The Model Context Protocol (MCP) quietly changed what “lead generation tool” means. Instead of exporting CSVs and importing them somewhere else, you hand your AI agent a set of tools and describe the outcome you want. The agent does the querying, filtering, enriching and drafting.
What an agent can do with database tools
With seven tools — search_stores, count_stores, get_store, get_store_catalog, get_store_ads, get_store_trend and get_quota — an agent can run a complete prospecting session: size a market, build a filtered list, pull each store's best-seller and Google reputation, then write outreach that references all of it.
The prompt that matters is not clever. It is specific: “Find French Shopify stores selling skincare with 50+ products and a Google rating under 4.3, then draft an email for each that references their best-selling product and their weakest recent review.” Specific inputs, verifiable outputs.
Setup in two minutes
In Claude Code: one command — claude mcp add eldb --transport http <server-url> with your API key as a header. In Cursor or Claude Desktop: five lines of JSON config. In n8n: an MCP Client node pointed at the same URL. The tools are discovered automatically, and every call is metered against your plan's quota — the agent can even check remaining quota itself.
Guardrails that keep it useful
Three rules from customers running this daily. One: make the agent count before it searches — count_stores is cheap and prevents burning search quota on a bad filter. Two: force it to cite the data field behind every personalized line, so hallucinated flattery can't slip in. Three: keep a human between draft and send. The agent builds the list and writes the copy; you press the button.
Prospecting used to be a job. Now it's a prompt. The people winning right now are the ones who understood that a month before their competitors.