The Long Read:
the ten reviews.

A long read · Live Intelligence · Earth

A space-weather page that fetches three raw NOAA feeds every five minutes. We timed them on a quiet evening and set the page beside NOAA's own product and SpaceWeatherLive.

By the LABS editors Tested 27 September 2026, 19:58–20:01 UTC A 4-minute read 829 words

Plate
Aurora on labs.llc, captured from build 537 for this review.

Chapter one · What it is

A desk for people who watch the sky

Aurora is a single page on labs.llc for anyone who has stood in a cold field wondering whether tonight is the night. It puts three NOAA measurements on one screen: the planetary K-index as a dial ruled with the G1 to G5 storm bands; the OVATION aurora model painted cell by cell over a world map; and a GOES satellite's X-ray flux, sorted into flare classes.

It is for aurora chasers who want the agency's own numbers without an app, an account or adverts, and for anyone curious whether the field is quiet or storming.

Chapter two · How it works

No relay in the middle

Everything happens in your browser. The script asks services.swpc.noaa.gov directly for three JSON files — Kp by the minute, the OVATION grid and GOES X-rays — with caching off and a fifteen-second limit on each.¹ No labs.llc server sits in the path; NOAA lets any site read these feeds, as the open CORS header we saw confirmed.² It reads again every five minutes, and when a hidden tab returns to view.³

The OVATION file arrives as 65,160 separate cells of longitude, latitude and probability. Aurora packs them into one 360 by 181 grid, finds the strongest cell in each hemisphere, rings it on the canvas and names its coordinates beneath.⁴

Storm levels are named only once Kp reaches 5; below that, the index is shown exactly as filed.⁵ Every panel carries its own read time, and a failed fetch is announced as an alert rather than passed over in silence.⁶

Chapter three · The test

A quiet evening, measured

At 19:58:54 UTC on 27 September 2026 the page, served from a local copy of labs.llc build 537, returned 55,761 bytes. Fifteen seconds later we requested the three NOAA files the script asks for.

The Kp feed held 357 one-minute rows ending at 19:55 UTC, with an estimated Kp of 0. The OVATION file weighed 918,537 bytes and contained exactly 65,160 cells; the strongest northern cell was 12% at 69°N, the strongest southern one 13% at 62°S. The latest GOES-18 X-ray reading, at 19:56 UTC, was 4.28 × 10⁻⁷ W/m², which the page's own bands class as B.⁷

Nothing much was happening, and a page reading these files had to say so. Aurora's code has no way to round a zero up into excitement: Kp 0 is drawn as Kp 0, and a 12% peak is ringed and named as 12%.

Chapter four · Against the field

NOAA's own page, and SpaceWeatherLive

NOAA SWPC Aurora – 30-Minute Forecast

Page read 27 September 2026, 20:01 UTC

This is the forecaster's own product page for the OVATION model. It draws separate northern and southern hemisphere maps, keeps the last 24 hours of frames for animation, explains the OVATION 2020 model and its 30-to-90-minute lead time, and offers the grid as JSON alongside the Hemispheric Power Index.

Where it is better Polar maps suit an oval centred on the pole, the frames show the aurora moving, and the model is explained by those who run it.

Where Aurora goes further It puts Kp, the aurora grid and the X-ray flux on one page, each with a read time, where NOAA gives one product per page;⁸ and it names the strongest cell for the hemisphere you pick, then remembers the pick.

SpaceWeatherLive

Page read 27 September 2026, 20:01 UTC

SpaceWeatherLive credits NOAA SWPC for its predictions. Its aurora page offers a three-day Kp forecast, probability figures for high and middle latitudes, a 27-day outlook, instant alerts, mobile apps and an ad-free paid tier.

Where it is better It looks ahead, and it will alert you when conditions change.

Where Aurora goes further It reads the minute-by-minute Kp estimate and the live OVATION grid itself, redrawing every five minutes, rather than a forecast summary refreshed daily; and it has no adverts and no paid tier.⁹

Chapter five · Where it falls short

It cannot see tomorrow, or wake you

Aurora is a nowcast. It has no three-day or 27-day Kp outlook, although NOAA publishes both and SpaceWeatherLive displays them — the figure a trip north is planned on.

It has no alerts either: it reads only while the tab is open, so nothing will tell you that Kp crossed 5 while you slept.

Its map is a flat world canvas at the grid's own 360 by 181 resolution, not NOAA's polar projection, so the oval stretches near the pole, where the aurora lives.

Chapter six · The verdict

Honest numbers for the present tense

Aurora's achievement is restraint. It fetches NOAA's raw measurements from your own browser, shows all 65,160 OVATION cells at their native resolution, names the strongest, and stamps every panel with the moment it was read. On a dead-calm evening it said, in effect, dead calm.

Use it to see what the sun and the field are doing now. For anything further ahead, or for a nudge in the night, open NOAA's forecast page or SpaceWeatherLive.

Space weather in the present tense, with the working shown.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. aurora/assets/js/app.js lines 1–11 and 15; fetchJSON, lines 60–69.
2. The open CORS header on the Kp feed, at 19:59:09 UTC.
3. app.js line 16 (REFRESH_MS 300000) and line 471 (read again on tab focus).
4. app.js lines 224–238, 250–258 and 312–318; the pick is kept at 324–329 and 445–453.
5. app.js lines 104–116, with NOAA's storm names from line 20.
6. app.js lines 118–124 (read stamp) and 50–57 (errors marked role=alert).
7. Page at 19:58:54 UTC; feeds at 19:59:09 UTC. Flare bands: app.js lines 21–24.
8. app.js lines 1–11 and refresh(), lines 71–100; the pick, 324–329.
9. aurora/index.html line 550; app.js lines 15–16.

A long read · Live Intelligence · Earth

A board of every current US wildland fire, read straight from NIFC's incident layer. We made its exact query on 27 September and set it beside Watch Duty, NASA FIRMS and InciWeb.

By the LABS editors Tested 27 September 2026, 19:58–20:01 UTC A 4-minute read 856 words

Plate
Wildfire on labs.llc, captured from build 537 for this review.

Chapter one · What it is

The fire register as a readable board

Wildfire presents the National Interagency Fire Center's 'Incident Locations (Current)' layer as a board: every current US wildland fire and prescribed burn, with its name, state, county, acres, containment, cause and discovery date, ranked biggest first, plotted on a drawn map and counted by state.

It is for people following the fire season — reporters, residents of fire country, the curious — who want the official register in a form they can read, rather than a map of perimeters.

Chapter two · How it works

One request, sixteen fields

The script asks NIFC's public ArcGIS feature service for every current incident as GeoJSON, naming sixteen fields instead of requesting all of them; a comment notes that asking for everything would weigh 1.25 MB.¹ That single response feeds the list, the map and the count by state. The coastline under the map is Natural Earth's public-domain 1:110m outline, bundled with the page.² It reads again every five minutes while the tab is visible, and warns if the service capped the result.³

Its central rule is written into the file's header: if an incident has not reported its size or containment, the cell stays empty and the incident is left out of every sum and every rank — never shown as zero.⁴ It separates wildfires from prescribed burns and labels which is which, and it stamps each read with the time, the host it came from and the newest record change in the layer.⁵

Chapter three · The test

Four hundred incidents at 19:59

We loaded the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026 (56,001 bytes), then made the exact query the script builds at 19:59:38 UTC. NIFC answered in 0.31 seconds with 238,238 bytes: 400 incidents and no transfer-limit flag — 306 wildfires, 91 prescribed burns and 3 others, across 28 states.⁶

Seventy-six incidents had no reported size and 241 had no reported containment. That second figure is why the page's rule matters: a board that printed zeros would have shown 241 fires as 0% contained, a claim nobody had made.

The three largest were all in Oregon: the Rowe Creek Complex at 373,927 acres and 99% contained, 0445 Crosswhite at 342,923 acres and 99%, and the Hay Creek Complex at 200,292 acres and 100%.

Chapter four · Against the field

Watch Duty, NASA FIRMS and InciWeb

Watch Duty

Page read 27 September 2026, 20:01 UTC; rechecked 20:45 UTC

Watch Duty is a 501(c)(3) nonprofit's real-time wildfire service, now covering floods as well, with verified evacuation and shelter information, alerts gathered into one feed, an aircraft tracker for paying members, and apps for iOS, Android and the web.

Where it is better Verified evacuation orders and shelters, real-time alerts, fire growth followed on live satellite imagery, and first-hand photos and updates from people nearby. For anyone in a fire's path, this is the one to open.

Where Wildfire goes further It needs no app, account or membership, and gives the whole current register on one page, with a state count and a size ranking built from reported values only.⁷

NASA FIRMS

Page read 27 September 2026, 20:01 UTC

NASA's Fire Information for Resource Management System maps satellite hotspots from the MODIS and VIIRS instruments worldwide, available within three hours of observation and in real time for North America, with email alerts and SHP, KML and TXT downloads.

Where it is better Global reach, and it sees fires that have not yet been logged as incidents.

Where Wildfire goes further It names each fire and gives its acres, containment, cause and discovery date from the incident record; a hotspot carries none of those.

InciWeb

Page read 27 September 2026, 20:01 UTC

InciWeb is the official US government incident information system: a map and table of wildfires, prescribed fires, burned-area response and other incidents, with a red-flag-warning layer.

Where it is better Incident pages written by the teams working the fire, and red-flag warnings on the map.

Where Wildfire goes further It counts every current incident by state and ranks the whole layer by size in one view, reading again every five minutes.

Chapter five · Where it falls short

A register, not a warning system

Wildfire plots one point per incident and says plainly that the size shown is the latest reported, not a measured perimeter.⁸ It carries no evacuation orders, shelters, red-flag warnings or alerts — the first things a resident under threat needs.

Its list shows the 150 biggest incidents only. With 400 in the feed on the evening we tested, 250 were reachable only through the map and the count.⁹ And it covers the United States alone, with nothing like FIRMS's worldwide satellite detections.

Chapter six · The verdict

Discipline you can count

Wildfire's quiet virtue is refusing to invent. It asks NIFC for exactly what it needs, a fifth of the full payload, and wherever the record is silent it stays silent too. On the evening we tested, that meant declining to fabricate 241 containment figures.

Read it to understand the season as a whole. If a fire is near you, open Watch Duty first.

A fire register that would rather show an empty cell than an invented zero.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. wildfire/assets/js/app.js lines 18–27 (service address and the sixteen named fields).
2. app.js header lines 7–8; coast.js lines 3–5.
3. app.js lines 42 and 489 (five-minute refresh) and 94 (cap warning).
4. app.js header lines 10–14; rank() at lines 76–82.
5. app.js line 43 (wildfires, prescribed burns) and lines 97–100 (read stamp).
6. Page at 19:58:54 UTC; the query from app.js line 27 at 19:59:38 UTC.
7. app.js lines 1–8, 18–27 and 76–82.
8. wildfire/index.html line 711.
9. app.js line 42 (ROWS_MAX) and lines 189–205.

A long read · Live Intelligence · Earth

Seventy-five USGS gauges ranked by how far today's flow sits from each gauge's own week, with NOAA's flood stages read live. We tested the Colorado at Austin and set it beside the National Water Dashboard and NWPS.

By the LABS editors Tested 27 September 2026, 19:58–20:02 UTC A 4-minute read 815 words

Plate
Rivers on labs.llc, captured from build 537 for this review.

Chapter one · What it is

Is the river high?

Rivers draws live hydrographs for a fixed register of 75 USGS river gauges. Pick one and you get seven days of flow and stage drawn against the National Weather Service's flood stages for that point, with water temperature where the site reports it. A board ranks all 75 rivers by how far their current flow sits above or below their own seven-day median.

It is for paddlers, anglers, riverside residents and weather watchers who want to know whether a big American river is running high right now, without first learning the USGS site-number system.

Chapter two · How it works

Two agencies, asked sparingly

The page talks directly to two government services from your browser: USGS Water Services for instantaneous and daily values, and NOAA's National Water Prediction Service for each gauge's record.¹ It is frugal with both. The board fetches the whole register's daily values in a single call and current readings in batches of 25 sites; a chosen gauge gets seven days of discharge, stage and temperature, plus its NWPS record. Everything is read again every fifteen minutes.²

The ranking is the idea. Each gauge's current flow is compared with the median of its own last seven days, so a creek in spate and the Columbia at normal read on the same scale.³ Flood stages are never stored in the page. They are read live from NWPS for each gauge, so a threshold cannot go stale, and drawn on the hydrograph as the NWS ladder — action, minor, moderate, major — in the categories' own colours.⁴ The register itself was checked by hand: 72 of the 75 gauges are NWS forecast points.

Chapter three · The test

The Colorado at Austin

We loaded the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026 (54,901 bytes), then requested the seven-day readings the page builds for the Colorado River at Austin, USGS 08158000, at 19:59:20 UTC. The service answered in 0.71 seconds with 322,778 bytes and an open CORS header: 2,015 discharge readings, the latest 222 cubic feet per second and marked provisional, and 2,015 stage readings, the latest 12.97 feet. This site reports no water temperature.⁵

At 19:59:23 UTC the NWPS record for the same gauge set the flood categories at 25 feet for action, 33 for minor, 38 for moderate and 42 for major, with 13.01 feet observed and no flooding. The river stood roughly half-way to action stage.

Chapter four · Against the field

The National Water Dashboard and NWPS

USGS National Water Dashboard

Page read 27 September 2026, 20:02 UTC, via its drought.gov entry

The Dashboard maps provisional real-time stream, lake, reservoir, rainfall, water-quality and groundwater data from more than 13,500 USGS stations, alongside NOAA radar, watches, warnings and drought layers, and links to WaterAlert notifications.

Where it is better Every reporting station rather than 75, weather layers on the same map, and threshold alerts.

Where Rivers goes further It gives one ranked board of how unusual each river is against its own week, and draws the flood ladder from the live NWPS record directly on each hydrograph.⁶

NOAA National Water Prediction Service

Page read 27 September 2026, 20:02 UTC, via drought.gov and its API

NWPS is the National Weather Service's river service: observations and forecasts for each gauge with hydrographs, flood categories, upstream and downstream hydrographs in one click, National Water Model guidance and long-range flood outlooks.

Where it is better Forecasts — where the river is going, not only where it has been — and long-range outlooks.

Where Rivers goes further It joins NWPS thresholds to a week of USGS discharge, stage and temperature on one canvas, and adds the cross-river ranking.

Chapter five · Where it falls short

The forecast it already holds

The most striking gap is one the page has already paid for. The NWPS record it fetches for the Colorado at Austin carried a forecast — 16.7 feet for 1 October — but the code uses only the flood thresholds, and the word 'forecast' appears only in a label.⁷ NWPS draws that forecast; Rivers does not.

Seventy-five gauges is a small share of the thousands USGS runs, so if your river is not in the register there is nothing to pick. There are no alerts of the kind WaterAlert sends. And if the USGS endpoints go down there is no fallback source in the code.

Chapter six · The verdict

A fair scale for every river

Rivers' good idea is to compare each river with its own recent week; its good manners are to batch its requests and to read flood stages live rather than copy them. On the evening we tested, it would have shown the Colorado at Austin at 12.97 feet against a 25-foot action stage — calm, and correctly measured.

For the forecast, and for a message when the water rises, go to NOAA's NWPS and the USGS dashboard.

Every river judged against its own week, with the flood ladder read fresh each time.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. rivers/assets/js/app.js lines 15–17.
2. app.js header lines 8–11; lines 18, 129, 141, 248 and 261.
3. app.js lines 85, 129–134, 167 and 170–173.
4. rivers/assets/js/gauges.js header lines 1–8; app.js lines 20–25 and 261.
5. Page at 19:58:54 UTC; USGS instantaneous values for 08158000 (discharge, stage, temperature; seven days) at 19:59:20 UTC; NWPS gauge ACRT2 at 19:59:23 UTC.
6. app.js lines 167–173, 20–25 and 261.
7. The only 'forecast' in app.js is the label at line 307; the endpoints at lines 15–16 have no fallback.

A long read · Live Intelligence · News

A reading room over Chronicling America that turns one word into a map of the fifty states and DC and paces itself under the Library's rate limit. We timed a search and set it beside loc.gov and Newspapers.com.

By the LABS editors Tested 27 September 2026, 19:58–20:02 UTC A 4-minute read 764 words

Plate
Press Room on labs.llc, captured from build 537 for this review.

Chapter one · What it is

A reading room for old newsprint

Press Room is a reading room over Chronicling America, the Library of Congress's collection of digitised historic American newspapers. Search a word or phrase and you get newsprint cards that open the page scans; filter by state and by years; or 'ink' a map of the United States by how often a word turns up in each state, decade by decade.

It is for family historians, students, writers and anyone curious how a word travelled through American papers before 1964 — people who want the Library's pages without a subscription.

Chapter two · How it works

Fifty-one questions for one word

There is no server component. Your browser talks directly to the loc.gov JSON interface for the Chronicling America collection, twenty results a sheet, with optional dates.¹

The ink map is the ingenious part. One word becomes fifty-one faceted questions, one for each state and one for the District of Columbia, and each carries a dates facet, so the decade scrubber can redraw the map without asking again.²

Then comes the patience. The Library rate-limits, and its refusals arrive without CORS headers, so a browser cannot even read them properly. Press Room spaces the start of each request 1.4 seconds apart, pauses the whole queue for 65 seconds after a refusal, and after three pauses settles for a partial map rather than failing.³ It also reads the interface's arithmetic correctly — 'of' counts page images and 'total' counts result sheets — a reading the code confirmed by checking that facet counts sum to 'of'.⁴

Your last search is remembered, and inked words are cached in your browser for six hours, ten words at most, so a return visit does not spend fifty-one requests again.⁵

Chapter three · The test

Twenty-four seconds for one word

We loaded the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026 (50,226 bytes). At 19:59:49 UTC we made the request the page would make for 'lighthouse'. The Library answered after 24.7 seconds with 86,306 bytes and an open CORS header.

It reported 223,150 hits over 11,158 sheets and returned twenty results. The first card was image 3 of the San Antonio Daily Light for 1 October 1895, linked to the Library's IIIF image service.⁶

Chapter four · Against the field

The Library's own site, and Newspapers.com

Chronicling America on loc.gov

Page read 27 September 2026, 20:02 UTC, via the Library's research guide

The Library of Congress and the National Endowment for the Humanities publish millions of newspaper pages through 1963, from every state, the District of Columbia, Puerto Rico and the US Virgin Islands, with a viewer for zooming and navigating and downloads as TIFF, JPEG2000, PDF with hidden text and ALTO text.

Where it is better It is the authoritative viewer for every page, with archival downloads and documented bulk data.

Where Press Room goes further It turns a word into a state-by-state, decade-by-decade map — a question the collection search answers one query at a time — and paces itself so a visitor is not locked out.⁷

Newspapers.com

Page read 27 September 2026, 20:02 UTC

Ancestry's commercial archive counted more than 1.25 billion pages from over 31,500 newspapers, 1690 to the present, with person and topic search, a clipping tool, and subscription access after a one-week trial.

Where it is better A vastly larger archive, including everything after 1963, and clipping tools built for genealogy.

Where Press Room goes further It is free and needs no account; every query is your own browser asking the Library of Congress.

Chapter five · Where it falls short

The cost is patience

Press Room is slow because the Library is slow. Our single sheet took 24.7 seconds, and the code itself records anything from 3 to 32 seconds a query. A full ink map is fifty-one requests spaced 1.4 seconds apart, so at least about seventy seconds before any pause.⁸

It can search only what Chronicling America holds, which stops at 1963. And it offers no clipping, saving or citation tools beyond the link to each page.

Chapter six · The verdict

A courteous guest in a public archive

Press Room treats a public collection with care. It reads the interface's figures correctly, gets a whole map's worth of history from fifty-one requests, spaces them so the Library is not hammered, and remembers what it has already drawn. The ink map is the reason to come.

Bring patience. For papers after 1963 or for clipping tools, Newspapers.com goes further; for the definitive page viewer, go to the Library itself.

A careful reading room for America's old papers, if you can spare the wait.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. press/assets/js/app.js header lines 1–5; line 19; lines 78–85.
2. app.js lines 195–199 and 228–230; states.js carries 51 state keys.
3. app.js header lines 9–15; lines 22–24 (1,400 ms gap, 65,000 ms pause, three pauses); line 111.
4. app.js header lines 6–9.
5. app.js line 98; lines 26–27 and 202–225.
6. Page at 19:58:54 UTC; the loc.gov request for 'lighthouse' (JSON, twenty a sheet) at 19:59:49 UTC.
7. app.js lines 195–199, 228–240 and 22–24.
8. app.js lines 10 and 25.

A long read · Live Intelligence · Sound

A world tuner on the open radio-browser.info register that plays each broadcaster's own stream and explains a dead or blocked stream instead of going silent. We tested a jazz shelf and set it beside Radio Garden and TuneIn.

By the LABS editors Tested 27 September 2026, 19:58–20:03 UTC A 4-minute read 749 words

Plate
Radio on labs.llc, captured from build 537 for this review.

Chapter one · What it is

A dial for the world's stations

Radio is a world radio tuner built on radio-browser.info, the community-kept register of internet stations. It has a drawn dial with a sweeping needle, shelves by genre and by country, a search by name, and streams played straight from the broadcaster — no key and no account.

It is for listeners who want to find and play stations from anywhere in the world in a browser, without an app, a sign-up or adverts.

Chapter two · How it works

The register, the mirror and the stream

Everything happens in your browser. At load, the page asks the register for its list of mirror servers and picks one at random, falling back to a named mirror if that fails, to spread its load across the volunteers' machines.¹ Shelves are built from the register's tags and countries, ordered by how many stations each holds; lists come from its search, ordered by votes, with broken stations hidden.²

A station plays through one audio element pointed straight at the broadcaster's own stream — no relay and no re-encoding — and nothing loads or plays until you press play.³ After a successful play, the page counts the listen with the register, as the project asks its users to.⁴

Its most useful habit is candour. A dead stream is reported as dead. A stream offered only over plain http is not attempted on the secure page; the page explains instead that a secure page is not allowed to open it.⁵

Chapter three · The test

Nineteen of thirty

We loaded the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026 (53,163 bytes). At 20:00:16 UTC the register's server list held a single mirror, de1.api.radio-browser.info, so the random pick had one candidate. At 20:00:30 UTC its statistics counted 59,622 stations, 6,967 of them marked broken, with 12,372 tags, 241 countries and 681 languages.⁶

At 20:00:32 UTC we asked for the jazz shelf exactly as the page does: thirty stations, ordered by votes, broken ones hidden. Nineteen of the thirty had a stream address beginning http:, which the page declines to play on https. The top entries were Classic Vinyl HD and Adroit Jazz Underground, both MP3 at 320 kbps, and 101 Smooth Jazz at 192.

Chapter four · Against the field

Radio Garden and TuneIn

Radio Garden

Page read 27 September 2026, 20:03 UTC, with its App Store listing

Radio Garden finds live radio by rotating a globe: each point is a place, and choosing it offers that place's stations. Its App Store listing adds search for stations, places and countries, favourites synced through iCloud, AirPlay, and free use with in-app purchases.

Where it is better A geographic way in, and apps with synced favourites and AirPlay.

Where Radio goes further It browses by genre and by country from the register's own counts, tells you plainly why a station will not play, and reads an open, keyless register straight from your browser.⁷

TuneIn

Page read 27 September 2026, 20:03 UTC

TuneIn is a streaming audio service for live sport, music, news, audiobooks and podcasts, available on hundreds of home, car and portable devices and working with Alexa and Google Home, with sign-in and a premium tier.

Where it is better Sport, news, talk and podcasts in one place, on smart speakers and in cars.

Where Radio goes further It needs no account, carries no adverts, and says in its status line that the stream is the broadcaster's own.⁸

Chapter five · Where it falls short

The limits of an open register

Many stations cannot play on a secure page — on the jazz shelf, nineteen of the top thirty — and all the page can do is explain why. It also depends on a volunteer register, which at 20:00 UTC had one mirror and itself marked 6,967 of 59,622 stations broken.

There is no favourites sync, no app, no casting and no sleep timer, and there are no podcasts, sport or talk catalogues of the kind TuneIn carries.

Chapter six · The verdict

Candour as a feature

Radio's lengths are honest ones. It reads an open register directly, shares the load across the project's mirrors, plays nothing until asked, plays the broadcaster's own stream, and when a station fails it tells you why — even when the answer is an awkward rule of browser security that rules out most of a shelf.

For a globe to spin, synced favourites, or sport and talk, Radio Garden and TuneIn go further.

A tuner that would rather explain a silence than hide it.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. radio/assets/js/app.js lines 13–14 and 59–68.
2. app.js lines 468 and 480 (shelves) and 77–81 (search by votes, broken hidden).
3. app.js header lines 3–8; line 19 (no preload); lines 390–392.
4. app.js line 407.
5. app.js header lines 7–10; play() lines 383–387.
6. Page at 19:58:54 UTC; server list at 20:00:16 UTC; register statistics at 20:00:30 UTC; jazz shelf at 20:00:32 UTC.
7. app.js lines 13–14, 468, 480 and 383–387.
8. app.js lines 402–405.

A long read · Studio & Tools · Sound

A microphone visualiser with thirty-three displays, a recorder that tests its formats first, and no network code at all. We read its 5,206 lines and set it beside audioMotion and Friture.

By the LABS editors Tested 27 September 2026, 19:58–20:05 UTC A 4-minute read 842 words

Plate
EQ on labs.llc, captured from build 537 for this review.

Chapter one · What it is

Seeing sound in a tab

EQ is a live equaliser driven by your microphone. Allow the microphone and the sound in the room is drawn in one of thirty-three ways, from Classic Bars and a scrolling spectrogram to VU needles, Note Warp, Truchet Tiles and Rose Pens.¹ You can record the picture with its sound, or the sound alone, straight to your own disk, or keep a single frame. There is no file to load and no player to learn.

It suits musicians, streamers, teachers and the curious who want to see sound — or make a visualiser clip — without installing anything, and who care where their microphone goes.

Chapter two · How it works

Raw input, and no way out

EQ is built on the Web Audio API and runs entirely in the browser. When it asks for the microphone it switches off echo cancellation, noise suppression and automatic gain — the processing browsers apply for calls — and falls back to a plain request only if that one fails.² The picture therefore shows the room rather than a telephone filter.

The signal goes to an analyser with an FFT size of 512, and every display draws from the same frequency and waveform arrays. The analyser is never connected to the speakers, so there is no feedback loop.³

The script's header states that the file makes no fetch, XHR, beacon or worker call, and a search of the file bears that out. The page's privacy section is careful to separate what the instrument promises from what the surrounding page loads, such as analytics and a font.⁴

Recording uses MediaRecorder on a stream taken from the canvas. At load, each container is put to MediaRecorder.isTypeSupported(), only the formats that pass are offered, and each option notes where the file will open — MP4 on iPhone and Windows, WebM not on Apple devices. A separate sound-only menu offers m4a or WebM.⁵

Chapter three · The test

Reading the instrument

We loaded the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026 (41,577 bytes), then the instrument script: 251,224 bytes over 5,206 lines. We counted 33 display names and found no occurrence of fetch(, XMLHttpRequest, sendBeacon or new Worker.⁶

We also worked out its resolution from the source. An FFT of 512 gives 256 frequency bins; on a device running at 48 kHz each bin is 93.75 Hz wide, so the lowest spans everything from 0 to about 94 Hz. Values are 8-bit, from 0 to 255. That is plenty for a lively picture, and coarse for measurement.

Chapter four · Against the field

audioMotion and Friture

audioMotion

Page read 27 September 2026, 20:03 UTC, via its project README

audioMotion is a media player and high-resolution spectrum analyser that runs as a browser app, taking local files, M3U playlists or the microphone. Its project README lists logarithmic, linear, Bark and Mel scales, up to 240 bands including octave bands, decibel or linear amplitude, and A, B, C, D and ITU-R 468 weighting; it is AGPL-licensed and asks for a Chromium-based browser.

Where it is better It plays and analyses your music files, not only the microphone, and offers measurement-grade scales, bands and weighting. It is open source.

Where EQ goes further It records picture and sound to a format your browser has proved it can write, and says where the file will open; and it is written for any browser with Web Audio, with a fallback for older Safari, rather than for Chromium alone.⁷

Friture

Page read 27 September 2026, 20:03 UTC

Friture is a free, open-source real-time audio analyser for Windows, macOS and Linux, with a spectrum analyser, a rolling spectrogram, a fraction-of-octave analyser, a scope, a pitch tracker, a delay estimator and a signal generator.

Where it is better These are real analysis tools, and the test-signal generator is something EQ does not have.

Where EQ goes further It asks you to install nothing, offers thirty-three designed displays, and records to MP4 or WebM for sharing.

Chapter five · Where it falls short

A picture, not a meter

EQ listens only to the microphone. There is no file input anywhere on the page, so you cannot visualise a song file directly, which audioMotion does.

It is not a measuring instrument: 256 bins of roughly 94 Hz, 8-bit values, no calibrated decibel axis, no weighting filters and no pitch tracker. For measurement, audioMotion and Friture are the tools.

There is also a small gap between word and code. The file's header says the only thing it stores is your palette name, but it also stores the recording mode under eq.recMode.⁸

Chapter six · The verdict

Careful with the one thing it asks for

EQ spends its care on the microphone. It asks for the unprocessed signal, never plays it back, contains no network code, and offers only the recording formats it has tested. Its thirty-three displays make that signal a pleasure to watch.

For seeing sound, and for a clip that stays yours, it is a fine instrument. For measuring sound, go to Friture or audioMotion.

An instrument for watching sound that never lets the sound leave.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. eq/assets/js/eq.js display names, lines 656–2804; eq/index.html line 406.
2. eq.js lines 3325–3331; fallback at line 3333.
3. eq.js lines 3340–3347; header lines 16–19.
4. eq.js header lines 8–13; the search for network calls found nothing.
5. eq.js lines 3755–3800 (candidates and notes), 3857 and 3883–3893 (sound only).
6. Page and script at 19:58:54 UTC; resolution worked out at 20:05 UTC from line 3341.
7. eq.js lines 3313–3336 (feature check, constraints, webkitAudioContext fallback).
8. eq.js header lines 19–20, against line 4793.

A long read · Studio & Tools · Sound

A sixteen-step drum machine with eight synthesised voices, a clock built for phones, and a whole groove in a 62-byte link. We read its three scripts and set it beside Drumbit and Splice Beatmaker.

By the LABS editors Tested 27 September 2026, 19:56–20:03 UTC A 4-minute read 818 words

Plate
Drums on labs.llc, captured from build 537 for this review.

Chapter one · What it is

A drum machine with its recipes showing

Drums is a sixteen-step drum machine whose eight voices are synthesised as they play. It has swing, tap tempo, eight pattern slots, a recorder, and a link that carries the entire pattern to whoever you send it to.

It is for beginners learning how a beat is put together, producers sketching an idea, and teachers — anyone who wants a drum machine in a tab with nothing to download and no account to make.

Chapter two · How it works

Oscillators, noise, and a clock that looks ahead

There are no samples. kit.js describes itself as having no sample, no fetch and no URL, and builds each voice from oscillators and noise shaped by envelopes, routed through voice buses to a master and out, with the recorder tapped off the master.¹ Each voice's recipe is written beside it: the kick is a sine swept from about 160 Hz down to 48 Hz in fifty milliseconds, with a noise transient; the clap is four bursts of noise near 1.1 kHz, about eleven milliseconds apart.²

Timing belongs to clock.js, a lookahead scheduler that writes each step into the audio clock at a sample-accurate time before it is due. Its pump runs twice — from the page and from a two-line Worker — because phones throttle idle page timers and leave a Worker's alone.³ Swing is defined exactly: every odd sixteenth is played late by the swing fraction of one step, with no running total, so the bar never stretches. Tap tempo averages the last eight taps and forgets any gap longer than two seconds.⁴

machine.js packs the whole machine — tempo, swing, length, kit, every cell and accent, each voice's level, tune and decay, and the mutes — into 62 bytes of base64url. The same encoding fills the slots and the share link, which carries it after #p= in the fragment, the part of an address browsers do not send to servers.⁵ The recorder, at 192 kbps, reads back the format the browser actually produced instead of assuming it.⁶

Chapter three · The test

Weighing the machine

At 19:56 UTC on 27 September 2026 we sized the three scripts in a local copy of build 537: clock.js 7,005 bytes, kit.js 25,946 and machine.js 35,397. At 19:58:54 UTC the page itself returned HTTP 200 and 66,235 bytes.

That is about 68 KB of code and not a single audio file — every sound on the page comes from those bytes.

Chapter four · Against the field

Drumbit and Splice Beatmaker

Drumbit

Page read 27 September 2026, 20:03 UTC

Drumbit is a free web drum sequencer with preset kits from Rock and Funk to Bossa Nova, Hip-Hop and House, a sixteen-step editor across four patterns, tempo and swing, a compressor, low- and high-pass filters, a keyboard entry mode, recording, and saving and loading as a drumbit file. It is funded by voluntary contributions.

Where it is better Four patterns rather than one, a compressor and filters on the output, and genre kits ready to play.

Where Drums goes further It carries its whole state in a 62-byte link fragment, so there is nothing to save as a file and nothing sent to a server, and it explains each voice as synthesis in plain words beside the grid.⁷

Splice Beatmaker

Page read 27 September 2026, 20:03 UTC

Splice's free browser beatmaker offers a 32-step grid of real drum samples, any tempo from 60 to 240, featured beats by named producers, sharing by link and a free MIDI download of the pattern. Saving and loading beats needs a Splice account, and more sounds need a subscription.

Where it is better Professionally recorded samples, twice the steps, and MIDI export.

Where Drums goes further It keeps your work without an account, in eight local slots and a self-contained link, and its Worker-driven second pump is engineered for phones.⁸

Chapter five · Where it falls short

One pattern, eight voices, no MIDI

Drums plays a single pattern of up to sixteen steps. There is no song arrangement, no chaining of patterns, no per-step probability and no MIDI — and the page says so itself.⁹

It is not a sampler: eight synthesised voices are all you get, where Splice offers producer kits of real recordings. And latency is whatever your browser and device provide; on a phone, the page admits, you can feel it when tapping along.

Chapter six · The verdict

Small, exact and entirely yours

Drums goes to the length of building every sound from arithmetic and publishing the recipe, of engineering its clock around the habits of phone browsers, of defining swing to the step, and of fitting a whole groove into a link the server never sees. It is a teaching instrument as much as a toy, and a good one.

When you need real samples, longer patterns or MIDI, Splice Beatmaker and Drumbit go further.

A beat machine that shows you how every sound is made, then lets you keep it.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. drums/js/kit.js lines 1–6 and 8–22.
2. kit.js lines 42–58 (the voice table).
3. drums/js/clock.js lines 1–18; the Worker at line 52.
4. clock.js lines 60–65 (swing) and 136–157 (tap tempo).
5. drums/js/machine.js lines 40–62 and 499–518; drums/index.html lines 915–920.
6. machine.js lines 567–572 and 595–598.
7. machine.js lines 40–62 and 499–518; kit.js lines 42–58.
8. machine.js line 15 (slots); clock.js lines 9–18 and 52.
9. drums/index.html lines 903–913.

A long read · Studio & Tools · Design

A generative art lab with forty-nine composition algorithms, a seed that replays any piece, and a recorder that asks for 60 Mbps. We counted its tables and set it beside Silk and Hydra.

By the LABS editors Tested 27 September 2026, 19:58–20:06 UTC A 4-minute read 862 words

Plate
DMT on labs.llc, captured from build 537 for this review.

Chapter one · What it is

Pictures you can steer and keep

DMT is a procedural geometry lab on labs.llc. Shapes are arranged on a canvas by one of forty-nine composition algorithms; you can let it run, steer it, freeze a moment, share a preset as a link, or record the result as video. Some modes react to a microphone.

It is aimed at visual artists, VJs and designers — anyone who wants generative art that moves, that can be replayed exactly, and that needs no code.

Chapter two · How it works

One page, one seeded generator

The whole engine lives inside one page, drawn to canvas in the browser. Nothing is fetched for the art; the only network request in the page belongs to its contact form.¹

Its most careful part is its randomness. A seed is hashed once into 128 bits and fed to an sfc32 generator, with independent salted sub-streams for different decisions, and picks are made with a seeded Fisher–Yates shuffle.² The comments explain why: the old habit of adding a counter to the seed made neighbouring seeds into shifted copies of each other. Now a seed replays a composition exactly and the next seed along is genuinely different.

Above that sits what the page calls a beauty engine, with three strategies. HARMONIA chooses colour relationships first — triads, split complements, analogous sets. GOLDEN walks an eight-dimensional composition space by a low-discrepancy sequence, so consecutive pieces do not clump. EVO keeps a population of fourteen compositions and breeds them by tournament selection, crossover and decaying mutation, steered by the pieces you keep.³

A preset is the full state as base64 JSON in the ?p= query, loaded through the same path as the undo history.⁴ The recorder captures the canvas at 60 frames a second and asks MediaRecorder for 60,000,000 bits a second, preferring WebM VP9 and then MP4, and names the file after the container it actually received.⁵

Chapter three · The test

Counting the tables

We requested the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026: HTTP 200 and 367,539 bytes of HTML, with the whole engine inline. At 20:06 UTC we counted its tables in the served source.⁶

The algorithm table holds 49 named entries, matching the '49 unique algos' in the page's own description. The shape table holds 21, where the description says 20; two of the 21 are slots for your own images, so the difference is one of counting rather than substance.

We did not measure the bitrate of a finished recording, and that leads to the first caveat in chapter five.

Chapter four · Against the field

Silk and Hydra

Silk

Page read 27 September 2026, 20:02 UTC

Silk, by Yuri Vishnevsky, turns a drag of the pointer into flowing strands, with mirror and spiral symmetry, undo, a choice of colour, and music and sound by Mat Jarvis. You keep a piece by saving its thumbnail image, and art shared through Silk carries a Creative Commons licence.

Where it is better It is instant and lovely with almost nothing to learn, and the art follows your hand.

Where DMT goes further Its forty-nine algorithms and seeded replay mean a piece can be regenerated exactly from a number, and it records moving video at a requested 60 Mbps where Silk saves a still image.⁷

Hydra

Page read 27 September 2026, 20:02 UTC, via its documentation

Hydra is a free, open-source video synth and live-coding environment that runs in the browser and compiles to WebGL. Its documentation describes chaining transformations like an analogue synthesiser, mixing camera, screen-share and video sources, audio-reactive visuals, streaming between browsers, and a community gallery.

Where it is better Code gives Hydra unlimited range; it renders on the GPU, takes live video inputs, and is open source.

Where DMT goes further It needs no code at all — the beauty engine chooses palettes and structures, and in EVO mode learns from what you keep — and it ships a high-bitrate recorder and one-link presets.⁸

Chapter five · Where it falls short

A request is not a guarantee

The 60 Mbps figure is what DMT asks the browser's encoder for; the code never reads back the bitrate it got, so the number describes intent rather than output.

Because the recorder prefers VP9, a recording made in Chrome is a WebM file, which Apple devices will not open without conversion; MP4 is used only when VP9 is unavailable.⁹

Browsers that cannot record get a blocking alert box rather than a message in the page. And DMT is not programmable: there is no code layer, no camera or video input to mix, and no streaming between browsers.

Chapter six · The verdict

An instrument for repeatable accidents

DMT goes to real lengths to make randomness trustworthy: a rebuilt generator so that seeds replay and neighbours differ, three stated strategies for choosing what looks good, and a recorder that asks the browser for 60 Mbps. For no-code generative video that you can reproduce and share, it is unusually thorough.

If you want to draw by hand, Silk is simpler and lovelier. If you want to write the visuals yourself, with cameras and shaders, Hydra is the deeper instrument.

Generative art you can reproduce from a number, no code required.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. dmt.php line 6741 (the contact form).
2. dmt.php lines 1366–1430 (seed hash and sfc32) and 1434–1445 (seeded shuffle).
3. dmt.php lines 5081–5088, 5137–5143 and 5173–5182; modes at line 5554.
4. dmt.php lines 2713, 2717–2720 and 2970–2979.
5. dmt.php lines 3265–3275 and 3283–3288.
6. Tables at lines 1488–1510 (shapes, 21) and 1512–1562 (algorithms, 49); description at line 186.
7. dmt.php lines 1512–1562, 1366–1430 and 3265–3275.
8. dmt.php lines 5081–5182, 2717–2720 and 3265–3275.
9. dmt.php lines 3268–3269; the alert at line 3261.

A long read · Consumer Brands · Cities

A register of 1,085 celebrated restaurants in 218 cities that deleted every field it could not stand behind. We counted it, and set it beside the Michelin Guide and the World's 50 Best.

By the LABS editors Tested 27 September 2026, 19:58–20:02 UTC A 4-minute read 800 words

Plate
Yum on labs.llc, captured from build 537 for this review.

Chapter one · What it is

A shortlist of celebrated rooms

Yum is a guide to notable restaurants around the world: 1,085 of them in 218 cities and 51 countries, drawn from the consensus of the Michelin Guide, Eater and the World's 50 Best. You can search it by name, city, cuisine and Michelin stars, have 'tables for two' dealt at random, and read best-of lists the page works out for itself. A separate register of 518 kosher restaurants sits alongside.

It is for travellers and diners planning a special meal who want an uncluttered shortlist by city and stars — and for anyone hunting for kosher tables in a strange town.

Chapter two · How it works

Two files, and a list of things removed

Yum runs in your browser from two static JSON files served next to the page, each fetched once; every count, filter, sort and list is computed from them.¹ The script's header records why that matters: an earlier version of the page claimed 671 restaurants and quietly served 315.

The more unusual story is what is missing. When the register was audited in August 2026, the fields nobody could vouch for were deleted: the older dataset's ratings, review counts, price bands and street addresses.² Each card now links to a Google Maps search on the restaurant's name, city and country rather than an address that might be wrong, and the best-of lists rank only by stars and place, because the old ratings had been generated rather than gathered.³

The kosher register is kept deliberately apart from the main collection, in a file of its own.

Chapter three · The test

Counting the register

We loaded the page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026, and downloaded both files at the same moment. The main file was 362,156 bytes and held 1,085 records across 218 cities and 51 countries. Stars broke down as 139 three-star, 129 two-star and 256 one-star rooms, with 561 carrying none. The kosher file held 518 records.⁴

Each record has a name, city, country, neighbourhood, cuisine, tags, a star count and a one-line description. There is no address, no date and no source on any record. The United States accounts for 657 records; France has 48, the United Kingdom 47, Italy 32 and Japan 31.

Chapter four · Against the field

The Michelin Guide and the World's 50 Best

The MICHELIN Guide

Page read 27 September 2026, 20:02 UTC

Michelin's own selection listed 19,745 establishments when we looked, filterable by distinction — 162 with three stars, 545 with two, 3,175 with one and 3,747 Bib Gourmands — and by cuisine, price, dietary needs and services, with photographs and a 'Reserve a table' button.

Where it is better It is the primary source for stars, it is vastly larger, and it helps you actually book, with prices and pictures.

Where Yum goes further It brings restaurants the stars do not cover into the same search — 561 of its 1,085 records carry no stars, with a 'Beyond the guide' list — and it keeps a 518-strong kosher register on the same page.⁵

The World's 50 Best Restaurants

Page read 27 September 2026, 20:02 UTC

The annual list ranks fifty restaurants, with an extended 51 to 100, each named with its city, plus individual awards; registered users can save venues and build their own lists. The old theworlds50best.com address now redirects to the50.com.

Where it is better It is a ranked list, set beside individual awards, which Yum does not attempt.

Where Yum goes further It is searchable across 218 cities by name, cuisine and stars, with every figure computed from its register rather than fixed at a hundred.⁶

Chapter five · Where it falls short

A snapshot, and a lean one

Yum's stars are given 'as last published', with no date or source on any record, while Michelin revises its selection. On the day we counted, Yum had 139 three-star rooms; Michelin's own filter showed 162.

There are no reservations, addresses, prices, opening hours or photographs — the price of deleting what could not be checked. And the register leans heavily American: 657 of 1,085 records, or 61%, are in the United States, against 48 in France.

Chapter six · The verdict

A guide that shows its subtractions

Most guides grow by adding. Yum's most admirable work was subtraction: it removed invented ratings and uncertain addresses, sent each card to a map search instead, and made every number on the page a count of its own file. What is left is smaller, but it is what it says it is.

Use Yum for a quick, honest shortlist by city and stars, and for its kosher register. To book a table, check a price or see the dining room, go to Michelin.

A smaller guide than the famous ones, and one that only says what it can stand behind.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. yum/assets/js/app.js header lines 1–7; fetches at lines 535 and 552.
2. yum/index.html line 720; app.js lines 24–31.
3. app.js lines 24–31 (map search) and 336–347 (lists by stars and place).
4. Page, restaurants.json and kosher.json at 19:58:54 UTC: 55,213, 362,156 and 116,647 bytes.
5. Star counts from restaurants.json; 'Beyond the guide' at app.js lines 343–347; kosher at line 552 and lines 40–41.
6. app.js lines 1–7; 218 cities counted in the file.

A long read · Live Intelligence · Earth

A register of 1,065 public EV charging stations in twelve US cities, sorted by your position without sending it anywhere. We read its data and set it beside the DOE's AFDC locator and PlugShare.

By the LABS editors Tested 27 September 2026, 19:58–20:01 UTC A 4-minute read 763 words

Plate
Cybertaxi on labs.llc, captured from build 537 for this review.

Chapter one · What it is

Chargers near you, and nothing invented

Cybertaxi lists 1,065 public electric-vehicle charging stations across twelve US metro areas, taken from OpenStreetMap, and sorts them by how far each is from you. Each entry gives the operator, connectors, power rating and access terms as mapped — and the site says outright that it cannot tell you whether a stall is free.

It is meant for EV and ride-hail drivers in the covered cities who want a quick, private list of nearby chargers, and for anyone curious about the shape of a city's public charging map.

Chapter two · How it works

Server first, then your browser

Unlike most tools in this issue, Cybertaxi is rendered on a server, in PHP, over a register built from OpenStreetMap's charging-station records, read on 7 September 2026 at 22:04 UTC.¹ The 'near' page prints the 120 fastest stations on the server, so the list works with JavaScript switched off.²

Share your position and the page re-sorts all 1,065 stations in your browser, against a copy of the register, by great-circle distance. Your position is never sent to the server.³ The code calls that distance straight-line, because road distance would need a routing service it does not have.

Two decisions shape the whole site. The first is a refusal: OpenStreetMap records where hardware is, never whether it is occupied, so no page claims a charger is available now.⁴ The second is a history. The site it replaced advertised live cabs, ratings and ride counts that were simply typed into the HTML; the rebuild publishes only what is real, which is the charging network.⁵

Chapter three · The test

What the register holds

We loaded the home page from a local copy of build 537 at 19:58:54 UTC on 27 September 2026 (67,821 bytes), and fetched the station file the browser sorts: 329,732 bytes. Its metadata records 1,065 stations in 12 cities from 80 networks. Of those, 259 are DC fast chargers, 312 carry a power rating and 523 have unknown access terms. Tesla runs 257, ChargePoint 209 and EVgo 73; Los Angeles has 269 stations and Milwaukee 21.⁶

At 19:59 UTC the near page returned 87,789 bytes, with 120 station cards already in the HTML before any script ran. Our local server does not apply the site's short /near/ address, so we requested the PHP file directly.

Chapter four · Against the field

The DOE's locator and PlugShare

AFDC Station Locator (US Department of Energy)

Page read 27 September 2026, 20:01 UTC

The Department of Energy's locator covers alternative fuelling and charging stations across the United States and Canada — electricity, and also biodiesel, ethanol, hydrogen, natural gas, propane and more — with route mapping, counts by state, summary charts, mobile apps and developer APIs.

Where it is better National coverage in two countries rather than twelve metros, plus route planning and APIs.

Where Cybertaxi goes further It sorts by your position without sending it anywhere, gives a working list before any script runs, and prints the date its register was read at the foot of every page instead of implying freshness.⁷

PlugShare

Page read 27 September 2026, 20:01 UTC, via its company site

PlugShare is an EV charging map and driver community covering more than 200 countries and territories, on iOS, Android and the web. Features we could not confirm on its company site are not described here.

Where it is better Worldwide coverage, a very large community of drivers, and apps that route you to a charger.

Where Cybertaxi goes further It needs no account or app, and its register is open data under OpenStreetMap's licence, with the date it was read published.⁸

Chapter five · Where it falls short

Twelve cities, one snapshot

There is no live availability or pricing, by design. The data is a snapshot read on 7 September 2026, twenty days before our test, and it covers twelve metro areas only.

The records are thin: only 312 of 1,065 stations carry a power rating, and 523 have unknown access terms. And distance is straight-line rather than driving distance; the AFDC locator will map a route.

Chapter six · The verdict

Useful because it is modest

Cybertaxi was rebuilt from something that faked its figures, and it shows. It publishes only what can be checked, dates its register on every page, keeps your position in your own browser, and still hands you a working list with JavaScript switched off.

If you drive in one of its twelve cities, it is a fast, private first look. For coverage and routing, the AFDC locator and PlugShare reach far further.

A charging list that keeps your position private and never oversells its data.

Notes on the evidence

Paths are in the labs.llc build 537 tree; times are UTC on 27 September 2026.

1. cybertaxi/data/meta.json (OpenStreetMap, ODbL, read 2026-09-07 22:04:06 UTC); lib.php.
2. cybertaxi/near.php lines 37–48.
3. cybertaxi/app.js line 26 (register copy), 36–44 (distance) and 114–130; near.php lines 4–9.
4. lib.php header lines 23–26; the caveat in meta.json.
5. lib.php header lines 12–21.
6. Home page, stations.json and meta.json at 19:58:54 UTC; near.php at 19:59:00 UTC.
7. near.php lines 4–9 and 36–48; app.js lines 114–130; lib.php lines 23–26, and the footer at lines 264–266, which prints the read date on every page.
8. cybertaxi/data/meta.json.