There is no official LinkedIn MCP server, so the right URL depends on which community server you pick. This article shows the local and hosted endpoints, the exact config for Claude Desktop, Claude Code and Cursor, the login flow, and the tasks worth automating, plus the account risks.
Type "LinkedIn MCP server URL" into a search bar and the first surprise is that there is no single address to copy. LinkedIn does not publish an MCP server of its own, and as of October 2026 every option you can install or connect to comes from a community project or a paid third-party platform. That changes how you choose one, how you set it up, and how much risk you take on with your account.
This article shows what the popular open-source server actually exposes, which URL belongs to which setup, the exact config for Claude Desktop, Claude Code and Cursor, and the job, research and outreach tasks worth handing to an assistant. It also marks the line between a useful shortcut and a restricted account.
MCP stands for Model Context Protocol, an open standard that lets an AI assistant call outside tools through one consistent interface. A LinkedIn MCP server is a small program that wraps LinkedIn actions (reading a profile, searching jobs, listing inbox threads) as tools your assistant can call in plain English.
People reach for one for the same few reasons. Recruiters want a shortlist without forty open tabs. Job seekers want listings ranked against their own background. Founders want a quick briefing on a prospect before a call. Each of those is a repetitive read-and-summarize loop, which is exactly what an assistant with the right tools does well, and exactly what a manual browser session does badly.
You type something like "find remote product design roles posted this week". The assistant calls the server, the server reads LinkedIn through a browser session or an API, and structured text comes back. The assistant then ranks, summarizes or drafts from it. No more copying between tabs.
Three families exist, and the URL question depends on which one you pick.
| Type | How it reaches LinkedIn | Where it runs | Typical cost |
|---|---|---|---|
| Browser automation (the popular open-source server) | A logged-in browser session on your machine | Local, stdio | Free |
| REST API wrappers | An OAuth app approved by LinkedIn, limited to granted scopes | Local or hosted | Depends on your access |
| Managed scraping platforms | The vendor's infrastructure | Hosted remote URL | Paid |
💡 The most used open-source option is stickerdaniel/linkedin-mcp-server on GitHub, published on PyPI as
mcp-server-linkedin(it was previously namedlinkedin-scraper-mcp, and the old name still forwards to the new one). Every command, flag and tool name below comes from its documentation, and "the server" means this one.
For a personal account, the local browser-session server is the simplest path: no developer application to apply for, no third party holding your data, and a cost of zero. The trade-off is the terms-of-service risk described below, which is why it suits low-volume personal use. An API-based wrapper is safer on paper, but it only returns what your approved application is allowed to read, and LinkedIn grants those permissions selectively. A hosted platform saves setup time and costs money, and you accept a vendor in the middle.
The default transport is stdio. Your client launches the server as a child process and talks to it through standard input and output, so there is nothing to paste into a "server URL" field. What you give the client is a command: uvx mcp-server-linkedin@latest.
If your client only accepts a URL, start the server with the streamable HTTP transport:
uvx mcp-server-linkedin@latest --transport streamable-http --host 127.0.0.1 --port 8000 --path /mcp
The address your client connects to is http://127.0.0.1:8000/mcp. The Docker image documents port 8080 for the same job, which gives http://127.0.0.1:8080/mcp. Keep the host on 127.0.0.1. Anyone who can reach that port can act through your logged-in session.
Streamable HTTP is also the easier route when the server runs on a different machine from the client, for instance a home mini PC that stays on all day while you work from a laptop. In that case keep the port closed to the open internet and reach it over a private network or an SSH tunnel.
💡 In Claude Code, one line registers the HTTP address:
claude mcp add --transport http linkedin http://127.0.0.1:8000/mcp
MCP directories also list remote endpoints, for example https://linkedin.run.mcp.com.ai/mcp, and https://gateway.pipeworx.io/linkedin_ads/mcp (that one targets LinkedIn Ads, not personal profiles). I have not vetted either. A hosted server sees whatever data and credentials you route through it, so read the publisher's privacy terms first and prefer a local server for anything tied to your personal account.
| Setup | Address or command | Use it when |
|---|---|---|
| Local stdio | uvx mcp-server-linkedin@latest | One machine, Claude Desktop, Cursor or Claude Code |
| Local HTTP (uvx) | http://127.0.0.1:8000/mcp | Your client only accepts a URL |
| Local HTTP (Docker) | http://127.0.0.1:8080/mcp | You want isolation from your main system |
| Hosted | Set by the vendor | You accept a third party in the loop |
The server runs through uvx, which ships with the uv package manager (version 0.4.0 or newer). On macOS or Linux:
curl -LsSf https://astral.sh/uv/install.sh | sh
On Windows, use PowerShell:
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
Run the login command a single time:
uvx mcp-server-linkedin@latest --login
A browser window opens and you sign in by hand, two-step verification included. The session is stored under ~/.linkedin-mcp/profile/, and your password never passes through the assistant. Related flags:
--import-from-browser reuses a session from Chrome, Brave, Edge, Arc, Vivaldi or another Chromium browser.--auto-import is on by default and imports from a logged-in local browser on the first tool call.--logout clears the stored session.For Claude Desktop, edit claude_desktop_config.json (on Windows it sits in %APPDATA%\Claude, on macOS in ~/Library/Application Support/Claude). Cursor reads the same block from .cursor/mcp.json:
{
"mcpServers": {
"mcp-server-linkedin": {
"command": "uvx",
"args": ["mcp-server-linkedin@latest"],
"env": { "UV_HTTP_TIMEOUT": "300" }
}
}
}
For Claude Code, register it from the terminal:
claude mcp add linkedin -- uvx mcp-server-linkedin@latest
Prefer containers? Mount the session folder into the image. On Windows, write the path with forward slashes, like C:/Users/Alice/.linkedin-mcp:
{
"mcpServers": {
"mcp-server-linkedin": {
"command": "docker",
"args": [
"run", "--rm", "-i",
"-v", "/absolute/path/to/.linkedin-mcp:/home/pwuser/.linkedin-mcp",
"stickerdaniel/linkedin-mcp-server:latest"
]
}
}
}
Restart the client fully, then send this:
Read my own LinkedIn profile and tell me which sections look thin.
It calls get_my_profile, reads only your own data, and changes nothing. If it answers, the wiring works. The first run can be slow while the server sets up its managed browser, which is why the config above raises UV_HTTP_TIMEOUT to 300 seconds.
💡
@latestmeans the server updates itself on every launch. For a setup you rely on, pin a version such asuvx mcp-server-linkedin@4.14.0(the PyPI version at the time of writing).
The server exposes roughly twenty tools, and every call runs one at a time through a queue.
| Area | Tools | What you get |
|---|---|---|
| Profiles | get_person_profile, get_my_profile, search_people, get_sidebar_profiles | Experience, education, skills, projects, posts |
| Connections | connect_with_person | Send or accept requests with an optional note |
| Messaging | get_inbox, get_conversation, search_conversations, send_message | Threads and search, sending needs confirmation |
| Companies | get_company_profile, get_company_posts, search_companies, get_company_employees | Company pages, posts, jobs, employee lists |
| Jobs | search_jobs, get_saved_jobs, get_job_details, get_job_apply_url | Listings, saved roles, full postings, apply links |
| Feed | get_feed, search_posts | Home feed and post search with recency filters |
| Session | close_session | Releases the browser |
This is where recruiters and sales teams save the most time. Ask for a shortlist and a one-paragraph summary per person instead of opening fifteen tabs.
Search for senior data engineers in Lisbon, open the first eight profiles, and list who has Kafka and dbt in their experience.
search_jobs takes search terms and a location, then filters for date posted, job type, experience level, work type, Easy Apply and sort order, with results paginated up to ten pages. Pair it with get_job_details and get_job_apply_url and the assistant can turn a noisy results page into a ranked table with the apply links attached.
Find remote UX research roles posted in the last week, skip anything under mid-level, and rank them by how closely they match my profile.
I would keep the applying itself manual. The assistant finds and ranks. You click.
A routine that works: every Monday morning, ask for new listings from the past seven days that match two or three saved search ideas, have the assistant drop anything you already saw, and keep the top five in a short note. That is three or four tool calls, far below anything that looks like bulk activity, and you stay in charge of every application.
Combine get_company_profile, get_company_posts and get_company_employees to build a one-page briefing: what the company announced lately, who works in the team you are joining, and which open roles hint at its priorities. search_posts and get_feed work the same way for topic monitoring, such as "what are people in my industry saying about this launch?"
get_inbox and search_conversations let the assistant find threads that went quiet, and it can draft a follow-up for each one using the context of the conversation. send_message asks for confirmation before anything goes out. Treat that confirmation as a hard stop: read the draft, edit it, and only then approve.
The server's own documentation is blunt: "LinkedIn's User Agreement prohibits automated access, and accounts using automated tools can be restricted or banned." The project is independent, not affiliated with LinkedIn or Microsoft, intended for personal use, and comes with no warranty. Using it is your decision and your risk.
Here is my own rule of thumb, not an official threshold:
| Task | Risk | Verdict |
|---|---|---|
| Reading your own profile | Low | Fine |
| A few job searches per day | Low to medium | Fine in moderation |
| Opening dozens of profiles in a row | Higher | Avoid |
| Bulk connection requests | High | Don't |
| Auto-sent messages | High | Draft only |
If a task would make you uncomfortable doing by hand fifty times in an hour, it should not be something an agent does for you either. The goal is to replace tab juggling, not to run a campaign.
uvx on its PATH. Run which uvx (or where uvx on Windows) and paste the absolute path into the command field, then restart the client fully.--login once in a terminal so everything is fetched before the client asks, and keep UV_HTTP_TIMEOUT at 300.--logout, then --login again. Sessions expire.--tool-timeout (180 seconds by default) and each page operation by --timeout (5000 ms). On a slow connection, add "--timeout", "10000" to args.--no-headless and the browser window stays visible.The MCP server fetches data. A writing model turns it into something you can send. You can paste what your assistant returned into Claude Sonnet 5 on PicassoIA and shape the output there.
💡 Posts with an image usually stand out in the feed. P-Image, Flux 2 Pro, Seedream 5 Lite and GPT Image 2 all generate a clean 16:9 header from a one-line prompt.
An assistant wired to LinkedIn saves hours of searching, but the result still needs words and visuals that sound like you. Take the next draft your assistant produces, open Picasso IA, and build the header image for it. Describe the scene in plain language, pick a model, generate a few variations, and keep the one that fits the post.
Start with something small: a photo-style banner for your next job update or a clean image for a company announcement. Two or three experiments will show you which prompts and models suit your style.
