linkedin-mcp

An MCP (Model Context Protocol) server that lets your AI assistant read and edit your own LinkedIn profile — by driving a real, logged-in Chrome through the Chrome DevTools Protocol (CDP).

Your assistant never sees your password: you log in once, by hand, in a visible Chrome window. Everything after that is the same UI you'd click yourself, automated.

AI client (Hermes / Claude Desktop / Cursor / Codex / ChatGPT)
        │  MCP (stdio or streamable HTTP)
        ▼
  server.py  ──CDP──▶  Chrome (your logged-in LinkedIn session)

⚠️ Read this first

Automating LinkedIn violates LinkedIn's User Agreement. Accounts that automate get rate-limited, checkpointed, or restricted. This project is for learning and personal, low-volume use (fixing your headline, adding a project) — not scraping, bulk actions, or engagement farming. You own the risk. Keep edits rare, deliberate, and human-approved (every write tool requires confirm=True for exactly this reason).

Related MCP server: linkedin-mcp-server

Features

All writes are gated: they refuse unless called with confirm=True, so your client must show you the change first.

Setup

1. Prerequisites

• Python 3.11+

• Google Chrome on the machine that will run the browser

• An MCP client (Hermes, Claude Desktop, Cursor, Codex CLI, ChatGPT — see CHATGPT.md)

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

2. Start Chrome with a debug port (dedicated profile)

Use a separate Chrome profile so automation never touches your personal browser and your LinkedIn login persists across restarts:

macOS / Linux

google-chrome --remote-debugging-port=9222 \
  --user-data-dir="$HOME/.linkedin-mcp-profile" \
  --no-first-run https://www.linkedin.com/login
# macOS: "/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" ...

Windows (headless/SSH-launched) — processes started from an SSH session are killed when the session closes and don't appear on your desktop. Use a one-shot scheduled task instead:

schtasks /Create /TN linkedin-mcp-chrome ^
 /TR "\"C:\Program Files\Google\Chrome\Application\chrome.exe\" --remote-debugging-port=9222 --user-data-dir=C:\linkedin-mcp-profile --remote-allow-origins=* --no-first-run https://www.linkedin.com/login" ^
 /SC ONCE /ST 00:00 /F
schtasks /Run /TN linkedin-mcp-chrome

Chrome 136+ note: remote debugging is ignored on the default profile — the custom --user-data-dir above is required, not optional.

3. Log in once

A Chrome window opens on LinkedIn's login page. Log in yourself. The profile directory persists — you won't do this again unless LinkedIn challenges you.

4. Configure

cp .env.example .env   # then edit

5. Wire into your client

Hermes (~/.hermes/config.yaml):

mcp_servers:
  linkedin:
    command: /path/to/linkedin-mcp/.venv/bin/python
    args: [/path/to/linkedin-mcp/server.py]
    enabled: true

Claude Desktop / Cursor / Codex (stdio):

{
  "mcpServers": {
    "linkedin": {
      "command": "/path/to/.venv/bin/python",
      "args": ["/path/to/server.py"],
      "env": { "LINKEDIN_PROFILE_URL": "https://www.linkedin.com/in/you/" }
    }
  }
}

Remote / HTTP mode — recommended config is OAuth (works with ChatGPT's "Connect" flow, Claude, and every serious MCP client):

OAUTH_USERNAME=you \
OAUTH_PASSWORD='generate-a-long-random-password' \
OAUTH_SECRET=$(python3 -c "import secrets;print(secrets.token_urlsafe(32))") \
LINKEDIN_MCP_PATH=$(python3 -c "import secrets;print(secrets.token_urlsafe(16))") \
PUBLIC_URL=https://your-tunnel-host \
.venv/bin/python server.py --http 8085 --public-url https://your-tunnel-host

Clients will pop your login page once, then hold a 30-day token. See CHATGPT.md.

Legacy alternative: LINKEDIN_MCP_TOKEN=<bearer> static token, or a bare secret LINKEDIN_MCP_PATH with no auth (not recommended).

How it works (and why the selectors look weird)

LinkedIn's 2026 "workspace" profile UI is adversarial to automation:

• edit dialogs are native <dialog open> elements — [role=dialog] attribute queries find nothing

• sections lazy-mount only when the main#workspace container scrolls (window scroll does nothing)

• per-card edit buttons are hidden until hovered or revealed by a generic control

• dialog inputs often have no aria-label — fields are identified by their current value

• project dialogs have no URL field anymore (put repo links in the description)

• skills list is newest-first: to reorder, delete the block and re-add in reverse

browser.py encodes all of this. When LinkedIn changes again, dom_dump is your friend.

Privacy

• state/, tokens, screenshots, and .env are gitignored — keep it that way

• the server stores nothing about you beyond the Chrome profile on your machine

• never commit your vanity URL into forks of this repo

Troubleshooting

License

MIT — see LICENSE. Not affiliated with or endorsed by LinkedIn.