lead-gen-agent

An open, agent-operated lead-generation loop. Open this repo in Claude Code and it finds you buyers: it scans Reddit, Twitter/X, Instagram, and TikTok for fresh posts by people actively looking for what you sell or complaining about your competitors, prioritizes the real asks, and writes a run report telling you exactly where to comment and who to reach. You do the replying, in your own words; the agent never contacts anyone.

Built and open-sourced by the team at xpoz.ai, based on the loop we run on ourselves. Social search is powered by the xpoz MCP server: billions of posts across the four platforms, queried in plain language by your agent.

Why this exists

Most lead-gen tooling sells you scraped contact lists and automated outreach, which is exactly the spam your buyers ignore. This is the opposite motion: find the people who are asking for your product right now, in public, and show up as a genuinely useful, disclosed human answer. The methodology is the product:

• Buying intent, not keywords. The query book is organized by intent buckets (Switching, Asking, Evaluating, Budgeting, Struggling), because "alternative to [competitor]" and "can anyone recommend" are the phrasings that precede a purchase, and each platform yields a different lead shape: Reddit produces threads to answer, X produces complaints and named prospects, TikTok and Instagram produce high-engagement comment spots.
• Freshness is a ranking criterion. The last 72 hours weigh heaviest; a three-week-old thread is backlog no matter how good the ask. Fresh threads are still watched, fresh askers still have the problem, and fresh threads become tomorrow's AI-cited surfaces.
• Judgment prioritization with a written why. Every lead lands as P1 (act now), P2 (worth engaging), or P3 (watch), each with a one-line rationale weighing intent strength, fit (including honest non-fits), freshness, and reach. No pseudo-precise scoring arithmetic.
• A query book that learns. Fixed core queries keep runs comparable week over week; a few freeform exploration slots per run discover new phrasings; queries that hit get promoted, queries that go dead twice get cut. Every run record ends with what worked and what was tuned.
• The agent never engages. It finds, prioritizes, and reports. Humans reply, publicly first, with their affiliation stated in the text. That constraint is what keeps the accounts, and the reputation, alive.

What you need

• Claude Code: the agent that operates the repo.
• An xpoz account: the search engine behind every scan. Connected via MCP: the first tool call opens a browser sign-in (OAuth), approve it and you're connected; a free key needs no credit card. A full scan across all four platforms costs on the order of 100 xpoz credits (the free tier covers your first scans; the $20/month plan covers weekly scans comfortably).

No API keys, no Python environment, no build step.

Quickstart

git clone https://github.com/XPOZpublic/lead-gen-agent.git && cd lead-gen-agent
cp .mcp.json.example .mcp.json
claude                      # open Claude Code in the repo

Then, inside Claude Code:

1. /setup: an interview plus web research about your product, personas, competitors, and pricing. Writes config.json, strategy/product-profile.md, seeds the query book with your competitor names and category phrases, and drafts your venue map.
2. /lead-scan: the run. Searches the four platforms, prioritizes what it finds, and writes memory/lead-scan-runs/<date>.md: this week's picks, per-lead reply angles, demand signals, and query tuning.
3. Reply to the picks in your own words, then tell the next scan what happened; outcomes are tracked run over run.

Weekly is the natural cadence. The first two or three runs are tuning runs: expect to cut dead queries and add phrasings discovered in real posts.

Layout

Path	What it holds
CLAUDE.md	Agent orientation: session ritual, structure, hard rules
strategy/lead-strategy.md	What counts as a lead per platform, the two search motions, freshness
strategy/engagement-playbook.md	Who to reach and where: personas and venue map (drafted by /setup)
strategy/product-profile.md	Your product's facts, ICP, and honest non-fits (written by /setup)
strategy/queries.md	The query book: intent buckets per platform, plus exploration slots
memory/journal.md	Session log, append-only
memory/lead-scan-runs/	Dated run records; conventions in its index
memory/seen-leads.json	Dedup ledger across runs
memory/competitors/	Competitor knowledge, researched into existence as leads surface them
examples/	A sample run record (fabricated data) showing what a scan produces

This is a fork-and-own repo: run records, the dedup ledger, and competitor files are committed alongside the code by design, so your copy becomes the durable record of your lead-gen loop. Clone to evaluate; before your first real scan, fork (or re-point origin to your own repo) so run records have somewhere to live.

The sibling project

geo-seo-agent runs the same philosophy on a different loop: AI-engine visibility (GEO) instead of lead generation. The two compound: the demand signals lead scans surface are content-brief fuel for a GEO program, and GEO-won visibility warms the buyers lead scans find.

License

MIT