ATX Layers — Austin, live from the city's open data

Where this comes from

Austin publishes a few thousand datasets at data.austintexas.gov. Most of them are spreadsheets nobody opens. A handful carry coordinates, and those are the interesting ones — they describe the city as a physical place, hour by hour. This page stitches a dozen of them — from both the Socrata portal and the city's GeoHub — onto one map and lets you cross-reference them against who actually lives where.

Live traffic incidents — dx9v-zd7x, refreshed by APD/ATD every few minutes. Pulsing pins are still open right now.
311 service requests — xwdj-i9he. Potholes, loose dogs, dead animals, noise complaints. These do carry coordinates, so the breakdown of what people are calling about can be scoped to a real bounding box or a radius around you.
Issued construction permits — 3syk-w9eu, with job valuation and contractor.
Crashes — y2wy-tgr5, the Vision Zero record. Runs a couple of weeks behind.
Live fire incidents — wpu4-x69d, the same CTECC dispatch feed as the traffic layer, both re-read every two minutes. Fire, rescue and hazmat; no medical calls, which the city withholds under HIPAA. Pulsing pins are still open right now.
Power outages — Austin Energy, through the KUBRA feed behind their public outage map. It reports by ZIP rather than by outage, so the shapes are the 51 ZIPs they serve and the counts are joined in the browser; only the ZIPs actually out are drawn. Customers out, customers served and a restoration estimate where there is one. Refreshed every five minutes against a source that republishes every ten. This is the one feed on the page with no archive behind it — the city publishes a map of what is true now and never writes it down — so a scheduled job in this repo has been recording it since 8 September 2026 to build the history that otherwise would not exist.
APD crime reports — fdj4-gpfu, 2.7 million of them going back to 2003. This one carries no coordinates at all — only a council district and a census block group, which is why it shades those units rather than dropping pins. All the counting happens on Socrata's side.
Traffic cameras — the city's live stills, which arrive at 1920 × 1080 — measured, because it decided the feature. Under about 640 px the right answer would have been showing two or three nearby cameras at once, since enlarging a small frame only enlarges the blur; at full HD the popup was simply showing one percent of the pixels the city was already sending, so tapping the thumbnail opens it near full width with a manual refresh. Nothing is stored. The dataset says plainly that these monitor live conditions and that no video is recorded or retained, and a still of a public street holds identifiable cars and people — an outage count is infrastructure state, this is not the same thing, and building the archive the city chose not to build is the one version of this page's habit that would create liability instead of value. Expect a share of any grid to be dark, frozen or pointed at a pole, and expect most of them after sunset to tell you only whether headlights are moving.
Fire stations, libraries, public art, traffic cameras — i8r8-6nhk, tc36-hn4j, uuk6-933w, b4k4-adkb.
Council district boundaries + demographics — w3v2-cj58 and puux-7swp, simplified and joined ahead of time so the shading is instant.
Census block groups, tree canopy and urban heat — the Tree Equity Score layer on Austin's GeoHub, the city's ArcGIS side, separate from the Socrata portal. 644 block groups, each carrying population, canopy cover and a summer surface-temperature reading against the city average. It's the only public source that puts all three in one place, and it's what lets this map shade 644 units instead of 10.
Parks, trails, floodplain, low-water crossings — also GeoHub, server-simplified into static files ahead of time.
Street condition — 28,006 segments, each graded A–F by Public Works off a measured roughness index. The drawable layer is the heavy one here, so it only loads if you ask for it; the per-area averages behind the two street metrics are joined to the block groups at build time and cost nothing to use.
Time — nothing extra is fetched to play a window back. Every point already carries its timestamp, so the scrubber is a filter. The layers don't share a clock, though: traffic is live to the minute while crashes stop wherever Vision Zero last published, so each layer sweeps the same relative position through its own range rather than pretending to one timeline. Anything without a clock — stations, libraries, art, cameras, and every ground shape — is standing infrastructure and stays put.
Live congestion — TomTom raster flow tiles, on when the page opens. They bill per tile and refetch on every pan, which runs to 185,000–212,000 requests a month against a 200,000 allowance — so they are the one layer here with a real marginal cost. The reason that is affordable is that the traffic archive no longer shares the risk: it rations itself against its own much smaller allowance, and the corridor board below is sampled once by a scheduled job rather than once per visitor. It sits deliberately under the APD/ATD dispatch pins, because traffic on its own is something every phone already has — the jam next to the wreck causing it is not. TomTom is the one vendor whose terms allow their traffic over someone else's basemap; Google's renders only on a Google map and their terms forbid it elsewhere, which would have meant discarding everything this page is built on.
Congestion as objects — the same TomTom congestion the flow tiles paint, from the Incident Details endpoint, drawn as lines you can tap and counted into a number. The tiles are pixels: you cannot count one, tap one, or ask it how many minutes of delay it represents. These are the same events the archiver has been recording for days. They also cost differently — tiles bill per tile and scale with map motion, which is how nine days came to 68,665 requests; this bills per request and scales with attention, one call per settled view on a snapped grid, cached for a minute. Off by default. Probe-derived, not reported: a jam here with no dispatch pin beside it means nobody called it in.
How fast is this road — TomTom Flow Segment Data, on a tap and never otherwise. Current speed against free-flow for the stretch of road under the incident you tapped, with the segment it measured highlighted so the number is not read as being about wherever your finger landed. Confidence is reported beside it, because it varies.
How far can I get — TomTom Reachable Range. Ten, fifteen and twenty minutes of driving from a point, as three nested polygons, in the traffic that exists right now. It is the only thing on this page that answers a question instead of describing a fact — everything else says what is; this says what you could do about it. Ask from the locate readout, or long-press anywhere on the map. Tap-only by design: three requests per ask is fine, three per pan would not be.
How the highways are moving — TomTom Flow Segment Data for nine named stretches: current speed against free-flow, as numbers. This is what the coloured flow tiles could never tell you — an overlay can say a road is red, it cannot say Mopac is doing 34 of 65, and you cannot read it without looking at the map. Sampled every twenty minutes by this site, not by your browser, so the reading costs the same whether one person is looking or a thousand — which is the only reason it can exist on a free allowance. Every coordinate was snapped to a motorway centreline and checked against OpenStreetMap: an earlier set chosen by eye landed on residential side streets, and the flow endpoint answers for whatever road is nearest without complaint, so it would have reported a cul-de-sac's speed under the name of a highway.
Rain — NOAA's base reflectivity ImageServer, which is time-enabled: each read asks for a narrow window at the end of its extent, because a request with no time on it returns the whole two-hour archive composited into one image, and three quarters of that rain has already fallen. None of the imagery is ever drawn. 565 m pixels are a fact about the instrument rather than about the weather, so the reflectivity is read as a density field and the map draws falling drops through it instead.
Cloud — GOES East infrared through Iowa State's mesonet WMS. Same treatment for a stronger reason: at 4 km a pixel is half of downtown, so it is drifting puffs sampled from the field, moving on the wind the Camp Mabry station actually reports. Infrared rather than visible on purpose — visible light is reflected sunlight, so that product is black for half of every day, and cloud at 2am is exactly when you would look. In infrared, brighter means colder means higher cloud top.
Satellite — Esri World Imagery, which over Austin is Nearmap orthophotography at about 8 cm. The panel reads the acquisition date out of Esri's own footprint layer for whatever you are looking at, so it says whose picture it is and when it was flown rather than just "satellite". It earns its place on detail rather than looks: the dark canvas has no tiles past zoom 16 and upscales a blurred one from there, so past 16 the imagery is the sharper map. It is filtered hard on the way in — most of the saturation removed, contrast up, and the same hue rotation that puts Esri's grey canvas into this page's navy — because raw aerial is the one basemap guaranteed to fight every colour on top of it. It rides the same day/night variables as everything else.
The light — computed, not fetched, and the only thing on this page that is not a feed. The sun's altitude and bearing come from the standard low-precision solar series for the map's own centre and the hour on display — no request, no library, nothing to be offline for. Checked against an independent almanac at both solstices and the equinox: within two and a half minutes on the sunrise instant, which for a colour ramp is nothing. It moves three things and deliberately not a fourth. The basemap stays dark always and only shifts along a warm/cool axis, deepest around 3am and warmest at noon — a light daytime map would break the contrast on every layer at once. The rain and the cloud take the sun's actual colour, gold when it is low and city-lit grey when it is down, dimmer at night but never invisible, because rain at 2am is still rain you came here to see. And a warm band leans in from the sun's own bearing while it is near the horizon — west at dusk, east at dawn. The hour it uses is the scrubbed hour, not the wall clock, so playing a window through walks the map across a day.
Buses and trains — CapMetro GTFS-realtime, about a hundred vehicles in service across thirty routes, reported within thirty seconds. This is the one feed the page cannot read directly: data.texas.gov answers with a redirect that carries no CORS header, and the file it points at rotates its name on every publish, so a small relay adds the missing header and changes nothing else. Vehicles in the yard are left off — where CapMetro parks is not the question. They move rather than blink: positions land about once a minute and each vehicle slides from where it was to where it is, on a canvas of its own so a hundred moving buses never repaint the sixteen thousand other marks.
When the next bus comes — CapMetro's trip updates and service alerts, through the same relay. Tap a stop for the next few arrivals by route, which stops are being skipped, and any detour or suspension affecting it. A caveat that matters: the feed publishes predicted arrival times and no delay field at all — checked across all 1,591 trips in a live pull — so this can say “next one in six minutes” and cannot say “four minutes late”. Saying the second would need CapMetro's static timetable to subtract from, which is megabytes. The alerts feed also carries the name and work email of the CapMetro staffer who filed each one; nothing here reads those fields.
Headlines — local newsroom RSS, read by a scheduled job in this repo rather than by your browser: the feeds do not send CORS headers, so the page could not fetch them directly even though they are public. The job reads them, keeps the item and drops everything else, and writes one small file to an orphan branch that the page loads cross-origin from raw.githubusercontent, which does send the header. An item only earns a dot on the map if its text names a place the gazetteer knows — a street, an intersection, a ZIP — so most headlines ride in the ticker and never land anywhere. A headline about the city is not a headline about a location.
Signals in fault — City of Austin traffic-signal status, the city's own record of which signals it cannot talk to or which are on flash. It is the nearest thing to a live map of where the grid is hurting, which is why it is worth having beside the outages — though the two do not always agree, and the City only operates City signals: TxDOT's, and other jurisdictions', are somebody else's dataset.
MetroBike docks — GBFS, the bikeshare interchange standard, straight from CapMetro's operator. Two files, one for where the docks are and one for what is in them, which is the whole reason the count can be live: the docks change yearly, the bikes change by the minute.
Wildfires — the national incident layer Esri hosts, filtered to Texas and then to what is near enough here to matter. Deliberately not clipped to the county: a fire 150 km upwind is the one you would want to know about, and it would never appear on a map of Austin.
Creek gauges — USGS instantaneous values, stage and discharge for every active gauge in the metro box, refreshed by them every fifteen minutes. No key, CORS open, and an endpoint that has outlived most of the web. It is the half the crossings layer is missing: crossings say what is closed, gauges say what is rising. An hour of readings is fetched rather than the latest one, because a stage on its own does not say whether the creek is coming up, and that is the part anyone acts on. Rising means a tenth of a foot an hour — measured, not guessed: across a dry hour the largest change on any of 37 gauges was 0.075 ft/hr, so a tenth is movement rather than the sensor breathing. And a stage is meaningless without a scale, so the National Weather Service flood thresholds are baked in beside it: 4.2 ft is alarming on Shoal Creek and nothing on the Colorado. Eleven of the 36 gauges NOAA lists here publish one, and they are the named creeks. Tap a closed crossing and it names the nearest gauge and what the water is doing — straight-line nearest, so it says which creek and lets you judge whether that is the one under the road.
School zones — three sources joined at build time: the city's zone register, the beacons themselves, and GeoHub centrelines. The register carries no geometry and the beacons carry the schedules, so the join is the feature: a beacon id reads “7322-529-”, and that 7322 is the zone, which is also GeoHub's. 161 of the 162 active zones come out with both a real window and a position; 27 have no published centreline and are drawn at their beacons instead, which the popup says. Only Active zones — the register also lists Inactive and Removed ones, and drawing a retired zone would be worse than drawing nothing. What it will not tell you is whether school is in session. The hours are published and exact, including the early-release window, so the page says a zone is scheduled when a weekday window is open and leaves the last step to you. No district here publishes its calendar in any form a page can read, and there are eight of them — Austin, Pflugerville, Round Rock, Leander, Eanes, Del Valle, Manor and a set of private schools — with different terms, different holidays, and one that runs year round. A school-zone layer that is confidently wrong on a teacher work day is worse than one that tells you what it actually knows.
Roadway work zones — open right-of-way permits rather than cones on the ground, which is why the count runs to thousands and some end dates run years out. Filtered to permits that have not expired, and drawn quietly except on the corridors the city itself flags as traffic-sensitive. Off by default: it is the heaviest ground layer and at city zoom it is the loudest thing on the map.
Routes and stops — CapMetro's static GTFS, which is a different feed from the live one: 34 MB zipped and 168 MB open, almost all of it in two files no browser should ever download. A weekly job streams them once — 3.3 million shape rows and 843,000 stop-time rows — and keeps only what is drawable. CapMetro publishes 7,323 shape variants for 71 routes, one for every detour and short-turn, so the most-used shape per route per direction is kept; that is what anyone means by “the 20” — unless that shape has a hole in it. Some do: 579 of the 7,323 skip more than a kilometre between consecutive points, and five routes were being drawn flying across open ground because of it. Those now fall back to the most-used sound variant, which in every case here is used by exactly as many trips. Route 491 has no sound variant in any direction, so it is drawn with a visible break rather than a confident straight line through country the bus does not cross. What is drawn stays within eight metres of the shape CapMetro published, and the build refuses to ship if it does not. Tapping a stop lights its routes and filters the live vehicles to them, because most stops serve several routes — Lavaca at 6th serves fifteen — and a stop therefore has colours rather than a colour.
Power history — outage state is not published as history anywhere, so a scheduled job here writes it down every quarter hour and the page reads that archive back: how long each ZIP was dark, and whether Austin Energy restored by the time it had published. Joined to census demographics by ZIP to ask whether the parts of town that stay dark longest are also the poorest — with the caveat, stated on the panel, that a few days of archive cannot answer that yet.
Current conditions — the Camp Mabry station (KATT) through api.weather.gov, which is the central Austin reading the city itself quotes, falling back to Austin-Bergstrom if it goes quiet. It rides in the news ticker every few headlines rather than in a bar of its own.
Next twelve hours, and air quality — Open-Meteo, free and keyless, and the one modelled thing on this page. Everything else here is an observation: a gauge read a stage, a station read a temperature, a bus reported where it was. A forecast is not, and neither is the air quality — the US AQI comes out of the CAMS model for this patch of the map rather than from a monitor standing in it. That is said plainly wherever it appears, because the claim this page makes is that it shows what is measured, and one row quietly not being that would cost the others their credibility. The hourly strip only appears when some hour reaches 20% — it exists to answer “which hour”, and twelve bars at the floor read as a broken chart rather than as a dry afternoon. The backlog suggested PurpleAir or AirNow; both now require an API key, and this page has nowhere to keep one a reader could not also read. Measured beats modelled, so if this earns its place the upgrade is AirNow behind the relay that already exists, where a key can live.
Weather alerts — api.weather.gov, the National Weather Service's own feed. Free, no key, and the only non-Austin source on the page. Every alert carries FIPS county codes whether or not it ships a polygon, which is what scopes it to Travis and its neighbors; warnings draw their real polygon, advisories that name only zones get the county outline instead, drawn dashed to mark the difference. Refreshed every three minutes. Most of the time there is nothing, and the section stays out of the way.
Your own location, if you tap the crosshair — read from the browser's geolocation API, used to work out which block group you're standing in, and never sent anywhere. There is no server here to send it to, and nothing is written to storage; close the tab and it's gone.

The per 10,000 residents shading is the part worth playing with. Raw counts just find the dense places. Normalizing by population surfaces where something is genuinely out of proportion to how many people live there — and at block-group resolution you watch it change street by street instead of across a tenth of the city.

Caveats worth knowing. 311 records reflect who calls 311, not where problems are. Crash data lags. Fire incidents are geocoded to the response area, not a doorstep. Demographics are the 2022 profile built on the 2020 census, so they age against everything else here. And per-capita rates break downtown: a block group with 200 residents and 40,000 daily workers will always look catastrophic per resident. The choropleth is shaded by rank rather than by raw value for exactly that reason — one runaway unit would otherwise render the whole city black — and the legend carries the real numbers.

One worth trying for yourself rather than taking on trust: shade the districts by percent of streets graded D or F and then by median household income. The pattern is messier than you would guess — downtown has the worst pavement in the city despite the money, because pavement follows traffic volume as much as it follows spending. Everything on this page is a link, so whatever you find you can send to someone.