Insights

Home → Insights is brAIn acting as your home’s analyst. Ask a question about your house and the answer comes back as a card: a one-line headline, the answer in a sentence, a handful of numbers, and one chart built for your data.

Ask, and get a card

Type a question into the bar at the top and press Ask. A minute or two later the answer is on a card.

Questions that make good cards

How quickly does the downstairs humidity drop when the AC turns on?
Why did the hallway light come on at 3 am?
Which rooms are coldest at night?
What's actually driving my standby power?
Which automations haven't triggered in 30 days?

Here is the card that the first of those produced on a real house:

Reading it top to bottom:

Part	What it tells you
The small heading	What the card is about — “HVAC · Humidity · Cooling”. For a card you asked, that’s its topics; for a recurring card, its name. Hover it to see the question you originally typed.
The title	The finding, in one line. It names what brAIn found rather than repeating your question.
The summary	The answer first, in bold (“Yes — the drying comes first and nearly free.”), then the reason with its numbers.
The tiles	The numbers the answer rests on, each with its unit and where it came from. They sit in even rows — four across, three for five or six, two on a phone.
The chart	One visualisation, drawn for this answer. Hover it for exact values. It sits straight on the card and stacks its panels on a phone.
The foot	When it was analysed, why it ran (you asked, on its refresh interval, a finding it reads changed), what it cost in tokens, and the run history.

On a phone the card rearranges itself: the title gets its own row under the buttons, and Refine and Share shrink to their icons.

The ask bar reads what you meant

Most things you type become a card, but brAIn reads the sentence first and sends it where it belongs — and one sentence can go two ways at once:

You type	What happens
A question — “why is the garage cold?”	A card with the answer.
learn about…, study…, figure out…	A study session. brAIn reads that corner of the house for a few minutes and files what it learns in Memory and Findings. No card.
A rule or a moment — “whenever the front door opens after dark…”, “when the guests leave, turn the porch light off”	A rule or a one-off, simulated against your history and offered on Proposals. Nothing runs until you accept it.
A fact — “the garage freezer is off in winter”	Kept in memory.
why… — “why did the porch light come on at 3 am?”, “why didn’t you tell me the garage was open?”	Answered right under the bar, in words, with what brAIn read to work it out — including what it decided not to say.
design my evening for the living room	Four scene moods for that room, offered on Proposals.

If brAIn can’t reach Claude at that moment, the bar falls back to deciding by the first words, the way it always did.

brain ask "why is the garage cold" in the terminal is the same engine, printed as text instead of a card.

Refine: change a card by saying what’s different

Every card has a ✎ Refine button. Type what should change and brAIn regenerates that card with the change, keeping everything you didn’t ask to change.

1. Press ✎ Refine on the card.

2. Say what should change. A few examples that work well:

   Compare it with upstairs
   Use the last 7 days, not one Tuesday
   Show costs in dollars
   Fewer numbers — just the three that matter
   Ignore the guest room sensor, it's in a drawer

   The chips under the box add common ones with a tap; you can combine several.

3. Leave Keep asking this on every future run ticked if the change should stick — the next scheduled refresh won’t quietly undo it. Untick it for a one-off look.

4. Press Regenerate with changes. The card shows its progress and redraws in a minute or two.

The version it replaced stays in the card’s history — page back with ‹ › in the foot. Everything you’ve asked a card to keep doing is listed at the bottom of the Refine dialog, each with a ✕ to stop it. Kept changes are also written into memory, so brAIn learns your preferences, not just this card’s.

Refine vs. Edit

Refine is for what the card says and is just a sentence. ⋯ → Edit on a recurring card changes its name, icon, schedule and the full analysis prompt — use it when you want to see and rewrite exactly what brAIn is asked each run.

Share: a picture, or a card on your dashboard

↗ Share takes a card somewhere else, two ways.

As a picture

The card’s face — heading, title, answer, tiles and chart, without any buttons or token counts — drawn as a PNG. This is what it looks like:

• Copy picture puts it on your clipboard, ready to paste into a message.
• Download saves it as a .png.

Some browsers — including Home Assistant’s own app on a phone — won’t let a page copy a picture. If Copy says so, press and hold (or right-click) the preview and copy it from there, or use Download.

On a dashboard

Pick a Dashboard, a View, and what to Show, then press Add card. brAIn adds the card to your dashboard itself — no YAML, no card editor.

Show	What appears on the dashboard
The whole card	Heading, title, answer, tiles and chart.
Just the chart	Only the visualisation, with the card’s title above it.

The dashboard card always shows the latest run of the card and updates whenever brAIn refreshes it. On a sections dashboard it arrives as a section of its own, so nothing you arranged moves.

Two kinds of dashboard can’t be written to, and the dialog says which:

• A dashboard kept in YAML — copy the YAML from the folded section underneath into it instead.
• A dashboard Home Assistant still builds automatically (the default Overview on a new install) — open it, choose ⋮ → Edit dashboard → Take control, then try again.

The YAML, for a YAML-mode dashboard

type: iframe
url: /local/brain/custom-1790000000-<your-card-token>.card.html
aspect_ratio: 110%

Cards are served by Home Assistant itself at /local/…, so they work on your LAN, over HTTPS and through Nabu Casa. The address contains a private token: anyone with the exact link can view that insight and nothing else. Dashboard cards made before 2.8 point at the chart-only file and keep working.

First use

If Home Assistant has never had a www folder, brAIn creates it — but Home Assistant only notices after one restart. The Share dialog tells you when that’s needed.

The card’s other buttons

Control	What it does
✎ Refine	Change the card by saying what should be different.
↗ Share	Copy or download it as a picture, or add it to a dashboard.
⤢ Expand	The chart at full size.
⋯ → Regenerate	Run the card again now. A card you asked for re-runs in place.
⋯ → Edit	Name, icon, analysis focus and schedule. On a card you asked for it’s Rename — the heading and icon.
⋯ → Edit tags	What the card can be filtered by in the chip row above the cards.
⋯ → Make this an automation	Puts the automation the card suggests — or just “When “ — into the ask bar for you to finish and send. It never sends from the menu.
⋯ → See the automation it suggested	Once a suggestion from this card is waiting on Proposals.
⋯ → Delete	The card and its history.

＋ Make recurring in the foot of a card you asked turns it into a scheduled card that refreshes on its own.

A card can suggest an automation

When a card points at an automation your house clearly lacks — “the hall light is left on most nights after everyone’s upstairs” — brAIn drafts it, replays it against your recorder and offers it on Proposals, exactly as if you had typed the sentence into the ask bar. Nothing is written until you accept it there, and the automation it makes carries a description saying an insight card suggested it.

Each suggestion stays listed on the card, saying whether it reached Proposals and why, so a suggestion that went nowhere doesn’t look the same as one waiting for you. Because this runs without you pressing anything, it follows the rules of every scheduled run — a sign-in, automatic insights on and budget left — and it’s limited to two suggestions per card and four a day. A card never re-offers what it suggested last time.

Where cards come from

A fresh install has no cards, deliberately — a generic card about a house brAIn has never looked at is noise on every run. Instead:

1. brAIn studies the home first — naming and areas, occupancy, energy, climate, device reliability — and files what it finds in memory. The screen lists what each study session learns as it lands (“The garage freezer is on a smart plug”), not just a progress bar.
2. It proposes cards grounded in what it found, each with the evidence that made it think so — and it waits until the study has actually found something before it does. You pick; nothing generates until you do. It may also suggest one automation to try, written as a sentence from something it saw you do; ticking it starts a week-long trial on Proposals and writes nothing to your house.
3. Later, Ideas keeps proposing. A house that grows a heat pump or a solar array gets suggestions for cards it’s now missing — costing nothing until you take one.

Deleted cards don’t come back on their own. brAIn proposes the cards your home should have, so the answer to “I want that card again” is to ask for it and let brAIn build it for the house it now knows.

How a card is made

For every card and every question, brAIn:

1. Looks up what the card needs rather than being handed the whole house. Claude gets a map of your home — how many entities of each kind, which areas they’re in — plus read-only Home Assistant tools, and fetches the rows, history and long-term statistics the question needs. If that search fails, it falls back to one slimmed snapshot of the house, so a card always appears — unless it failed because of your sign-in, a usage limit or an overloaded service, where the snapshot would fail the same way.
2. Is handed the memory that matters — the facts about the rooms and devices the card reads and its own kind of thing, with facts that share words with your question first, rather than the whole memory document. It can look memory up itself too. Free text from your house (media titles, calendar entries, notification bodies) reaches it marked as data, never as instructions.
3. Draws it with a fixed design system — one palette, one y-axis per chart, reserved status colours, light and dark mode, and a layout that works from a phone to a desktop.
4. Stores the result and shows the chart in a sandboxed frame that cannot touch your Home Assistant session.

Nothing on this path can change your house. An insight run holds reading tools only, and an automation a card suggests is only offered, on Proposals. The one path from a model to a change is a finding’s Fix it — a plan you read and an Apply a person presses. See Permissions & safety.

A card that declares live entities keeps those readings current while it’s on screen — the foot says “3 readings live · just now”, and the prose above it is labelled Analysed with its own age, because the words are from the last run and the numbers are from now.

Schedules, history and tags

• Per-card schedules — an interval, or fixed daily times (07:00, 19:00). Fixed times cost far less than a short interval. 0 means manual only.
• Run history — past runs are kept, with ‹ › in the foot to page back and prev: … under each number.
• Tags — the analyst tags each card by what it found, so #batteries surfaces every card that hit a battery problem. Edit them from ⋯ → Edit tags.

When a card actually re-runs

A schedule only says when a card may run. By default it also has to have something new to say: a scheduled card re-runs only when what it is actually told has changed — a finding it reads, the facts from memory it would be given, the state of a measurement it uses, feedback you added, a rewritten focus. A card regenerated over identical inputs would be the identical card at full price. Its own findings don’t count, so a card can’t trigger its own refresh, and a measurement that was merely rebuilt overnight with the same result doesn’t either.

The foot says why each run happened — on its refresh interval, a finding it reads changed — and a card that skipped says what it is waiting for, so a quiet dashboard and a stopped scheduler never look the same. When to refresh a card in ⚙ Settings picks the rule:

Setting	A scheduled card runs when…
When something it reads has changed	The interval has passed and something it reads has moved. The default.
Always, on the interval	The interval has passed.
Never on a schedule	Never on its own. Regenerate still works.

If brAIn can’t tell whether anything changed, the card refreshes anyway — “I could not tell” must never become a card that never updates again.

A card that fails waits longer each time — from 15 minutes, doubling up to 12 hours — and the next success clears it. If your Claude plan’s usage limit is hit, scheduled cards pause (30 minutes, doubling up to 4 hours) until the first run that succeeds, rather than failing every minute. Regenerate is never held.

What it costs

A card is the most expensive thing brAIn does — typically 25k–45k tokens — so every card says what it cost in its foot (hover for the in/out split and how long it took), and the usage pill’s popover itemises brAIn’s runs biggest first.

An asked question costs more than a recurring card. A recurring card reads its own slice of the home; a question could be about anything. Refine costs about the same as a regenerate.

⚙ Settings caps how much of each 5-hour session window brAIn may spend on scheduled work. Automatic runs pause at the budget; things you press never do. The meter uses your real Anthropic account usage, so brAIn backs off when you’re using Claude elsewhere. Every control: Reference.

Beyond cards: the weekly report and a deep review

Two things read the whole house rather than one question, and both are documented with their settings in Reference:

• The weekly report (Send a weekly report) ends on one thing to do this week. That pick is made across the week: brAIn reads the open problems alongside the week’s numbers and chooses the one whose dealing-with would make the most difference, with a sentence saying why — so three warnings about one flaky hub can outrank a serious row that is simply how your house is. It can only pick something it was shown, it never passes over something critical, and if it can’t run (no sign-in, automatic runs paused, the budget spent) the report falls back to the worst-severity, longest-open pick, exactly as before. Its counts match what the Findings tab shows.
• A deep review is a button on House → Knowledge. brAIn’s most capable model, reading only, sits with everything brAIn knows — the open problems, the measurements, the memory, the last weekly report — and writes up a handful of observations: patterns across rooms, a cause behind several symptoms, what is worth doing next. It files nothing; the review is a page you read (the last six are kept). It never runs on a schedule, and the button shows roughly what it will cost before you press it — from the reviews you’ve run before, or a first guess until then.

Insight jobs are the other half

Cards are interactive deep dives in the panel. Insight jobs are the dashboard-native sibling: scheduled reports rendered to sensors, with markdown in their attributes and an event on completion — good for glanceable tiles and spoken briefings. Like cards, they run with reading tools only, pause with Automatic insights and the usage budget when they run on their schedule (brain.run_insight always runs), and a run that fails shows as an error on the sensor beside the last good report — it is never sent to your phone as the report.