 Security & Privacy Invariants

• Zero Plaintext Password Exposure: The server never prompts for, reads, or stores your LinkedIn password.

• Local Storage Only: All session state, cookies, and identity caches are saved strictly to your local machine at ~/.linkedin_mcp. No external telemetry or cloud servers are used.

• Mathematical Impossibility of Impersonation: Self-profile editing tools are hard-coded to navigate to /in/me/. There is no target_profile_url parameter, preventing prompt-injected LLMs from altering other members' profiles.

📄 License

MIT License. See LICENSE for details. Authored with ❤️ by Chimbueze (David) Okoroji.