How to use MCP clients for LinkedIn prospecting

Connect Claude, ChatGPT or Cursor to your LinkedIn lead data with an MCP server. Find leads, sort replies and draft answers without leaving the chat.

Get started

Short answer

An MCP client connects an AI app like Claude, ChatGPT or Cursor to a server that holds your lead data. You ask in plain language, the model calls the server, and you get current leads, replies and drafts instead of a stale paste.

What an MCP client actually does for prospecting

Most people use AI for sales by pasting a profile or a message thread into a chat window. The answer comes back fast, but it's a snapshot of a moment that has already passed. By the time you act on it, three new replies have landed and the lead you were excited about has gone quiet.

An MCP client keeps the connection live. MCP stands for Model Context Protocol, and it's a way for an AI app to call tools on a server instead of guessing from training data. Ask for today's replies and the model asks the server for today's replies. Ask it to find agency founders in Manchester and it runs a real search rather than inventing plausible names.

That matters most on LinkedIn, where the useful signal is always the newest thing. Who replied an hour ago. Who commented on a competitor's post yesterday. Who viewed your profile this morning. None of that is in a CSV you exported last week, and none of it is in the model's weights.

Copy and paste

  • Export a CSV of leads, upload it, watch it go stale in a day
  • Paste each reply thread by hand before asking for a draft
  • Rely on memory to tell the AI who is warm right now

MCP client

  • Ask for leads matching a description and get current results
  • Ask which replies came in today and get them labelled
  • Ask for a draft with the real thread already in context

Setting up the connection

You need two things: a client that speaks MCP and a server that holds your lead data. The client is the window you type in, so Claude Desktop, ChatGPT, Cursor, Claude Code or Codex. The server does the fetching and keeps the credentials.

Setup is a config file, not code. You add the server address and a credential, restart the app, and the tools appear in the chat. Chat apps usually connect through OAuth, which means a browser sign-in instead of a pasted key. Scripts and terminals use API keys.

  1. 1

    Pick a client you already use

    Claude Desktop and Cursor are the quickest to set up on a laptop. Claude Code and Codex suit people who already work in a terminal.

  2. 2

    Find the config file

    In Claude Desktop it sits in the app settings folder. In Cursor it's the MCP section of settings. The path differs by operating system, so check the client's own docs rather than a blog post.

  3. 3

    Add the server entry

    Paste the server URL and your credential. If the client supports OAuth, sign in through the browser window it opens and skip the key entirely.

  4. 4

    Restart and check the tool list

    Quit the app fully, reopen it, and ask what tools are available. If the server isn't listed, the config has a typo or the key is wrong.

If the tool list is empty, the problem is almost always the config file, not the model. Check the JSON before you rewrite your prompt.

Finding leads by describing the customer

You don't need Boolean strings. Describe the person you want the way you'd describe them to a colleague, then narrow from there. The model turns that into a search, runs it, and comes back with names you can check.

The first pass is usually too broad. That's normal and cheap to fix. Ask for a tighter description rather than a longer list of exclusions, because exclusions fight each other and a sharper description doesn't.

Example

  1. You

    Find founders of agencies with 20 to 50 people in the UK who posted about hiring a salesperson in the last month.

  2. Suggested reply

    Running the search now. Filtering by company size, location and recent post keywords.

  3. Prospect

    I found 14 people. 9 look like agency founders, 3 are sales leads at agencies, 2 are recruiters. Want me to drop the recruiters?

  4. You

    Drop the recruiters and keep the rest in a group called UK agency founders.

  5. Suggested reply

    Done. 12 leads saved to UK agency founders.

Sorting replies without leaving the chat

Triage takes more time than sending. With a client connected you can ask for the state of the inbox in one sentence and get an answer that reflects what's actually there, not what you remember from Monday.

Then pick one reply and ask for a draft that answers the specific thing the person raised. Generic drafts come from generic prompts, so paste the actual message into the request or let the tool pass it in for you.

  1. 1

    Ask for the day's replies

    Something like: list every reply from the last 24 hours with the person's name and what they asked. One request, one table.

  2. 2

    Ask for intent labels

    Interested, not now, wrong person, question. Ask the model to put its reason next to each label so you can check the reasoning instead of trusting the tag.

  3. 3

    Draft against the real message

    Pick a reply and ask for an answer that addresses the point they made. Give the model your goal and tone, or paste your own script so it sounds like you.

  4. 4

    Send from the app, not the chat

    The client can draft. The send happens in your outreach tool, where the message sits in a queue and you can edit it before it goes.

Check the labels on polite rejections. Sounds great, let's talk in Q3 is not an interested reply, and a model reading only tone will call it one.

Where the chat window stops being useful

MCP is good at reading and drafting. It's a poor place to run the sending, because sending needs pacing, daily limits and a record of what went out. Those live in the outreach tool, not in a chat session you'll close in an hour.

The split is worth knowing before you build a routine around the chat window. Use the client to think and look things up. Use the app to act, because that's where the limits and the record live.

Finding leads

Task
Good for exploring a new customer description
In the chat window
Needed when leads must be saved, scored and reused

Reading replies

Task
Good for a summary of the day
In the chat window
Needed for one inbox across every campaign

Sending messages

Task
Drafting only
In the chat window
Queuing, pacing and pausing happen here

Checking what went out

Task
You ask and hope the log is complete
In the chat window
The actions page lists every queued message

Working with a teammate

Task
Awkward at best
In the chat window
Shared workspaces with separate logins

Doing this with Kinlead

Kinlead runs an MCP server, so Claude, ChatGPT, Cursor, Claude Code and Codex can find leads, check replies and draft answers from the chat. Chat apps connect with OAuth and scripts use API keys. You describe the customer you want or point at your website, and each lead is scored from 0 to 100 against that profile. Leads below the fit line aren't kept.

Replies from every campaign land in one inbox, sorted by intent with interested ones first, and you get an email when someone answers. Drafts are written from the campaign goal, tone and your own sales playbook. Sending stays in the app, where each LinkedIn account sends at most one action every 5 minutes inside the limits and send window you set. The Agent API and MCP page covers the connection details.

A config you can copy

Most clients take the same shape: a named server, a way to reach it, and a credential. Replace the URL and the key with your own and save the file.

Restart the client after every edit. Most connection problems are a config file that was saved but never reloaded.

Paste this job

{
  "mcpServers": {
    "leads": {
      "url": "https://your-server.example.com/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Frequently asked questions

Related

Get started with Kinlead

Prefer markdown? How to use MCP clients for LinkedIn prospecting.md