Letting an AI Agent Pitch Me to Podcasts, Safely

Getting booked as a podcast guest means researching shows, finding the right contact and writing a pitch that proves you’ve actually listened. It’s hours of work per pitch. And unlike most automation, a mistake here is irreversible: an email goes to a real person.

The challenge

Build an agent that could run outreach end to end, without ever embarrassing me.

What I did

I built Megapitch, a web app and command-line tool that:

• searches Podcast Index and Listen Notes for relevant shows,
• finds the right contacts,
• drafts a tailored pitch with an LLM,
• and sends it from Gmail.

The interesting part is the guardrails:

• An LLM-as-judge check scores show fit, pitch quality and how tailored the pitch really is, and nothing sends without passing it.
• A daily send cap is enforced on the server, not just in the interface.
• Atomic locks make sure follow-ups never double-send.
• Tracking follows every pitch from opened to replied to booked.

I also wrote an AGENTS.md so other agents can run it with structured --json output, and deployed it on DigitalOcean.

What changed

Research and drafting that used to take hours per pitch now happens in the background, and I review outcomes instead of writing every email.

What I took away

The more irreversible the action, the more of your design effort belongs in the checks, not the clever part.