Find Google Maps leads from Claude with MCP: setup, prompts and limits

You can find local business leads from an AI assistant by connecting it to a lead tool's MCP server: then you ask "find roofers in Denver with verified emails" in plain English, and the assistant starts the search, checks on it, and reads back businesses with their emails and grades. MCP (Model Context Protocol) is the open standard that lets assistants like Claude call outside tools. With Leadraft's server, Claude Code, Claude Desktop, Cursor and VS Code can connect today with an API key. ChatGPT can't connect to it directly yet, for a reason explained below.

What is MCP, in plain terms?#

MCP is an open-source standard for connecting AI applications to external systems. The official site compares it to a USB-C port for AI applications: one standard plug, many tools. Anthropic introduced it in November 2024, and clients including Claude, ChatGPT, VS Code and Cursor now support it.

An MCP server offers a set of tools, each with a name, a description and parameters. An MCP client (your assistant) reads those descriptions and decides when to call a tool, based on what you asked. You don't write code or call the tools yourself.

For lead generation, that means the assistant can do the clicking for you: start a search, wait for it, filter the results, and turn them into a table or a summary.

Which AI assistants can connect to Leadraft's MCP server?#

Any MCP client that can connect to a remote server over streamable HTTP and send an API key in the Authorization header.

On ChatGPT: its connectors use OAuth, and clients that only connect through OAuth can't send an API key header, so ChatGPT can't connect to Leadraft's MCP server yet. If you live in ChatGPT, the REST API is the alternative for now.

How do I connect Claude Code or Cursor to Leadraft?#

Three steps, about a minute:

1. Create an API key. In Leadraft, open Settings and create an API key. Use a read-only key if the assistant should never start searches.
2. Add the server to your client.
3. Ask for businesses in plain English.

The server details:

• URL: https://app.leadraft.co/api/mcp/
• Transport: streamable HTTP (stateless, JSON responses)
• Auth: Authorization: Bearer <your key>, the same keys as the REST API

For Claude Code, one command:

claude mcp add --transport http leadraft https://app.leadraft.co/api/mcp/ \
  --header "Authorization: Bearer lg_your_key"

Claude Code stores this for the current project by default. Add --scope user to use it in all your projects.

For Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "leadraft": {
      "url": "https://app.leadraft.co/api/mcp/",
      "headers": { "Authorization": "Bearer lg_your_key" }
    }
  }
}

For Claude Desktop (claude_desktop_config.json), bridged through mcp-remote:

{
  "mcpServers": {
    "leadraft": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://app.leadraft.co/api/mcp/",
               "--header", "Authorization: Bearer lg_your_key"]
    }
  }
}

Snippets for VS Code and the rest are in the MCP server docs.

What tools does the assistant get?#

Six. The assistant picks which to call; you only ask. Behind start_search, the same pipeline runs as in the web app:

One limit to know: start_search uses the default scope, which is everything the search returns. To count only businesses inside a place, or inside it plus a ring around it, start the search in the app or through the API.

What can I ask? Example prompts#

Start simple, then chain steps. These all work with the six tools above:

• "How many business credits do I have left this month?"
• "Find roofers in Denver, CO with verified emails. Tell me when the search is done."
• "Is my roofers search done yet? How many have an email?"
• "Show me the 20 best-rated dentists from that search, with their emails, as a table."
• "Start searches for HVAC companies in Phoenix, Tucson and Mesa, AZ."
• "Which of yesterday's plumbers have a rating above 4.5 and a website?"

Where an assistant earns its keep is the work around the list:

• "From my Austin plumbers search, group the Verified leads by category and draft one first line per group, using their city and review count."
• "Compare the dentist searches I ran in Tempe and Mesa: how many businesses, how many with a Verified email, and the average rating in each."
• "Get the full record for the top five leads and tell me which have a contact form on their site but no email."

Check what the assistant writes before you send anything. It works from the data the tools return, but it can still summarize wrongly.

Is this "vibe prospecting"?#

Yes, if you mean the same thing most people do by it: describing the leads you want in plain English and letting an AI assistant run the search, rather than building filters by hand. The term is informal. What decides whether it works is not the prompt but the tool behind it: whether it covers the whole area, keeps to your trade, and checks emails before they reach your sequencer.

What guardrails stop an assistant from overspending?#

The same rules as the app, plus a few habits.

• Same credits. Searches use business credits as businesses are found, and each business is charged once, ever, however it was found. Reading a Verified email uses a checked-email credit the first time.
• Read-only keys. A read-only key can call everything except start_search. Give one to any assistant that only needs to read and summarize.
• Ask for a budget check. Start sessions with "Check my credits first, and confirm the number of search lines before starting anything."
• Rate limit. Tool calls count against the key's limit of 120 requests a minute.
• Your data only. A key belongs to your organization; the assistant can only see your searches and businesses. Revoke a key any time in Settings.

MCP lead tool or a web-browsing agent: which should I use?#

Use a web-browsing agent for a handful of businesses or one-off research, and a lead tool over MCP for lists. Browsing agents (assistants that drive a real browser) can open Google Maps, read a page of results and visit websites. That's useful for "tell me about these three bakeries". It's a poor fit for "every bakery in Chicago":

There's a third option worth knowing: Google's own Maps MCP server, Maps Experience agent tools (formerly Grounding Lite). It's built for answering questions about places, weather and routes, not for lists: as of October 2026 its place search returns up to 5 places per call on the Essentials tier (6 to 10 on Pro, for select customers), at $7 per 1,000 requests after a free monthly allowance, and Google requires that its Maps content isn't cached or stored by the model. If you want an assistant to recommend a restaurant, use Google's. If you want a list of every restaurant in a county, use a lead tool.

Is it legal to use AI agents for lead generation?#

The same rules apply as for any other prospecting. Google's Maps Platform terms say customers won't "export, extract, or otherwise scrape Google Maps Content for use outside the Services", which is a contract matter. Data-protection law applies to personal data, even when it's public. And email law, such as CAN-SPAM in the US, applies to whatever you send. An AI assistant doesn't change any of that. Read our take on the legal questions and get legal advice for your situation.

For the rest of the process, from keywords to sending, see the Google Maps lead generation playbook.

FAQ#

What is MCP?#

The Model Context Protocol is an open standard, introduced by Anthropic in November 2024, for connecting AI applications such as Claude or ChatGPT to external tools and data. An MCP server exposes tools; the assistant decides when to call them.

Can I use ChatGPT to find leads with MCP?#

Not yet. Leadraft's MCP server needs an API key in the Authorization header, and ChatGPT's connectors use OAuth, which can't send one. Use Claude Code, Claude Desktop (through mcp-remote), Cursor or VS Code, or the REST API.

Does finding leads through an AI assistant cost extra?#

Not with Leadraft. Searches the assistant starts use the same business credits as the app, and reading a Verified email uses a checked-email credit the first time, exactly as in the app. Ask the assistant to check credits before a big search.

Can an AI agent just browse Google Maps for leads instead?#

It can open Maps in a browser and read a page of results, which is fine for a handful of businesses. For a full list it has to scroll and copy one result at a time, it can mis-copy details, and it doesn't verify emails. Google's Maps terms also restrict scraping. A lead tool exposed over MCP returns structured rows with graded emails.

How do I stop the assistant from starting expensive searches?#

Give it a read-only API key, which can read leads and searches but not start new ones. With a write key, ask it to call credits first and to confirm the number of search lines before starting.

Sources#

1. Model Context Protocol, What is the Model Context Protocol (MCP)?
2. Anthropic, Introducing the Model Context Protocol (November 25, 2024).
3. Anthropic, Connect Claude Code to tools via MCP (claude mcp add --transport http, --header, scopes).
4. Google, Maps Experience agent tools (formerly Grounding Lite) (tools, result limits, pricing, storage requirements).
5. Google, Google Maps Platform Terms of Service, section 3.2.3.
6. Leadraft, MCP server docs (URL, transport, auth, tools, client snippets).