mysmart.garden: An AI Gardener for People Who Kill Plants
I am a novice gardener. The kind who buys four pepper plants with real optimism and then stands in the yard a week later wondering if the wilted one is thirsty, drowning, sunburned, or just dramatic. Last year I killed a strawberry jar through what I can only describe as confident neglect.
So this spring I did what I always do when a problem annoys me enough: I built software for it. mysmart.garden is a total garden-management app for my whole property - the two raised beds, the tomato grow bags, the herb boxes, the strawberry jar that betrayed me, and eventually the flowers and bushes too. It started life as a static HTML watering tracker and a pile of phone reminders. It's now an AI companion that actually knows my garden.
The Idea
Most garden apps want you to fill out forms. Plant name, planted-on date, sun exposure, watering interval, dropdown, dropdown, dropdown. I have eleven-plus plants and the patience of a man holding a hose. Forms are where good intentions go to die.
So I flipped it. The primary interface is a conversation. I tell it "I put four mini bell peppers and four rainbow peppers in the long raised bed, and the end ones got frost-bitten" - in plain English - and an AI agent pulls that apart into actual structured data: plants, the bed they're in, their positions, even a note about the freeze damage. Forms still exist, but only as an escape hatch for when the AI gets something wrong. I describe; it files.
The mascot - a little sprout named Sprout - is the agent's face. It waves on the way in, looks concerned when there's a frost warning, and celebrates when I log a harvest. Seven expressions, no more. It's the whole emotional vocabulary of the app and I'm weirdly proud of how much restraint that took.
What It Actually Does
Morning briefing. Every morning I get an email that already knows the weather. "It's going to hit 91° today and you haven't watered the grow bags in two days - give the tomatoes a deep soak." It fires per-person at whatever hour each household member picks, so my wife and I can get ours at different times.
It watches the weather so I don't have to. The app pulls the forecast from the National Weather Service, cross-checks it against a second source (Open-Meteo), and reconciles the two - and if both disagree it tells me the forecast is uncertain instead of lying confidently. If it rains overnight, it auto-logs the watering for me, because the sky already did my job. And it watches for frost and extreme heat, sending an alert when peppers are about to get bitten - without spamming me the same warning twice.
Snap a photo, get an answer. I can take a picture of a sad-looking leaf right inside the chat, and the agent looks at it, tells me what it thinks is going on, and offers to file the photo against that specific plant so there's a visual history.
Adaptive watering. It doesn't just count calendar days since the last watering - it weights them by how hot it's been, so a stretch of 90° afternoons moves the needle faster than a mild week. The reminders bend to the actual weather.
The whole household - and the helpers. Data belongs to a household, not a single login. My wife sees the same garden I do. And when we go on vacation, my father-in-law gets a stripped-down email with a one-button "Done" link to confirm he watered - no app, no account, no password. He clicks one button and the log updates.
It also keeps a journal of everything that's happened, draws illustrated scenes of each bed, tracks harvests, suggests companion plantings, and remembers facts about my garden across conversations the way a good assistant would.
The Stack
The fun part. This one's a real application, not a weekend toy:
- Next.js 16 (App Router, Turbopack) + React 19 + TypeScript - the genuinely-new Next.js, which meant reading the actual docs instead of trusting muscle memory
- Supabase (Postgres + magic-link auth) - no passwords, row-level security everywhere, household-scoped data
- Vercel AI Gateway via the Vercel AI SDK - Haiku 4.5 for the fast stuff (chat, care tips, briefings), Sonnet 4.6 for the heavier reasoning (onboarding, scene generation, season planning)
- Vercel Pro for hosting and unlimited cron - five scheduled jobs run the briefings, the rain auto-logging, the weather alerts, and the helper reminders
- smtp.com + Nodemailer for the email that makes the whole thing feel alive
- Tailwind v4 + Radix for a mobile-first UI that's locked to a tiny six-color palette and exactly two fonts
The agent has a toolbox of functions it can call - create a bed, move a plant, log a watering, analyze a photo, plan a season - so the conversation isn't just chat, it's the steering wheel for the entire database.
Why I Built It
Partly because I wanted to learn the new Next.js and AI SDK on something I actually cared about. Partly because every existing garden app felt like a spreadsheet wearing a costume.
But mostly because I'm a novice, and being a novice is lonely. The plants don't tell you what's wrong. The internet gives you forty contradictory answers. What I wanted was something that knew my garden specifically - my beds, my zone, my west-facing fence that only gets a half-day of sun - and could just tell me, in the morning, what to do today. Not a database I have to maintain. A gardener I can text.
That's the whole bet: that the right interface for a garden isn't a form, it's a conversation with something that already knows the answer.
Where It's At
This is very much an active build. The foundation is in - onboarding, the weather pipeline, daily briefings, helpers, care tasks, photo analysis, and weather alerts are all shipped and running in production against my real garden this season. The illustrated scene generation works but is still rough, and I've got a list a mile long: hooking it up to the weather station on my own roof instead of a 2km grid cell, richer onboarding, and eventually the flowers and bushes that turn this from a veggie tracker into a true whole-property garden brain.
For now, it's getting me through pepper season. The frost-bitten ones, by the way, pulled through.