# How to manage LinkedIn leads with an MCP client
> Connect Claude, ChatGPT or Cursor to your LinkedIn lead data with an MCP server. Check replies, sort intent and draft answers without leaving the chat.
- HTML: https://getkinlead.com/using-mcp-for-linkedin-lead-management
- Markdown: https://getkinlead.com/using-mcp-for-linkedin-lead-management.md

**Short answer:** An MCP client puts your LinkedIn lead data inside Claude, ChatGPT or Cursor, so you can ask about replies, intent and drafts in one place instead of pasting profiles by hand. You connect the server once, then work in plain English and check the records the assistant quotes back to you.

## What an MCP client actually changes about your day

A chat window with nothing connected knows only what you paste into it. You copy a profile, paste a thread, describe your offer, then wait. Do that five times and you've spent more time feeding the assistant than selling.

An MCP server closes that gap. It's a small service that exposes tools the assistant can call: search leads, fetch a reply, read a profile. You ask in plain English, the assistant picks a tool, the data comes back, and the answer is built from records rather than your memory of them.

The real change is scale. One lead in a chat is a toy. Forty leads in a chat, where the assistant can read each one and sort by who asked about price, is a morning's work finished in ten minutes.

It also changes what you can ask. Without a connection, every question has to fit in one paste. With one, you can ask across your whole pipeline: who replied twice and got no answer, which leads went quiet after I sent the offer, who mentioned a competitor.

**Pasting by hand**

- Copying a profile into the chat and asking for a message
- Retyping three replies so the assistant can hear the tone
- Opening five tabs to check who answered yesterday

**Connected client**

- Asking which replies mention a budget
- Asking for a draft that answers the thread the lead actually sent
- Asking who replied twice and never got a follow up

## Setting up the connection in Claude, ChatGPT or Cursor

You need two things: a client that speaks MCP, and credentials for the server you want to reach. Chat apps usually connect through a sign-in flow that opens a browser tab and asks you to approve access. Editors and scripts usually take an API key you paste into a config file.

Keep those paths apart. A key sitting in a config file on your laptop is fine there and wrong on a shared machine. Check the client's own settings page before you guess which route it wants.

Name each server for the job it does. When an assistant lists its tools to you, leads and replies should be obvious at a glance. Two servers called server1 and server2 tell you nothing about which one just ran.

1. **Open the connectors or MCP screen.** In a chat app this usually sits under settings, sometimes labelled connectors or integrations. In an editor, look for a JSON config file with an mcpServers block.
2. **Add the server.** Paste the server URL and finish the sign-in, or paste the key next to the server name. Save the file.
3. **Restart and list the tools.** Ask the assistant to list the tools it can see. If the lead tools aren't there, the connection failed and you're about to get a confident answer invented from nothing.
4. **Run one cheap test.** Ask for your most recent reply or a count of your lead groups. A real record with a real name proves the path works before you trust a long answer.

> If the assistant can't name the tool it just used, assume it didn't call one. Ask it to quote the record it read.

## Checking replies and finding who is actually interested

Replies are where this pays off fastest. Your inbox holds people who asked a real question, people who said not now, and people who replied thanks and meant nothing. Reading it top to bottom is slow and it puts the wrong person first.

Ask for recent replies, then ask for each one to land in a bucket: asked about price, asked about timing, asked for a call, not interested, unclear. Then ask it to quote the line it used for the sort. Quoting is what stops the assistant from guessing.

One habit makes the output worth reading: ask for one action per lead. Tell me the next step for each of these nine gives you a list you can work. What should I do about my replies gives you an essay.

- **you:** Pull my LinkedIn replies from the last two weeks. Group them by what the person wants and quote the line you used for each one.
- **draft:** Called the replies tool. Nine replies. Three asked about price, one quoted: what does this cost for a team of five. Two asked for a call, one quoted: can we do Thursday. Four are unclear or not interested.
- **you:** For the two who asked for a call, draft a short reply with two time slots and use the detail they mentioned.
- **them:** Thursday works. Also, we're a 12-person agency so the seat question matters.
- **draft:** Thursday at 10 or 2 works on my side, whichever suits you. I'll put the seat question first on the agenda since 12 people changes the answer. Shall I send the invite?

## Drafting replies that don't sound like a template

The assistant writes badly when you give it nothing. Write a friendly follow up returns the same message everyone else sends. Hand it the thread, the lead's own words and the outcome you want, and the draft gets specific.

Draft one message at a time. Batch drafting pulls every draft toward the same shape and you delete most of them.

Keep a short house rule in your prompt: no greeting formula, no line about how busy they must be, one clear next step. That single sentence removes most of the stiffness you'd otherwise edit out by hand.

**Prompt that gets ignored**

- Write a follow up to this lead
- Make it sound professional and friendly
- Mention that we help teams like theirs

**Prompt that gets a usable draft**

- They asked how onboarding works for a team of 12. Answer that in two sentences.
- They posted about hiring SDRs last week. Use it once, in the second line.
- End with a question that takes a yes or a no.

> Read every draft before it sends. Your name is on it, and the assistant has no idea which of your claims are true.

```
Read the last four messages with this lead and the note I sent with the connection request. Then write a reply that answers the question they actually asked, uses one detail from their profile or a recent post, and ends with one clear next step. Keep it under 60 words. No greeting formula. Mention our product name once at most.
```

## Finding new leads without leaving the chat

The connection that reads replies can also search. Describe the buyer the way you'd describe them to a colleague: title, company size, the thing they post about. The assistant turns that into a search and comes back with matches.

Say what you don't want. Marketing leads at B2B software companies, 20 to 200 people, not agencies, not current customers gives it something to filter on. Good leads gives it nothing.

Ask why each person matched. You'll spot a bad search term in the reasons before you spot it in the names.

| | Task | What the assistant can do | What you still decide |
| --- | --- | --- | --- |
| Find people | Turn your description of a buyer into a search and return matches with a reason for each | Whether that description is the customer you actually want |
| Read a profile | Summarise the role, the company and anything the person wrote recently | Whether they're worth a message this month |
| Sort replies | Group by intent and quote the line that decided the group | Who you answer first |
| Draft | Write a reply from the thread and your goal | Whether anything in it is true |

> Export the list before you build on it. A CSV of profile URLs is easy to check, easy to hand to someone else, and hard to lose when a chat window closes.

## Where this setup goes wrong

A handful of failures cover most of it.

- It invents a detail. A shared connection that doesn't exist, a company that changed its name last year. Ask for the source line and the invention drops out.
- It answers about the wrong person. Two leads named Sarah at two companies. Put the company name in the question.
- It sounds certain about old data. If the connection dropped mid conversation, the assistant may keep answering from what it read earlier. List the tools again when an answer feels off.
- It drafts claims you can't stand behind. Check anything touching price, timeline or what the product does before you paste it into LinkedIn.

> The fix for all of these is the same: make the assistant name the record and quote the line. An answer you can trace is usable. One you can't is a guess with good grammar.

## Doing this with Kinlead

Kinlead has a REST Agent API and an MCP server, so Claude, ChatGPT, Cursor, Claude Code, Codex and other MCP clients can find leads, check replies and draft answers inside the chat. Chat apps connect with OAuth and scripts use API keys. Lead search runs on LinkedIn's standard search, so a Sales Navigator seat isn't required.

Leads sit in lead groups and each one is scored 0 to 100 against the customer you described, so what your assistant pulls is already filtered rather than everything LinkedIn returned. Replies from every campaign land in one inbox, classified by intent with interested replies first, and you get an email when someone replies. Details are on the [Agent API and MCP page](/features/agent-api-and-mcp).

A note on sending: Kinlead sends from your own LinkedIn account through Unipile, at most one action every 5 minutes inside the daily limits and send window you set. The AI writes messages and follow ups from your campaign goal, tone and sales playbook, and queued messages can be read and edited on the Actions page before they go.

## Related

- [Agent API and MCP](https://getkinlead.com/features/agent-api-and-mcp.md)
- [MCP integration setup](https://getkinlead.com/integrations/mcp.md)
- [Reply drafts](https://getkinlead.com/features/reply-drafts.md)
- [Unified inbox](https://getkinlead.com/features/unified-inbox.md)

## Common questions

**Do I need to write code to use an MCP client?**

Not much. Setup is usually a pasted URL with a sign-in, or a key added to a JSON block. After that you type in plain English.

**Can the assistant send LinkedIn messages through MCP?**

Treat the client as a reader and a drafter. Writing the message in the chat and sending it from your own LinkedIn account keeps you in control of what goes out under your name.

**Why did my client show no lead tools after setup?**

The connection failed or the server isn't authorised. Ask the assistant to list its tools, and if the list is empty, redo the sign-in or check the key before asking it anything about your pipeline.

**How do I stop the assistant inventing details about a lead?**

Ask it to quote the record behind every claim, and give it the company name in your question. When it can't quote a line, the detail didn't come from your data.

**Can I use this for research rather than replies?**

Yes. Describe your buyer, ask for matches with a reason each, then export the list and check it before anyone gets a message.

[Create a Kinlead account](https://getkinlead.com/signup) or [book a call](https://getkinlead.com/demo) to talk through pricing and setup.
