# How to find leads using Claude or ChatGPT
> Connect AI agents like Claude or Cursor to a lead source via MCP and REST APIs to find and qualify LinkedIn prospects using natural language.
- HTML: https://getkinlead.com/find-leads-via-mcp-server
- Markdown: https://getkinlead.com/find-leads-via-mcp-server.md

**Short answer:** You can find LinkedIn leads by connecting an MCP client like Claude or Cursor to a lead source that exposes an MCP server or a REST API. The agent then runs searches, reads the profiles and scores each person against the customer profile you describe, so you review a shortlist instead of browsing.

## Why use an AI agent for lead sourcing

LinkedIn's search box matches words. Type "founder" and you get the people who wrote founder in their headline. You miss the ones who call themselves CEO, owner or managing partner, which is most of them. An agent reads the profile and decides whether the person runs the company, which is a different job from matching a string.

Agents also handle the filtering that eats your morning. Checking whether a company actually sells B2B software means opening the company page, reading the description and making a judgment about the business model. That judgment is exactly what a language model is good at, and it can do it for forty profiles while you make coffee.

The bigger change is what happens to your time. Searching and vetting feel productive and mostly aren't. Reviewing a scored list of forty people takes ten minutes. Building that same list by hand takes an afternoon, and you'll still be unsure about half of them.

**Manual search**

- Guessing keywords like founder or owner and hoping the right people used them
- Opening fifty tabs to check company size and industry
- Copying names into a spreadsheet with no record of why they qualified

**AI agent sourcing**

- Asking for people who run small agencies that sell to ecommerce brands
- Filtering by team size, industry and recent posting activity in one pass
- Getting a score and a one-line reason for every person on the list

## Setting up your AI environment

MCP is the piece that makes this work. It's a protocol that lets a chat client call tools on a server, so the model can run a search and read the result instead of answering from training data. Without it, the model guesses. With it, the model reads your lead source.

You need two things: a client that speaks MCP and a lead source that exposes one. Claude Desktop, Claude Code, Cursor and ChatGPT all take MCP connections. On the lead side, look for a provider that publishes an MCP server, or a REST API you can wrap yourself if you're comfortable writing a small script.

1. **Install an MCP client.** Claude Desktop, Claude Code, Cursor or ChatGPT all work. Pick the one you already have open all day.
2. **Get your credentials.** Find the server URL and either an OAuth connection button or an API key, depending on whether you're using a chat app or a script.
3. **Add the server to your config.** Paste the URL and key into the MCP settings file or the connector screen in your client, then save.
4. **Confirm the tools are visible.** Restart the client and ask it to list the available tools. If you see search and profile functions, you're connected.

> Keep API keys out of shared repos and screenshots. A leaked key on a lead source is a leaked contact list.

## Writing prompts that find qualified leads

Vague prompts produce vague lists. "Find me marketing leaders" returns a thousand people and no way to choose between them. Write the prompt the way you'd brief a new SDR on their first day: who the person is, where they work, what the company sells, and what they've done recently that makes them worth contacting now.

Start broad, then narrow. Ask for a first pass, look at what comes back, and add constraints based on what's wrong with it. If half the results are agency people and you sell to in-house teams, say so and exclude agencies. Two rounds of this usually gets you to something usable.

- **you:** Search for marketing managers in New York at B2B SaaS companies with 50 to 200 employees who posted about lead generation in the last 30 days.
- **draft:** I found 42 people matching those criteria. Do you want them scored against your customer profile, or filtered further first?
- **you:** Drop anyone who started the role in the last six months, then score the rest against a seed-stage founder profile and give me the reason for each score.
- **draft:** Down to 31. Top of the list is a head of growth at a 90-person SaaS company who posted about referral programs twice this month. Score 88 because the company sells to agencies and she owns the partner channel.

```
You are sourcing leads for [your product].

Find people on LinkedIn who match all of these:
- Title: [titles, including synonyms like owner, managing partner, head of]
- Company: [industry], [size range] employees, sells to [buyer type]
- Location: [regions]
- Recent activity: posted about [topic] in the last 60 days

For each person return: name, current title, company, a one-line description of what the company sells, and a fit score from 0 to 100 against the criteria above. Give the reason for each score in one sentence.

Skip anyone whose company description doesn't make clear what they sell. Return "unclear" instead of guessing.
```

## Qualifying leads against your customer profile

Finding names is the easy half. Deciding who's worth a message is where most lists fall apart, because the decision needs context the search filters don't capture. Does this company sell to businesses or consumers? Is this person in the buying group or adjacent to it? Did they just start the job, which usually means a budget reset?

Give the agent a written profile of your best customer and ask it to act as a gatekeeper. It reads the profile data and returns a decision with reasoning, and you check the reasoning rather than the number.

1. **Write the profile down.** One page: industry, company size, the buyer's job title, the problem they have, and two examples of companies that fit.
2. **Pull full profiles.** Ask the agent to fetch the profile and recent posts for each search result, not just the headline.
3. **Run the comparison.** Have it compare each profile against the written description and flag anything that contradicts it.
4. **Ask for reasons, not just scores.** Request a table with name, score from 0 to 100, and one sentence explaining the score. Reject any row with no reason.

> A score with no explanation is impossible to argue with or improve. Insist on the reason every time.

## Mistakes that produce bad lead lists

Most disappointing results come from the prompt, not the model. These are the four that show up again and again.

- Asking for a hundred leads at once. You can't review a hundred, so you won't. Ask for twenty, work them, then ask again.
- Accepting a score with no reason behind it. If the agent can't say why someone scored 80, the number is decoration.
- Ignoring when the profile was last active. A perfect-fit person who hasn't posted in two years is a cold contact with good credentials.
- Letting the agent fill gaps. When a company description is blank, the model will invent something plausible. Tell it to return unclear and move on.

## What to do with the leads once you have them

Export to CSV if your team lives in a spreadsheet or a CRM. A CSV of profile URLs is also the way to bring a list back into a tool later, so keep the file even after you've worked it.

Write the first line from the qualification reason. The agent already knows why the person is on the list, so ask it to open with that. "Saw you're hiring a second SDR" gets read. "I came across your profile" gets deleted.

Keep the list small enough to actually work. A hundred scored leads you never contact is worse than twenty you do. Set a fit threshold, keep everything above it, and archive the rest without guilt.

| | Stage | What you do | Why it matters |
| --- | --- | --- | --- |
| Export | Download the scored list as CSV | Your team works where it already works |
| First message | Open with the reason the lead scored high | The reason is the one part the person can verify |
| Follow-up | Reference something new from the profile | A second message that repeats the first gets ignored |
| Review | Check which prompts produced replies | The prompt is the part you can actually fix |

## Doing this with Kinlead

Kinlead runs an MCP server and a REST Agent API, so Claude, ChatGPT, Cursor, Claude Code and Codex can find leads, check replies and draft answers without you opening the app. Chat apps connect with OAuth. Scripts use API keys.

Lead search uses LinkedIn's standard search, so a Sales Navigator seat isn't required. Each lead gets a score from 0 to 100 against the customer profile you describe or your website, and leads below the fit line aren't kept. Leads sit in lead groups, and any campaign can point at any group.

> Test new prompts against a lead group you're willing to delete. Throwing away a bad list is easier than cleaning one.

## Related

- [Agent API and MCP](https://getkinlead.com/features/agent-api-and-mcp.md)
- [Claude integration](https://getkinlead.com/integrations/claude.md)
- [Lead groups and scoring](https://getkinlead.com/features/lead-groups-and-scoring.md)

## Common questions

**Do I need Sales Navigator to find leads with an AI agent?**

It depends on the lead source you connect. Some tools run on LinkedIn's standard search, so a Sales Navigator seat isn't required. Check what the provider uses before you buy a seat.

**How do I stop the AI from inventing lead details?**

Use a server that pulls real data rather than relying on the model's memory, and tell the agent to return unclear when a field is missing. Then spot-check a few profiles against what the agent reported.

**Can the agent send messages after it finds the leads?**

Only if the connected tool has an outreach API and you've set that up. Finding and qualifying leads is the part every MCP lead source can do. Sending is a separate capability you should confirm before relying on it.

**What's the difference between a REST API and an MCP server?**

A REST API is the raw data connection with endpoints you call yourself. An MCP server sits on top and tells the AI client which endpoints exist and how to use them, which is what lets you search in plain language.

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