jakub@quanetra:~/projects$ ssh pi@homestats
Raspberry Pi · Matter · SQLite
A collector polls your Matter sensors every minute and appends the readings to SQLite on the Raspberry Pi. The same computer serves ten views over your home network: from “what is happening now”, through traces, rhythms and relationships, to energy cost, events, a report and raw rows. No cloud, account or subscription.
A coloured text illustration: four measurement traces from the living room, an airing entry below them, and a cloud-request counter reading zero.
// features
The collector asks every Matter sensor once a minute and stores the reading with its device name, endpoint and unit. Names live in a table of their own, so renaming a sensor covers the whole history — it never splits into two lines on a chart.
Temperature and humidity give dew point, absolute humidity and apparent temperature — no new sensor, and for the whole history backwards. Socket power gives consumption, cost at your tariff, standby draw and the appliance’s duty cycle.
6 h, 24 h, 7 days, 30 days, or any interval. One control bar drives every data view: comparison with the previous period, a profile from the preceding seven days, aggregation and CSV export.
The interface is divided by question: Now, Traces, House, Rhythm, Comparisons, Energy, Events and Report. Every panel has a number and a place in the side rail, while the digest beneath the filters gives the key findings for the current page.
When the oldest sensor has been silent longer than the threshold set in Configuration, a red banner comes up over a dimmed page: it stops pretending to show the present. The same banner, in amber, reports a failed fetch.
Overnight cooling yields one number: how well the house holds heat. It is computed against a measured temperature difference once you mark which sensor stands outdoors — Matter does not report that, and the readings alone cannot settle it.
// screenshots
The screenshots come from a running server rather than a mockup. The demo database contains four devices, twenty-two series and enough history to show profiles, anomalies, seasons, energy cost and deliberate data gaps — without one reading from anybody’s home.
The expanded gallery now crosses twelve frames: the opening view, live state, device cards, five analytical pages, events, rhythm, the calendar and two responsive layouts.
// right_now
The dashboard describes a window, and the house is happening now — two different things, and the live section is for the second. Every series gets a sentence: what it reads, which way it is going, and how many minutes until it reaches the threshold at the current rate. A pulse says how many sensors are answering, and when they all go quiet at once it says plainly that this is the Raspberry Pi or the network, not the sensors.
// analysis
A line says what happened in sequence. These panels ask about time over a limit, shared movement, mould, the bill, airing and the physics of humid air.
How much time a value spent above a line — not how often it appeared. Those are two different things, and only the first answers “how many hours was CO₂ over a thousand this month”.
Every pair of measurements at once, sensor by sensor. This is where you see that bedroom humidity follows its temperature, and that the dust outside has little to do with either.
Humidity computed at the coldest surface in the house rather than in the middle of the room: mould grows from 80% up, which is long before the water that condensation reports appears on the wall.
Kilowatt-hours, cost at your tariff, standby draw and the appliance’s duty cycle — with the bill split into what went on merely sitting in the socket and what went on working.
A fast, sustained fall in CO₂ is marked as probable airing. The panel gives its duration, effectiveness and heat cost, while keeping the honest “probable”: a draught or everybody leaving can make the same trace.
Measurements on a psychrometric plot, constant-dew-point lines and the comfort rectangle. It shows whether the air actually gained water or merely became colder.
Beyond them: gradients between sensors, a measured air-change rate from CO₂ decay, day-by-day distributions, weekday box plots, an overlay of consecutive days, the rate at which the house warms up and the difference between two sensors over time.
// event_channel
Alongside deviation from the daily profile and a comfort threshold being crossed, there are: a lasting shift in level, calibration drift between two sensors, a combination of individually normal values, a weather front read from pressure, an open window inferred without a CO₂ sensor, and occupancy estimated from how fast CO₂ rises. Events from several sensors sharing a moment fold into an incident, and those still running at the edge of the window stand apart under “happening now”, with a minute counter instead of an end time.
// report
The report sums the window up in prose, under one rule with no exceptions: no sentence without a number. Beside it stands data quality — completeness, gaps, resolution and frozen values for every series in one table, worst first. Because that is what decides how much of the period being described was measured at all.
// beside_the_dashboard
Nine metrics on separate charts, with min–max bands, the previous period and a seven-day profile. Every chart can zoom and open full screen.
Panels derived from building physics: gradients, air exchange, mould, airing, the thermal time constant and a joint view of temperature and humidity.
An hour map, daily profile, weekdays, period overlay, day-by-day distributions and a yearly calendar separate what repeats from what happened once.
Seven panels from a table and histogram to correlation and the difference between two sensors. This is where hypotheses are tested, not merely where traces are viewed.
The flat’s climate, the silence-alarm threshold, where each sensor stands, per-sensor comfort thresholds and room volumes. All stored in the database on the Raspberry Pi, so the phone and the laptop read the same settings.
A table of raw readings straight from the database, with paging and sorting done by SQLite. Plus an export of the current window to CSV, one button in the bar above the dashboard.
One setting changes five things at once: a sensor marked as outdoor stops being judged by room thresholds, drops out of the “conditions normal” tile, stops standing in for a surface that condensation could form on, and stops filling the event channel with the weather — and in return it brings the building’s time constant and the hint about when to open a window.
// privacy
$ homestats --network-check
✓ accounts and signing in to a service 0
✓ readings sent to the cloud 0
✓ telemetry and analytics 0
✓ a subscription for your own data 0
status: everything stays on the Raspberry Pi
Readings go to homestats.sqlite in the project directory. A backup is a copy of one file, and retention for raw readings and for the hourly aggregate is set separately — an hour weighs sixty times less than a minute, so the history can stay for years after the minutes are gone.
Anyone who can see the server’s port on your network can see every reading. That is a deliberate decision for a flat, and equally the reason that port never gets exposed to the internet.
One script sets up the systemd units, builds the server and checks /api/health. If an update fails halfway it restores the services that were running before it started — the house is not left unmonitored because an install went wrong.
jakub@quanetra:~/projects/homestats$ status