Shipped 1.5.0 of my local-first Mac app last night and walked back a key decision 12 hours later

I spent the last 24 hours shipping bundled AI in Runwita, and then walking the decision back partially. The build-in-public version of the story:

Runwita is a Mac app for tracking long projects and customer relationships across many meetings. Local-first, SQLite on disk, no cloud. AI extraction handles the structuring (topics, actions, decisions) so users don't have to.

The friction problem: setting up an Anthropic API key. A non-technical user (the one I'm actually building for) bounces 70%+ of the time when asked to create a developer account.

So I shipped bundled AI for everyone. Free trial gets it cheap, Pro gets it generous, app routes through a Cloudflare Worker that proxies to Claude Haiku.

Twelve hours later I noticed two problems:

1. Cost math. $0.06/user-action uncached. A typical Pro user doing 5+ actions/day eats >50% of $12 revenue, with margin going negative for heavy users.

2. Privacy claim. Routing every Pro user's prompts through api.runwita.com quietly weakens the "your data stays local" promise that's the whole reason a security-conscious user pays.

So I pulled bundled AI back to Free-trial-only. Pro and Believer plans now BYOK. The acquisition story stays clean (free trial = no setup), the privacy story stays defensible (paid users' prompts never transit my infrastructure), and the cost math becomes "Anthropic spend is a tiny acquisition cost capped at the org level, not a recurring per-user expense."

The honest reflection: I was too attached to the "no API key, ever" pitch. Took a customer conversation and a 6 a.m. realisation to notice that two-tier (Free bundled, Pro BYOK) was a better answer than one tier.

Things I shipped along the way: a Cloudflare Worker with D1 for usage tracking, tier-aware gating on the proxy, a usage meter in the app's settings panel, full SEO Tier 1 work on the marketing site, first founder essay on the new blog at runwita.com/blog.

If you've shipped a paid tier with bundled AI: how do you handle the "non-technical user vs API-key wall" problem? I'd be especially interested in anyone who's tried both directions.

— B