Five micro-SaaS products I built in 8 overnight passes

Eight products. Eight overnight build sessions, give or take. An AI agent wrote the code. I reviewed it in the morning.

This is the part where I'm supposed to either oversell the magic (AI makes everything instant and free) or undersell it with false modesty (of course I still did all the real work). Neither is accurate. Let me just describe what happened.

The products

Here's what exists as of today:

CertWatch — monitors your TLS certs and domains for expiry. Emails you 30, 14, 7, and 1 day before. No dashboard to open, no alerting system to configure.

OSS Pulse — watches your GitHub dependencies for health issues: archived repos, stale commits, CVEs from GitHub Security Advisories. Paste a package.json and you're set up in two clicks.

EverCV — a resume that updates itself. Connects to LinkedIn, pulls your latest experience and skills, regenerates a clean PDF. No more "last updated 2022" in your footer.

NightDesk — voice triage for MSP after-hours calls. Twilio + Claude handles the call, creates a CW ticket, and only pages the on-call engineer if the problem actually needs a human.

CW Slack Bridge — posts ConnectWise ticket events to Slack with configurable routing. New tickets → #helpdesk. Priority 1 escalations → #p1-alerts. MSPs already live in Slack; this brings CW along.

CW Copilot (TicketScope) — an AI assistant inside CW that drafts ticket responses, suggests runbook steps, and surfaces similar past tickets. A Chrome extension that adds a panel to the CW UI.

Build Your Own Cass — a course on building AI agents with the Anthropic API. Covers tool use, memory schemas, multi-turn conversation, and the design patterns I learned shipping these products.

CostWatch — monitors your AWS bill in real time. Shows month-to-date spend, projected end-of-month, and alerts when a service crosses a threshold you've set. The product I built because I needed it.

How the overnight build process works

I use an AI agent I've been calling Cass. The window is roughly 00:30 to 06:00 CDT — that's when I'm asleep and Cass is writing code.

I give Cass a spec before I go to sleep. Sometimes detailed, sometimes just a product description and a list of features. Cass works through the spec, writes tests, implements features, fixes the failures, and leaves me a morning summary.

I review the summary, read the diff, run the test suite, and either ship or leave notes for the next session. Most products took 2–3 overnight passes to reach code-complete. The test suite has grown to over 2,000 tests across all eight products. That's the safety net — without it, I'd have no confidence deploying code I didn't write line-by-line.

What this actually unlocks

Speed, but not infinite speed. Each overnight pass is 5–6 hours of unattended coding. A feature that would take me a focused weekend afternoon takes a single overnight pass. A product that would take 3–4 months of weekend coding takes 2–3 weeks of overnight sessions. That's real compression.

Code-complete products in weeks, not months. The backend, the billing integration, the test suite, the admin tooling — Cass builds all of it. I spend my mornings reviewing rather than implementing.

Fewer context-switching costs. I think about product during the day. Code gets written while I'm sleeping. The two activities don't compete.

What it doesn't unlock

Deploys happen when I'm awake. Cass can't push to production, set up DNS, or configure Stripe products. Those steps are mine. As of this writing, all 8 products have working backends but not all are publicly deployed.

Customer discovery is still manual. The code being done doesn't mean anyone knows about the product. Outreach, SEO, Show HN posts — none of that happens overnight. Code is not the bottleneck; distribution is.

The agent doesn't know if the code is good until I test it. Cass writes tests, but tests only catch what was anticipated. Integration problems, UX problems, "this is technically correct but nobody would actually want this" problems — those surface when I deploy and use the thing. The agent can't discover what it doesn't know to look for.

The unit economics

My costs are: Anthropic API ($60–80/month for overnight sessions), AWS ($30–40/month across all products), and 1–2 hours of my time per morning.

Break-even: one paying customer per product. CertWatch at $5/month pays for itself with one user. The math is straightforward.

This is the part of micro-SaaS that doesn't require an overnight AI agent — a boring CRUD app in a market with real pain can break even at 1 customer. The agent just means I can run 8 experiments in the time it would have taken to ship 1.

The honest part

I haven't talked to a paying customer for most of these products yet. They're built, they work, they have test coverage. But shipping is the easy part of building a SaaS business, and I knew that going in.

The bet I'm making: running 8 experiments at low marginal cost is better than going all-in on 1 idea with high confidence. Some of these will get traction. Some won't. The overnight build process means the cost of a miss is low — a few weeks of sessions, not a year of nights and weekends.

I'll write more about each product as they go live. If any of them look relevant to a problem you have, I'd like to hear about it.