AgentScout — automated "sell the screenshot" outreach for Agent Readiness

Fully automatic cold-outreach machine on GitHub Actions, $0/month: discovers EU businesses in your niche → audits how AI assistants talk about them (buyer-intent prompts across a panel of free models) → drafts a cold email led by the concrete finding ("you appeared in 0 of 8 AI answers; X was recommended instead") → opens a GitHub issue telling you exactly which screenshot to take → sends via Gmail with EU country gates, caps and suppression → detects replies over IMAP and opens a 🔥 issue when a prospect answers.

Successor to gtm-lite, but on autopilot: no human approval step (switchable).

Setup (15 minutes)

1. Create a private GitHub repo, push this folder.
2. Gmail: use a fresh/secondary Google account (protect your main domain's reputation). Enable 2FA → create an App Password (Google Account → Security → App passwords).
3. Repo → Settings → Secrets and variables → Actions, add:
   • GMAIL_ADDRESS, GMAIL_APP_PASSWORD
   • OPENROUTER_API_KEY (free at openrouter.ai — optional but recommended: it's the 2nd audit-panel model) and/or NVIDIA_API_KEY (build.nvidia.com, free)
   • GITHUB_TOKEN is automatic — it powers GitHub Models (free LLM) and issues.
4. Edit config/config.yaml: your from_email, postal_line (legally required), and set auto_send: false for the first 2 runs — check the drafts in data/leads.csv, then flip to true.
5. Actions tab → run "AgentScout pipeline" manually once. Done — it now runs twice every weekday.

Your only manual job

Watch for issues labeled screenshot-todo: each gives you 3-4 exact prompts to paste into ChatGPT/Gemini. Screenshot the worst fail, save it as screenshots/<lead-id>.png (filename is in the issue), commit. The follow-up email (day 4) attaches it automatically. First emails go out text-only with the statistics — the screenshot arms the follow-up punch.

And issues labeled hot-reply — that's a prospect answering. Close them yourself; that part will never be automatic.

Niche strategy

config/niches.yaml ships with 4 EU-tuned niches; run.active_niches round-robins between them. Give each niche ~60-80 sends (2-3 weeks) before judging. Kill anything with zero replies; double down where you get 2+. Dental tourism and relocation law are the highest-value-per-client starts; B2B SaaS is the biggest pond.

The legal line (read once, seriously)

B2B cold email with accurate sender info + opt-out is workable in most of the EU, but DE/AT/NL/DK/CH effectively require consent — they're blocked in policy and should stay blocked. The system prefers generic inboxes (info@) over named people, keeps volume at 16/day, includes your postal address and an opt-out line, and auto-suppresses anyone who replies "no thanks". Fill in postal_line before sending anything. This is engineering hygiene, not legal advice.

Costs

Piece	Free tier	Our usage
GitHub Actions	2000 min/mo	~10 min/day
GitHub Models	free w/ GITHUB_TOKEN	audits + drafts
OpenRouter :free	50 req/day	2nd audit model
DuckDuckGo (ddgs)	keyless	discovery
Gmail SMTP/IMAP	free	16 sends/day, reply scan

Files

src/discover.py   find companies (DDG) + extract emails/country from sites
src/audit.py      buyer-intent prompt panel + visibility scoring
src/outreach.py   compose / screenshot issues / send / reply detection
src/run_all.py    orchestrator (niche round-robin)
config/           sender identity, caps, country policy, niches
templates/playbook.md  the rules every draft must obey
data/             state, committed back by the workflow