The Self-Healing Local-Business Lead Scraper: An Agentic Claude Code Build That Harvests Leads (Even on No-API Sites) Straight Into Your CRM

The problem

Cold outreach and local-service lead-gen need a steady, hands-off pipeline of qualified prospects with real contact data — name, phone, website, rating, and ideally an email. The usual options all have a tax. Paid lead databases are expensive and often stale, and the moment you target a niche (dentists in one metro, HVAC firms by zip, real-estate offices on a strip) the rented list thins out fast. Scraping it yourself is worse: many target sites have no API, contact details are buried on a separate "Contact" or "About" page, and search sources throw bot-blocks the second you query at volume. Manual copy-paste is slow, breaks constantly, and quietly produces duplicates that pollute the CRM. The result is a sales team that either overpays for a database it doesn't trust or spends its mornings building lists instead of working them.

Who it's for

Agencies selling into local niches (dentists, HVAC, real estate, salons, law firms), outbound sales teams, SDRs, and solo operators building cold-outreach lists who want an owned, repeatable lead pipeline instead of a per-seat lead-database subscription — especially anyone whose target prospects sit on small no-API business sites that off-the-shelf scrapers choke on.

How it works

1. 1

   Define the job in plain English: tell Claude Code the niche, geography, volume, and destination — e.g. 'every Monday, find 25 dental practices in Austin and add them to my ClickUp Leads list with name, phone, website, rating, and email.' This plain-language spec becomes the agent's brief.

2. 2

   Have the agent build two tools: a find-leads task that queries SerpAPI / Serper or the Google Maps results page for businesses matching the niche and area, and a create-lead task that writes a clean record into your CRM or sheet. Claude Code scaffolds both and wires them together.

3. 3

   Enrich each lead from its own site: for every business returned, the agent drives a headless browser (Playwright) to open the website, click into the contact or about page, and extract the phone and email that the search result alone doesn't give you — including on small business sites that have no API.

4. 4

   Make the scrape self-healing: when a search source bot-blocks the agent, it fails over from Google to DuckDuckGo and retries instead of dying, and it caps retries and timeouts so one bad page never stalls the run. This self-correcting loop is what lets it survive no-API and bot-protected sites unattended.

5. 5

   Dedupe before writing: key every record on a stable identifier (Google Maps place_id or business ID) and check it against an existing-leads ledger so the same practice is never added to the CRM twice across runs.

6. 6

   Schedule it and wrap it into a reusable skill: deploy the agent on Trigger.dev or a cloud cron to run on your cadence with no human in the loop, and package the whole build as a reusable skill so you can re-point it at a new niche or city in one line.

Tools

The result

You get an owned, scheduled lead pipeline instead of a rented database: on whatever cadence you set, an agent discovers local businesses in your niche, opens each one's site to pull the phone and email that search results omit, skips anything it has already captured by place_id, and lands clean records in your CRM or sheet — with no one watching it run. The distinctive payoff over commodity Google-Maps scrapers is the agentic self-correction: because Claude Code drives a real headless browser and fails over from Google to DuckDuckGo when blocked, it keeps harvesting from small no-API and bot-protected sites that simpler scrapers skip, and the place_id dedupe keeps the list clean across runs. Honest caveat: the underlying "Google Maps to enriched contact to CRM" capability is heavily commoditized — paid tools like Outscraper, Apify's lead finders, and Scrap.io do the basic version cheaply, and search APIs bill per call — so the real value here is owning a self-healing, schedulable build you can re-point at any niche, not the raw data itself. Expect contact-field coverage (especially email) to vary by how much each business publishes, and respect each source's terms of service and rate limits.