TouchGrass: an offline garden planner powered by TabPFN (open-weight tabular AI)

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

TouchGrass is an offline garden planner. You tell it your USDA hardiness zone and today's date; it tells you which crops will thrive if you sow them this week — and then you close the laptop and go outside.

The entire decision core is TabPFN, Prior Labs' open-weight tabular foundation model, running on my laptop's CPU with the Wi-Fi off. No API keys, no cloud endpoint, no per-request cost.

🌱 TouchGrass — zone 5, week of 2026-04-20
   last spring frost Apr 15 · first fall frost Oct 15
   this week: 25.4 weeks before first fall frost

   ✅ spinach    thrive probability 84%
   ✅ radish     thrive probability 84%
   ✅ lettuce    thrive probability 83%
   ✅ zucchini   thrive probability 83%
   ✅ peas       thrive probability 83%

   Computed locally with TabPFN. Now go outside. 🍂

And in the fall, asked five weeks before first frost, it correctly gets cautious:

🌱 TouchGrass — zone 7, week of 2026-10-06
   this week: 5.0 weeks before first fall frost

   🤔 spinach    thrive probability 49%
   🤔 peas       thrive probability 49%
   ...

Who is it for? Anyone with a garden bed and a terminal — the goal was to make the screen the shortest part of the experience: one command, five crops, go plant.

Demo

uv sync
uv run touchgrass.py --zone 7 --date 2026-10-06

That's the whole UX. The heavier demo is turning off networking first and watching it still answer — that's the point.

Code

linhlban150612 / touchgrass-planner

Repo: https://github.com/linhlban150612/touchgrass-planner (MIT).

How I Built It

Three small pieces:

1. data/frost_dates.csv — median last-spring / first-fall frost dates per hardiness zone (zones 3–10).
2. data/plantings.csv — a training set of historical planting outcomes: crop, weeks from last spring frost, weeks before first fall frost, soil temperature, daylight hours → outcome class (thrived / stunted / failed).
3. TabPFN as the whole model. For each candidate crop I build one probe row for this week's frost window and ask TabPFN to classify the outcome. There is no training loop, no hyperparameter tuning, no gradient descent — a tabular foundation model doing few-shot classification over 55 rows and just working.

clf = TabPFNClassifier(device="cpu",
                       categorical_features_indices=[0])  # crop
clf.fit(X_train, y_train)
proba = clf.predict_proba(X_probe)  # one row per candidate crop

My first run had a fun bug: I forgot to include the crop as a feature, so every crop scored an identical 51% — TabPFN was correctly telling me all my probe rows were the same row. Adding the crop as a categorical feature gave it the signal it needed.

Why Does Open Innovation Matter?

• It runs on a laptop with no internet. TabPFN's weights are downloaded once and run locally. A gardener standing in a field with zero signal gets the exact same answer as one on fibre. A closed API returns a DNS error.
• It keeps your data off servers you don't control. Your zone, your dates, your planting history never leave the machine.
• It costs nothing to run. Every planting plan is $0.00. No key to rotate, no quota to watch, no endpoint to be deprecated under you.
• You can swap it. Because the model is a local artifact, swapping TabPFN for another open tabular model is a one-line change — the tool's value lives in the data and the idea, not in a vendor lock-in.

Where the open approach worked better than a closed one: latency and availability. Local inference returns in well under a second on CPU with zero network round-trips, and the failure mode of "airplane mode" simply doesn't exist. For a tool whose explicit goal is to get you away from connectivity, closed inference would be self-defeating.

My Agent Session

The whole project — code, data, debugging the identical-probabilities bug, git history — was built in a single agent session. The commit history in the repo (including the "crop as categorical feature" fix above) is the honest trace of that session.

Prize Categories

• Prior Labs — Best Use of TabPFN: TabPFN isn't bolted on; it is the product. A tabular foundation model forecasting planting success from historical data, fully local, inside a CLI.