Custode

The wrong question

Every plant-ID app answers what, never whether

Point a phone at a leaf and a dozen free apps will tell you what it is and guess what's wrong with it. That part is solved, and it's free — it will stay free. It also isn't the question anyone who actually cares about a tree is asking.

One-shot diagnosis

Answerable from a single photo
Every competitor already does this
Tells you what it is, today
No reason to come back

Needs a registry, an anchor point, and history
Nobody else builds the boring part underneath it
Tells you whether it's changing
Worth more every month you keep it

The record

How it works

Register the tree once

A name, a spot on the map, a tag on the trunk. GPS alone isn't accurate enough at typical planting spacing — this is what keeps every later photo pointed at the right tree.

Same three shots, every visit — from any source

Canopy, foliage, trunk — from the same anchor point, with a ghost overlay of last time so framing stays consistent. Consistency of capture matters more than sophistication of model. The phone in your pocket is the default, but it isn't the only producer: a drone flyover and Custode's sibling grove-robot, Il Guardino, write into the exact same per-tree record — one registry, whichever of the three actually took the photo that visit.

A structured assessment, not a guess

A health score, specific findings with the evidence behind them, and a confidence level — never a bare verdict. A vision-capable Claude model reads all three shots together in one call (so a leaf symptom and what's underneath it get connected, not scored separately), forced through a fixed schema rather than free text — the model can't wander off into prose instead of an answer, and a malformed reply gets retried, not silently parsed-and-guessed.

It says when it doesn't know

Marginal photos and low-confidence findings get flagged for a human look, not quietly resolved. Triage, not diagnosis — a wrong "it's fine" costs you the tree, and that's the one mistake this is built to avoid.

Why the on-grove test comes first. Before this is anyone else's tool, it has to prove one thing on real trees: does the change detection actually catch something you'd have otherwise missed by eye? That's running now — 80 olive and fruit trees at Piccolo Paradiso, near Collie, Western Australia.

The interface

Every tree on the map, colour-coded by the last visit

Design mockup for the 80-tree grove — six olive rows, two fruit rows. Colour never carries meaning alone: every non-healthy pin also pairs a glyph and a word, so it isn't lost to colour blindness or a glance in bright sun.

Custode grove map screen: 80 trees laid out by row, colour-coded green/amber/red by health status, with a Needs Attention list below sorted worst-first. — Grove overview — 62 healthy, 12 to monitor, 5 needing action, 1 urgent.

Two real assessments from the mockup — an olive cultivar and a citrus one, because the findings that matter differ by species and Custode tracks the sub-type, not just "tree".

Tree detail for OL-R1-05, an Olea europaea Frantoio olive planted 2016, showing a declining health score and a water stress finding at 59% confidence. — Olea europaea — olive

Sub-type: Frantoio, planted 2016

Tree detail for FR-R1-03, an Imperial Mandarin citrus planted 2021, marked healthy with a low-severity leaf curl finding and confirm/not-right buttons. — Citrus reticulata — mandarin

Sub-type: Imperial, planted 2021

And the point of keeping the record: not one score, a trend — grove-wide, month over month, with the specific findings driving it.

Custode insights screen: 62 of 80 trees healthy, 18 needing attention (up 1 versus July), an 8-month line chart of trees needing attention by month, and a most-common-findings list led by black scale on 8 trees. — Eight months of visits, grove-wide — trending the wrong way in August, which is exactly the alert this is built to raise.