LeadHub — B2B Lead Discovery for AI Agents

📋 Project Overview

When we use Cursor or other agents for outreach, the bottleneck is rarely “writing the email” — it’s finding who to contact with a real business email and enough context to sound relevant.

LeadHub is an open-source MCP server that gives agents a freer, tool-native way to hunt B2B leads: search the web, score fit, pull public emails, verify MX records, and export contacts — without baking an LLM API into the tool itself. The agent keeps the reasoning; LeadHub does the legwork.

🚀 Key Features

Core Implementation

• One-shot leadhub_find_leads — discover → qualify → harvest → fetch → save → export in a single MCP call
• Agent-native design — no embedded LLM; the agent reads page excerpts and writes names / pitch fit
• Rule-based ICP scoring — filters listicles, media sites, and low-fit domains before enrichment
• Email pipeline — classify (personal / role / business), junk filtering, MX verification
• Private campaigns — local YAML configs stay gitignored; public templates live under examples/

Technical Highlights

• FastMCP tool surface for Cursor and compatible clients
• theHarvester OSINT subprocess for domain-level email / subdomain signals
• Deterministic fetch layer — contact / about / team pages for agent-readable excerpts
• SQLite + CSV/JSON export for downstream CRM or manual review

💻 Project Detail

Typical flow:

1. User tells the agent a cooperation brief (what we sell, who we want)
2. Agent confirms once, then calls leadhub_find_leads
3. LeadHub searches keywords (DuckDuckGo via ddgs), extracts domains, scores ICP
4. For qualified sites: OSINT harvest + contact-page fetch → pick best email → verify → save
5. Results land in output/contacts_*.csv with company snippet, email, sources, cooperation-fit stub

CLI batch mode is still available for scripted runs via campaign YAML.

📊 Project Impact

• Cuts MCP tool spam from 10+ per-domain calls down to one for common “find leads” requests
• Works across verticals (we’ve tested export / industrial buyers and SaaS-style outreach)
• Fully open source (MIT) — agents can run it locally with no vendor LLM key inside the server

🎮 Links

• GitHub: https://github.com/Harvey-Yuan/leadhub

🛠️ Technology Stack

Agent Integration:
  - Model Context Protocol (MCP)
  - FastMCP
  - Cursor Agent Skills

Discovery & OSINT:
  - DuckDuckGo Search (ddgs)
  - theHarvester
  - HTTP fetch + BeautifulSoup

Data & Quality:
  - Python 3.12+
  - SQLite
  - dnspython (MX verification)
  - Rule-based ICP / email classification

CLI:
  - Click
  - YAML campaigns (local, private)

LeadHub is our take on “give agents hands, not another chatbot”: public-data lead discovery that stays composable with whatever model the user already runs.