Awesome Grow Room Parts

This repo is free. My 2am dryback debugging is not. If it saved you a crop, a weekend, or a nervous breakdown — buy the rabbit a bag of nutes. If it didn't, keep your money. I respect a tight nutrient budget.

A parts catalogue for automated grow rooms, where the quality ratings are earned from evidence instead of asserted.

Most "awesome" hardware lists are a pile of links someone thought looked good. This one starts from a different question: which of these parts has actually survived in a production room, and how do we know?

The seed data comes from a real commercial indoor growing operation — 149 live ESPHome device configs across three flower/veg rooms, a drying room and back of house. A part tagged field-proven is named in a stated number of those live configs, and the filenames are recorded in the entry so you can check the claim. A part tagged avoid has a documented reason, usually "someone already tried this here and it did not work".

Everything beyond that seed is contributable. If you have run a part in anger, send it in.

→ Browse the catalogue site · searchable, filterable, with a copy-paste config example per part and an I²C address collision map.

→ Read the cost comparison

A sourced comparison of AROYA, Growlink, TrolMaster, Pulse Grow and Argus against a DIY build, with an interactive break-even calculator. Every figure carries the confidence level and the source it came from, and modelled numbers are not presented with the same weight as published ones.

The table groups by what each product actually does, because a monitor and a controller are not substitutes and ranking them together implies they are. Runs the room: TrolMaster ~$6.4k · Growlink ~$27k · Argus ~$51k at four rooms over five years (Argus's own published 9-zone cannabis example is ~$195k, 2018 dollars). Analytics rather than control: AROYA ~$76k at 8,000 sq ft canopy. Monitoring only, and not a substitute for a controller: Grow Sensor ~$3.8k · Pulse ~$12k.

Some of what the research turned up:

• AROYA's commercial subscription tiers are published in their ROI calculator's page JavaScript — $150–$3,900/month, billed per canopy square foot facility-wide. They corroborate three independent user-reported quotes, including one that an AROYA employee publicly disputed.
• "No subscription" is not the same as no recurring cost. TrolMaster's app is free, but their own API Gateway page charges $15/month per device for programmatic access to your own data.
• Only two of five vendors will say what happens to your hardware if you stop paying. TrolMaster and Pulse both answer it. AROYA's FAQ declines and routes to a sales demo. Growlink's Terms of Service contains no "Effect of Termination" clause at all.
• Monitoring is not control. Pulse Grow and Grow Sensor are monitoring and alerting platforms with no relay outputs and no equipment control — they cannot run a room, and the page groups them separately rather than listing their cost beside a controller's.
• Priva and iUNU could not be priced at all — not published anywhere, including through distributors. iUNU has also exited cannabis: their own contact form no longer lists it as a crop.

The page also reproduces, in full and unsoftened, the case against DIY — no support contract, no warranty, you are the integrator, calibration is your problem, no compliance audit trail, key-person risk. You should be able to read that section and reasonably decide to buy. If it were not there, the rest would not be worth trusting.

The tier is not an opinion field. CI enforces that anything claiming field-proven carries deployment_count ≥ 3 and the list of config filenames backing it, and that anything marked avoid states its failure_modes.

A five-star review tells you someone liked a part on the day it arrived. A deployment count tells you the part is still in a wall, in a hot humid room, eighteen months later, and nobody has ripped it out. Those are very different claims. This catalogue only makes the second one.

The single most annoying failure when building a sensor node: two parts that both want the same I²C address, discovered after everything is mounted.

The site's I²C page lists every part by address, flags every collision, and shows which parts can be strapped to an alternate. It also handles composite modules properly — an M5Stack ENV.III is two chips occupying 0x44 and 0x70, and that second address is what silently collides with a TCA9548A multiplexer's default. That is exactly the kind of thing you want to find on a web page rather than with a logic analyser.

Known collisions in the current catalogue: 0x28, 0x44, 0x51, 0x62, 0x69, 0x70, 0x76.

The catalogue is data first. One YAML file per part in parts/, validated in CI against schema/part.schema.json. The README tables, the website and the recipe totals are all generated from it — nothing is maintained by hand in two places.

parts/<id>.yaml        one part, schema-validated
recipes/<id>.yaml      a BOM referencing parts by id
schema/                the JSON Schema
assets/parts/          self-hosted thumbnails (never hotlinked)
docs/                  generated static site (GitHub Pages)
scripts/               validate · build site · generate README · check links

Vendor links rot, so every entry also records canonical part_numbers. When a shop URL dies, the part is still findable — and scripts/check_links.py only fails the build if a part has no reachable vendor and no part number to search on.

python scripts/validate.py        # schema + cross-refs + secret scan
python scripts/gen_thumbnails.py  # placeholder art for any part missing an image
python scripts/build_site.py      # -> docs/
python scripts/gen_readme.py      # -> the tables below
python scripts/check_links.py     # vendor/driver link rot

66 parts. Generated from parts/*.yaml by scripts/gen_readme.py — edit the YAML, not this table.

Prices: each part records one price - what was actually paid, in the currency it was paid in. Figures marked ~ are derived, not quoted, converted at 1 NZD = 0.5879 USD (rate date 2026-08-01; see data/fx.yaml). Rates move, so treat the band as the durable signal. Parts with no observed price show a band only. ^r = from a purchase receipt (what was actually paid); ^l = a vendor listing; unmarked = an estimate.

Parts named in the stated number of live production ESPHome device configs.

Documented dead ends. Listed so nobody repeats the work.

• Deployment counts were derived by scanning 149 ESPHome device configs for active (uncommented) component blocks. Archived and backup configs were excluded. Where a count looked ambiguous it is explained in the entry's evidence field.
• Prices are indicative, mostly NZD, with the observation date recorded. They will drift. Treat the band ($–$$$$) as the durable signal.
• Config examples are distilled from real production configs and have been scrubbed of credentials — see SECURITY.md. They use !secret placeholders; they are not drop-in copies of anyone's live system.
• This is one facility's experience. It is a strong signal, not a universal truth — a part that failed here may work fine in your conditions, and vice versa. Contributions that contradict the seed data with evidence are welcome.

Thumbnails are self-hosted, never hotlinked — a vendor reorganising their CDN must not break the catalogue. Each image is downloaded, normalised to a square white-backed PNG at two sizes, and committed.

Product images remain the property of their respective manufacturers and vendors, and are reproduced here for the sole purpose of identifying the part being catalogued. Every image records image_source and image_source_url in its part file, so the origin of any picture is traceable in one step.

Preference order when sourcing: the manufacturer's own product page or press kit, then a vendor listing, then nothing. Where no usable image is found the entry keeps a generated placeholder — a truthful placeholder beats a wrong or misleading picture.

Takedown: if you hold rights to an image here and would like it removed, open an issue on this repository, or contact the maintainer through the GitHub profile linked from it. Images are removed on request, no argument — the entry falls back to a generated placeholder and nothing else about it changes.

Catalogue data and documentation: CC BY 4.0. Scripts and site code: MIT.