io.github.Raffaele86/local-leads-mcp

local-leads-mcp

Find local businesses that need a website — and prove it — from your AI agent.

One MCP server that discovers local businesses via OpenStreetMap and then verifies them with the checks that separate a real lead list from a pile of false positives. Built for freelancers and agencies that use Claude, Cursor or any MCP client for prospecting.

No API keys required for the core workflow: discovery, website verification, email deliverability check and weak-site scoring all run on free, open infrastructure (Overpass, Nominatim, DNS-over-HTTPS, plain HTTP).

Why this and not a maps scraper?

The hard part of local prospecting isn't getting a list of businesses — it's not embarrassing yourself. Lead lists built from raw map data are full of traps this server checks for explicitly:

• "No website" that isn't. OSM data can be stale; a missing website tag proves nothing. verify_website checks apex and www and redirects (an apex domain that doesn't resolve while the site lives happily on www. is a classic false "no website"), and confirm_no_website uses a web search as the final word.
• A Facebook page is not a website. Social profiles, booking aggregators (Fresha, TheFork, Treatwell…), directory listings and free subdomains (*.wixsite.com, *.business.site…) are classified as not_own_site, not as "has a website".
• Emails that bounce. check_mx predicts deliverability before you ever write to a lead (no MX = guaranteed bounce; PEC-only domains are flagged).
• "The site is bad" needs evidence. score_website returns citable issues (no HTTPS, no mobile viewport, missing title/meta/H1, no sitemap…), not vibes — optionally with Lighthouse scores via the PageSpeed API.

Every verdict ships with the raw checks that produced it, so your agent can show its work.

Quickstart (60 seconds)

{
  "mcpServers": {
    "local-leads": {
      "command": "uvx",
      "args": ["local-leads-mcp"]
    }
  }
}

That's it — no credentials. Then ask your agent things like:

Find 20 restaurants in Turin with no website, verify each one, and give me a CSV of the confirmed ones with a phone number.

Optional environment variables

Tools

Discovery (OpenStreetMap — no key)

• find_businesses(area, category, website_filter, phone_prefix, require_phone, limit) — businesses in a city/district, filtered by website presence. ~35 friendly categories plus any raw OSM key=value tag.
• list_categories() — the category → OSM tag map.

Verification (no key)

• verify_website(domain_or_url) — verdicts: live_own_site, not_own_site, domain_exists_no_site, no_website_found; includes per-URL evidence.
• check_mx(domain) — email deliverability pre-check via DNS-over-HTTPS.

Qualification (no key; PageSpeed optional)

• score_website(url) — 0–100 weak-site score with named issues + CMS/tech detection (WordPress, Wix, Shopify, …).

Confirmation (needs a search key)

• confirm_no_website(business_name, city) — the final word on "really no website?", classifying search results into own-site candidates vs listings/socials.

Export

• export_leads(leads_json, format) — CSV, JSON or Markdown table.

Ethics & fair use

• Data sources are public: OpenStreetMap (© OSM contributors, ODbL), public DNS, and the businesses' own public pages.
• OSM requests are rate-limited locally (min 1 s between calls) and carry a descriptive User-Agent, per the Nominatim and Overpass usage policies. For heavy workloads, run your own Overpass instance.
• This server discovers and verifies — it does not scrape personal emails and it does not send anything. How you contact a business is on you: follow your local rules (GDPR and friends).

See also

seo-stack-mcp — the same idea for SEO data: Google Search Console, GA4, Bing Webmaster and Microsoft Clarity in one MCP server. Prospect with local-leads, then monitor the sites you build with seo-stack.

License

MIT.

Built by Raffaele Nocera — web consulting for local businesses. If this tool helps you win a client, I'd love to hear about it.