Screen Test:
the ten reviews.

Channel 01 · Dentist · labs.llc/dentist/

We put the labs.llc dental register on screen at desktop and phone width, weighed the files it downloads, and set it beside the NHS, the ADA and Zocdoc.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:13:23Z Phone capture 20:13:28Z Publisher LABS, which also makes this tool

CH 01 labs.llc/dentist/ captured 20:13:23Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The search box; its placeholder shows the kinds of question it takes — a specialty in a city, a postcode, a name.
2. Callout 2: Near me, open now and — judging by its telephone icon — a has-phone filter.
3. Callout 3: The scope note: “Search downloads only the countries of this region” — printed where you choose one.
4. Callout 4: Specialty tiles counted for the region: 13,318 general practices and 948 orthodontists in Latin America.
5. Callout 5: Two city totals for one scope in the same frame: 2,184 in the summary line, 2,256 on the tile beneath it.

Segment 1 · On screen

What it is, and who it is for

Dentist is a register of dental practices, clinics and specialists drawn from OpenStreetMap: 109,030 of them in 69 countries and 21,077 cities across Europe and the Americas. You search it by city, postcode, specialty or name, in Spanish, Portuguese or English.

It is for anyone who needs a dentist in a city they do not know, and for practice owners who want to see, and correct, what the map says about them. It is not a booking service and does not act like one.

Segment 2 · Behind the picture

How it works

Almost all of the work happens on your device. The client engine, dentist.js (107,621 bytes), first loads data/meta.json — counts, city aliases and fourteen specialties — then fetches only the country files for the region you pick, indexes them in memory and searches there (lines 218–226). The country files are an extract of every OpenStreetMap object tagged amenity=dentist or healthcare=dentist, built on 4 September 2026. Crawlable country and city pages come from place.php, which counts every figure at request time (lines 12–16).

The effort shows in the search code. A Damerau–Levenshtein check (dl1, line 253) forgives a one-letter slip, “ortod” finds Ortodoncia, and city aliases for eleven countries let CDMX reach Mexico City. Open now uses each practice’s own mapped hours (lines 640 and 699); a practice with no hours is never guessed open.

• OpenStreetMap (ODbL) extract built 2026-09-04: data/meta.json plus per-country shards, 21 MB on disk
• Your browser’s own location, for near me — used on the device only

Segment 3 · The test

What we measured, and when

At 20:13:18 UTC on 27 September we asked the local copy of build 537 for the page and its register. The page weighed 32,281 bytes; meta.json 417,660; the Mexico shard 747,173; the United States shard 4,891,172. It agreed with the headline figures and added the numbers a patient cares about: 45,895 practices with a postcode, 37,766 with a phone, 27,660 with opening hours.

One test failed on our side: the place pages need an Apache rewrite our local server does not apply, so they answered 404 and 301, and we have not reviewed them. The desktop capture shows something a reader will notice (callout 5): two city totals for Latin America on one screen. The frame does not say which definition each uses.

390 × 844 20:13:28Z

Callouts · phone capture, 390 × 844

1. Callout 1: The near-me, open-now and phone buttons stay one tap each at phone width.
2. Callout 2: The download note survives at 390 pixels, directly under the region buttons.

Segment 4 · Other channels

Beside NHS Find a dentist, ADA Find-a-Dentist and Zocdoc

Each comparator was checked on the date shown. We say what each does better before we say where Dentist goes further.

NHS Find a dentist (England)

Checked 2026-09-27 20:17 UTC

What it is

The NHS search for dental practices by town, city or postcode in England, or by your location.

What it does better

It says which practices accept new NHS patients or offer urgent care — a fact OpenStreetMap does not hold — and it is the health service’s own record.

Where Dentist goes further

Dentist covers 69 countries rather than England alone, and searches in three languages with typo tolerance and city aliases.

ADA Find-a-Dentist

Checked 2026-09-27 20:17 UTC · the site refused our fetch, so we read the ADA’s own article on its redesign and its pages via search

What it is

The American Dental Association’s directory of member dentists, searched by address or ZIP code.

What it does better

Filters for insurance accepted and languages spoken, and profiles the dentists maintain themselves.

Where Dentist goes further

Only ADA members who opt in appear there; Dentist lists whatever has been mapped, and computes open-now from each practice’s own hours.

Zocdoc (dentists)

Checked 2026-09-27 20:18 UTC · refused our fetch; verified from its pages via search

What it is

A US booking marketplace: dentists by location and insurance, with open appointment slots.

What it does better

Real availability, online booking, an insurance filter and reviews from patients who booked through it.

Where Dentist goes further

Nobody pays to be listed or ranked on Dentist (index.html line 475), and every record links to OpenStreetMap so anyone can view or correct it.

Segment 5 · Dropouts

Where it falls short

• Coverage is only what volunteers have mapped, and the page says so. Contact details are thin: 35% of practices carry a phone number and 25% carry opening hours.
• It is a snapshot. The extract was 23 days old on test day; a correction made in OpenStreetMap appears only in the next edition.
• A search in the United States downloads the whole 4.9 MB US file first — a heavy first search on a phone connection.
• No booking, no insurance filter, no patient reviews, and no check that a practice is still trading.

Segment 6 · End card

The verdict

Dentist is a wide, candid dental register: 109,030 practices in 69 countries, searched on your own device, with its gaps counted rather than filled. Use it to learn who practises in an unfamiliar city; use the NHS or a booking service to learn who will see you.

Open Dentist on labs.llc How we tested

Channel 02 · Grid · labs.llc/grid/

Europe’s fuel mix and Britain’s carbon intensity on one desk — captured, measured through its relay, and set beside Electricity Maps and Energy-Charts itself.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:13:34Z Phone capture 20:13:40Z Publisher LABS, which also makes this tool

CH 02 labs.llc/grid/ captured 20:13:34Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow states the job: “The power desk · Europe’s grid, read live”.
2. Callout 2: The lede promises “quarter-hours for the big grids, whole hours for a few small ones” — a claim our test checked.
3. Callout 3: “Fetched through this page’s own five-minute relay”: the desk’s one server part, named on its first screen.
4. Callout 4: Britain’s carbon intensity drawn as fourteen regions straight off NESO’s open API, with no key and no account.
5. Callout 5: The inventory line: 15 countries, 10 fuel inks, 48 hours drawn, 14 UK regions.

Segment 1 · On screen

What it is, and who it is for

Grid is a power desk with two instruments. The first draws the last 48 hours of electricity generation, fuel by fuel, for any of 15 European countries, with the latest mix in percentages. The second shows Great Britain’s carbon intensity — grams of CO₂ per kilowatt-hour — as a national figure, a 24-hour forecast and fourteen regions on a tile board.

It is for anyone curious how their country’s power is being made right now, for energy writers and students, and for people in Britain choosing a cleaner hour to run a heavy load.

Segment 2 · Behind the picture

How it works

The two halves travel differently. Fraunhofer ISE’s Energy-Charts does not send the header a browser on another site needs, so Grid runs a relay, grid-proxy.php, on its own origin. The relay is deliberately narrow: one upstream route (public_power, the last 48 hours), a whitelist of fifteen countries, a window the server chooses, and a five-minute cache per country (lines 11–25, 49–56, 64–66). A crowd watching Germany costs Energy-Charts one call every five minutes, and the page refreshes on the same rhythm (app.js line 30). The British half needs no relay: your browser reads NESO’s API directly for the national figure, the forecast and the regions (lines 401–405).

On the way to the chart, the code declines to guess. It reads the time step off each series instead of assuming one (resLabel, line 182), trims quarter-hours nobody has reported into yet rather than drawing them as zero (line 217), and leaves out of the legend any fuel a country does not use. The regional British figures are labelled as the forecasts they are, on a drawn board rather than a survey map.

• api.energy-charts.info /public_power (Fraunhofer ISE), through grid/grid-proxy.php
• api.carbonintensity.org.uk /intensity, /fw24h and /regional (NESO), read by your browser

Segment 3 · The test

What we measured, and when

At 20:14:02 UTC we asked the relay for Germany. It answered in 0.0015 seconds — from its cache — with 27,202 bytes: 186 points at 15-minute steps and 21 series, from lignite and hard coal to offshore wind, solar, load and the renewable share. Switzerland, asked in the same second, came fresh from upstream in 0.44 seconds: 44 hourly points and 12 series, nuclear and run-of-river hydro among them. So callout 2 holds up — one country quarter-hourly, the other hourly, each labelled as it is.

We then asked for Great Britain, which is not on the relay’s list. It refused with HTTP 400 before fetching anything: this is not an open proxy. At 20:14:07 UTC the British regional feed returned 18 entries for the 19:30–20:00 window: Great Britain at 84 g/kWh (“low”), North Scotland at 0 and South Wales at 331 (“very high”).

390 × 844 20:13:40Z

Callouts · phone capture, 390 × 844

1. Callout 1: At 390 pixels the lede sets cleanly, with the relay named in the first paragraph.
2. Callout 2: In this first-screen capture the floating Ask launcher sits over the inventory line. We did not test whether it moves aside on scroll.

Segment 4 · Other channels

Beside Electricity Maps and Energy-Charts

Each comparator was checked on the date shown. We say what each does better before we say where Grid goes further.

Electricity Maps

Checked 2026-09-27 20:18 UTC

What it is

A free live map of carbon intensity and electricity production by zone, worldwide, with a paid API and datasets.

What it does better

Zones on every continent, carbon intensity for each of them, and history and forecasts beside the live signal.

Where Grid goes further

Grid names the single upstream it reads and prints each series’ own time step and read time on the page (app.js lines 167–172), and draws Britain’s regional forecast as a fourteen-tile board in the API’s own bands.

Energy-Charts (Fraunhofer ISE)

Checked 2026-09-27 20:18 UTC · the site returned only its title to our fetch; verified from Fraunhofer ISE’s press release on its tenth year

What it is

The source itself: generation, trade, emissions, exchange prices and capacity for 42 European countries, with an open API.

What it does better

Far deeper — 42 countries, years of history, prices and emissions — and it is the primary record Grid reads.

Where Grid goes further

Grid sets the continental mix beside Britain’s carbon intensity from two different public sources on one page, and holds its own load on the source to one call per country per five minutes.

Segment 5 · Dropouts

Where it falls short

• Fifteen countries, while Energy-Charts covers 42.
• Forty-eight hours only: the relay fixes the window and a visitor cannot lengthen it.
• Carbon intensity for Great Britain alone; the continental countries get a fuel mix but no g/kWh figure.
• No prices, cross-border flows or generation forecasts, all of which Energy-Charts publishes.

Segment 6 · End card

The verdict

Grid is a small, exact window on two large datasets. It shows less than either source, but every number it does show carries its own step, its own read time and its own origin. Watch it for the shape of today; go to Energy-Charts for the long view.

Open Grid on labs.llc How we tested

Channel 03 · Rx.Luxury · labs.llc/rxluxury/

A clinical notebook, a DICOM viewer and two federal directories — the relays measured, the entry screen captured, and the whole set beside Medicare Care Compare and OHIF.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:13:45Z Phone capture 20:13:51Z Publisher LABS, which also makes this tool

CH 03 labs.llc/rxluxury/ captured 20:13:45Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The desk lists its parts: scheduling, physiology, medications, telehealth, literacy and an in-browser MRI/DICOM lab.
2. Callout 2: “Everything stays on this device” — a promise the code backs, as the next segment shows.
3. Callout 3: The gate asks only how to address you. No account, no email address.
4. Callout 4: “Not a medical device. Not a substitute for professional care.” — printed under the button, not buried.

Segment 1 · On screen

What it is, and who it is for

Rx.Luxury is a personal clinical desk that runs in the browser: intake notes, symptoms, medications, vitals, a printable visit packet and a viewer for DICOM scan files. Beside it sit two live US directories — clinicians from the federal NPI Registry, and every Medicare-certified hospital with its CMS overall rating where CMS publishes one.

It is for patients and carers in the United States who want their own notes in one place, a tidy packet to take to an appointment, a look at their own scans, and clinicians and hospitals from the official registers rather than an advertiser’s list.

Segment 2 · Behind the picture

How it works

Your side of the desk stays in your browser. State lives in localStorage (app.js lines 64–73); the PDF and ZIP exports are written by hand in save.js, so no outside library is fetched (lines 1–13); scans are parsed locally by a bundled dicom-parser 1.8.21 and drawn with window presets for brain, bone, lung and soft tissue (dicom.js lines 1–25). Mail is opt-in, and only to an address you have confirmed with a six-digit code.

The directories pass through two same-origin relays because neither federal service sends a CORS header, and what those relays leave out is the telling part. An earlier build listed eight invented clinicians with ratings and “next available” times; npi-proxy.php replaced them with the NPPES register and dropped every field it does not hold — no ratings, no insurance networks, no availability (lines 7–31). For hospitals, cms-proxy.php skips the NPI-2 search, which returns billing entities rather than buildings, and reads the CMS facility file, spelling its filters out in full because the short form is silently ignored and returns all 5,419 rows (lines 6–40).

• NPPES NPI Registry API v2.1, through rxluxury/npi-proxy.php
• CMS Hospital General Information (dataset xubh-q36u), through rxluxury/cms-proxy.php
• OpenStreetMap Nominatim and Overpass, from your browser, for Germany and Panama
• Bundled locally: dicom-parser 1.8.21

Segment 3 · The test

What we measured, and when

Our capture (20:13:45 UTC) shows the entry gate, not the desk behind it, so the callouts describe the first screen only. The relays we measured directly. At 20:14:27 UTC we searched Boston clinicians the way a person might type it — “Cardiology”. The relay answered in 0.083 seconds with six rows, every one a pharmacist with a cardiology specialty. At 20:14:35 we used the register’s own term, “Cardiovascular Disease”: 114 rows in 0.097 seconds, each with an NPI number, credential, address and phone. The page’s own picker supplies the register’s terms; a reader typing plain words would be misled.

Boston hospitals, asked at 20:14:27 UTC, took 1.70 seconds: twelve hospitals, each with CMS id, address, county, phone, type, ownership and an emergency-department flag. Six carried a CMS overall rating — Massachusetts General, Brigham and Women’s and New England Baptist at 5, Boston Medical Center and Beth Israel Deaconess at 3, Tufts at 2 — and six had none.

390 × 844 20:13:51Z

Callouts · phone capture, 390 × 844

1. Callout 1: The same gate at 390 pixels: one field and one button.
2. Callout 2: The disclaimer keeps its place beneath the button.

Segment 4 · Other channels

Beside Medicare Care Compare and OHIF Viewer

Each comparator was checked on the date shown. We say what each does better before we say where Rx.Luxury goes further.

Medicare Care Compare

Checked 2026-09-27 20:19 UTC · the page is built by script and returned no text; verified from medicare.gov pages via search

What it is

Medicare’s official tool for finding and comparing approved providers — clinicians, hospitals, nursing homes, home health, hospice, dialysis and more.

What it does better

Many more provider types and quality measures, and clinician profiles with board certification, education and hospital affiliations.

Where Rx.Luxury goes further

Rx.Luxury keeps your own intake, medications and vitals beside the directory and turns them into a visit packet in the browser, and it opens your own DICOM scans with windowing presets.

OHIF Viewer

Checked 2026-09-27 20:19 UTC

What it is

An open-source, browser-based medical imaging viewer, a programme of Massachusetts General Hospital, with annotation, segmentation, PET/CT and volume rendering.

What it does better

A far more capable imaging workstation, open source, with a large community behind it.

Where Rx.Luxury goes further

Rx.Luxury’s viewer sits inside a patient-side desk with intake, medications, vitals and the federal directories, and exports PNG, PDF and ZIP without a third-party library.

Segment 5 · Dropouts

Where it falls short

• The hospital relay’s title-casing capitalises the letter after an apostrophe: our Boston answer printed “Brigham And Women’S Hospital” and “Boston Children’S Hospital” (cms-proxy.php line 56).
• Clinician search depends on the register’s wording: “Cardiology” found six pharmacists and no cardiologists.
• No insurance, availability or quality data for clinicians — by design, but Care Compare gives more.
• Outside the US there is only an OpenStreetMap radius search for Germany and Panama, which the code itself calls “a real downgrade from NPPES”.

Segment 6 · End card

The verdict

Rx.Luxury’s best feature is a deletion: it threw out invented doctors and now shows only what the federal registers hold, while your notes stay in your browser unless you choose to mail yourself a packet, and your scans never leave the machine. Keep your records here; check credentials and quality on Care Compare.

Open Rx.Luxury on labs.llc How we tested

Channel 04 · Weather · labs.llc/weather/

Open-Meteo forecasts, National Weather Service warnings and a satellite-and-radar map, captured and measured, then set beside Windy.com and weather.gov.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:13:56Z Phone capture 20:14:02Z Publisher LABS, which also makes this tool

CH 04 labs.llc/weather/ captured 20:13:56Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow: “The weather desk · any city, read live”.
2. Callout 2: “The numbers are Open-Meteo’s open forecast wire — model data, fetched by this page itself, no key, no account.”
3. Callout 3: US cities also get the National Weather Service’s active warnings — and the page says plainly when there are none.
4. Callout 4: The inventory: any city by name, 48 hours drawn, 7 days, 42 cities stamped live, °F or °C remembered.
5. Callout 5: “The sky” in the navigation leads to the satellite, radar and lightning map below this first screen.

Segment 1 · On screen

What it is, and who it is for

Weather is a live desk for any city by name: current conditions, 48 hours of temperature, seven days drawn from low to high, National Weather Service warnings for US points, a board of 42 world cities, and a sky map of satellite, radar and lightning with a model forecast clock.

It is for anyone checking a place, travellers comparing several, and people who want the satellite and radar picture without installing an app or opening an account.

Segment 2 · Behind the picture

How it works

There is no server in the middle: the page’s script fetches everything from your browser (app.js lines 1–8). One refresh is three separate reads — the Open-Meteo forecast for your city, with temperature, humidity, apparent temperature, weather code, wind and gusts, pressure, 48 hours and seven days (lines 155–161); the NWS alerts at that point if it is in the US (line 191); and a single multi-point Open-Meteo request that stamps all 42 world cities at once (line 208). Each read settles on its own, so a slow warnings wire never blanks the forecast. Weather codes are printed in the WMO’s own words (lines 18–26). It refreshes every ten minutes, only while the tab is visible.

The sky map goes furthest. It loads a self-hosted MapLibre and assembles layers from national services — NASA GIBS satellite, NOAA nowCOAST radar and lightning, Canada’s MSC GeoMet radar, Germany’s DWD radar — plus Open-Meteo model grids (skymap.js lines 40–43). Its list of frames comes from each service’s own capabilities, never from the clock; every frame prints its UTC and local time, an old layer says “held” and how old, and a step with failed tiles says “partial” (lines 33–39).

• api.open-meteo.com /v1/forecast and geocoding-api.open-meteo.com, from your browser
• api.weather.gov /alerts/active, US points only, from your browser
• Sky map: NASA GIBS, NOAA nowCOAST, MSC GeoMet, DWD, Open-Meteo grids and NOAA tropical layers, on OpenFreeMap base styles

Segment 3 · The test

What we measured, and when

The page came back from the local copy at 20:15:01 UTC at 57,309 bytes. In the same second we made the forecast read the page makes, for Miami: 5,024 bytes in 0.40 seconds, 168 hourly points, seven daily rows, the America/New_York time zone, and a current 28.6 °C under weather code 0 — clear sky — on a 15-minute interval.

The NWS read at the same point took 1.02 seconds and returned two active alerts: a Coastal Flood Advisory and a Rip Current Statement. That is the kind of official notice a model forecast alone would never show, and it arrives beside the forecast without holding it up. Our desktop capture (20:13:56 UTC) is the first screen, where the page describes its own wiring; the code matches the description. We did not capture the sky map itself, which needs WebGL.

390 × 844 20:14:02Z

Callouts · phone capture, 390 × 844

1. Callout 1: The warnings paragraph at 390 pixels, readable without zooming.
2. Callout 2: In the first screen the Ask launcher covers part of the inventory line. We did not test it on scroll.

Segment 4 · Other channels

Beside Windy.com and National Weather Service

Each comparator was checked on the date shown. We say what each does better before we say where Weather goes further.

Windy.com

Checked 2026-09-27 20:19 UTC · the page returned only its title; verified from windy.com pages and its App Store listing via search

What it is

An interactive forecast map with the leading global models and many local ones, more than fifty layers, radar and satellite.

What it does better

Far more layers and models, including local high-resolution ones, a richer point forecast, and 15-day ECMWF and GFS on Premium.

Where Weather goes further

Weather states the age of every map layer and marks held or partial frames instead of looping a picture that looks current; it puts the NWS’s own warnings beside the forecast; and every source it reads is keyless, straight from your browser.

National Weather Service (weather.gov)

Checked 2026-09-27 20:19 UTC

What it is

The US government’s forecast service: a warnings map, local forecasts by city or ZIP, radar, satellite, rivers and air quality.

What it does better

The authority for US warnings, with forecaster detail, hydrology, fire weather and air quality the labs desk does not carry.

Where Weather goes further

Weather works for any city the geocoder knows and keeps a live board of 42 world cities, and its single map combines GOES, Himawari and VIIRS satellite with US, Canadian and German radar and lightning.

Segment 5 · Dropouts

Where it falls short

• The city board charts hourly temperature only — no hour-by-hour rain, wind or cloud.
• Official warnings exist only for US points; a city in Europe or Asia shows none.
• Seven days of forecast, against fifteen on Windy Premium.
• The moving map needs WebGL; without it the map says it cannot draw, leaving the boards.

Segment 6 · End card

The verdict

Weather’s distinction is candour about time: every frame dated, every stale layer called stale. Its forecast is thinner than Windy’s and its warnings stop at the US border, but it will not pass off an old sky as a new one.

Open Weather on labs.llc How we tested

Channel 05 · Quake · labs.llc/quake/

The survey’s own GeoJSON feed turned into a wire, a map and a set of meters that cannot disagree — measured, captured, and set beside the USGS map and EMSC.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:14:08Z Phone capture 20:14:13Z Publisher LABS, which also makes this tool

CH 05 labs.llc/quake/ captured 20:14:08Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow names the source outright: “USGS, read live”.
2. Callout 2: “Fetched by this page directly and read again every five minutes while you watch.”
3. Callout 3: “A magnitude here is whatever USGS says it is” — often automatic at first, revised once an analyst reviews it.
4. Callout 4: Three windows: hour, day and week.
5. Callout 5: “The meters” — the counts by magnitude band and depth.

Segment 1 · On screen

What it is, and who it is for

Quake is a seismic desk built on the US Geological Survey’s own feed. Choose the past hour, day or week and it shows every event the survey has posted: a newest-first wire, a world map with events ringed by magnitude, counts by magnitude band and by depth, and an optional push alert above a magnitude you set.

It is for people who felt something or saw a headline and want the survey’s list rather than a news summary, for students, and for anyone who wants a quiet notice when a large earthquake happens.

Segment 2 · Behind the picture

How it works

Your browser reads one USGS GeoJSON summary feed per window — keyless and open to any site — and draws the wire, the map and the meters from that single read (app.js lines 1–15). The code’s own comment puts the rule plainly: three renderings of one fetched window, never three fetches, so the list and the counts cannot disagree. It reads again every five minutes. The coastline is Natural Earth’s public-domain 1:110m outline, simplified and carried inside the page, so no map tiles are fetched. Events the feed marks as something other than an earthquake, such as quarry blasts, carry their type on the wire (line 126), and when the wire is cut to the newest 100 it says so while the map and meters still count everything (lines 135–136).

The alerts show the same restraint. alerts.php only stores and removes Web Push subscriptions, after checking the endpoint belongs to one of four push vendors (lines 17–19). A scheduled script reads the hourly feed itself, compares it with what it saw last and sends an empty, signed push; its first run seeds and sends nothing, and no figure ever travels in the notification. Tapping it opens the desk, which reads the wire live.

• earthquake.usgs.gov summary feeds all_hour, all_day and all_week (GeoJSON), from your browser
• Natural Earth 1:110m coastline (public domain), carried inside the page
• For alerts: the same hourly feed, read by the scheduled sender on the server

Segment 3 · The test

What we measured, and when

The page came back from the local copy at 20:15:16 UTC at 57,324 bytes. In the same second we read the past-day feed the page reads: 143,309 bytes in 0.04 seconds, 199 events — 8 at magnitude 4.5 or above, 34 between 2.5 and 4.5, and 157 below 2.5. The largest was M5.1, 90 km east-north-east of Tadine, New Caledonia. With 199 events the wire would show the newest 100 and say so, while the meters counted all 199.

The alerts door answered in 90 bytes: no devices, no runs. The sender is a scheduled job that does not run on the local copy, so we did not exercise the live alert path and make no claim about delivery. Our first-screen capture dates from 20:14:08 UTC.

390 × 844 20:14:13Z

Callouts · phone capture, 390 × 844

1. Callout 1: “No key, no account, no server in the middle” stays in the first screen at phone width.
2. Callout 2: The Ask launcher sits over part of the figures line in this capture; we did not test it on scroll.

Segment 4 · Other channels

Beside USGS Latest Earthquakes and ENS and EMSC

Each comparator was checked on the date shown. We say what each does better before we say where Quake goes further.

USGS Latest Earthquakes and ENS

Checked 2026-09-27 20:19 UTC · the map page returned only a browser notice; verified from the survey’s feed page and usgs.gov via search

What it is

The survey’s own near-real-time map of up to 30 days of earthquakes, feeds in five formats, a catalogue API, and ENS, a free email and phone notification service.

What it does better

The primary source, with 30 days on the map, full catalogue search, per-event products, and notifications by email or text for regions you choose.

Where Quake goes further

Quake counts the window by magnitude band and depth on the same screen as the wire, and its push carries no numbers at all — it opens the live desk — with a deliberately silent first run.

EMSC (European-Mediterranean Seismological Centre)

Checked 2026-09-27 20:19 UTC

What it is

A world map and list of latest earthquakes from member institutes, eyewitness felt reports, and the LastQuake apps and messenger bot.

What it does better

Crowdsourced felt reports and testimonies, and strong Euro-Mediterranean coverage from many national institutes.

Where Quake goes further

Every figure on Quake comes from one read of one feed, the page says when its list is cut, and it runs as a single light page with the coastline built in.

Segment 5 · Dropouts

Where it falls short

• The map is a fixed 1:110m world outline with no zoom, so a cluster of small events in one region cannot be pulled apart.
• Three windows only; the USGS map reaches back 30 days and its catalogue allows any search.
• None of the survey’s derived products — Did You Feel It?, ShakeMap, tsunami flags — beyond a link from each row.
• One network: outside the US, regional agencies such as EMSC may publish first or with a different magnitude.

Segment 6 · End card

The verdict

Quake promises little and keeps all of it: one feed, one read, three views that agree. For depth, history or felt reports go to the USGS or EMSC; for a glance you can trust, this is enough.

Open Quake on labs.llc How we tested

Channel 06 · DNS propagation · labs.llc/domains/dns/

A propagation checker that asks the domain’s own nameservers first and measures every resolver against them — captured, run on wikipedia.org, and set beside two familiar map-based checkers.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:14:19Z Phone capture 20:14:24Z Publisher LABS, which also makes this tool

CH 06 labs.llc/domains/dns/ captured 20:14:19Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The promise in the lede: fourteen public resolvers, and the domain’s own nameservers asked directly.
2. Callout 2: The receiver’s header: UDP 53 · 14 resolvers · the authority, recursion off.
3. Callout 3: Ten record types, A to HTTPS.
4. Callout 4: Three instrument skins — T3, ET66 and AB1 — for the same reading.
5. Callout 5: What a reading holds: the truth, fourteen resolvers, TTL countdowns, who differs.

Segment 1 · On screen

What it is, and who it is for

DNS propagation is one of ten instruments on the labs.llc domains desk. You give it a name and a record type; it puts the question to fourteen public resolvers and, beside them, to the domain’s own authoritative nameservers, so every answer is judged against the source. Each cached copy’s remaining lifetime is counted down live.

It is for site owners and administrators who have just changed a record and need to know who still holds the old answer, and for how much longer.

Segment 2 · Behind the picture

How it works

This instrument does its asking on the server, through a same-origin relay (domains/whois/tools.php, op=dns). It does not trust a single answer when it looks for the authority: it asks the parent zone’s servers, takes a referral from a parent server with recursion off, resolves every nameserver, checks that each address is public, and asks each one with recursion off (tools_dns.php lines 16–24). Only then are the fourteen resolvers asked the same question over UDP port 53, with EDNS NSID requested so the page can name the anycast node that answered.

The list reaches past the usual Western set: Google, Cloudflare, Quad9, OpenDNS, Level3, AdGuard, Hurricane Electric, CleanBrowsing, UltraDNS, Comodo and DNS.SB, plus Yandex in Moscow and 114DNS and Baidu inside mainland China. The browser script, dc-dns.js, then applies three rules: the authoritative answer is the truth the others are measured against; a resolver that did not answer counts as no answer, never as agreement; and a different answer is not automatically an old one — location-aware DNS, filtering resolvers and the national firewall are named as likelier reasons, marked as probable (lines 18–24).

• Fourteen public resolvers, asked directly over UDP 53
• The domain’s own authoritative nameservers, found through the parent zone and asked with recursion off
• All through domains/whois/tools.php on the labs.llc server

Segment 3 · The test

What we measured, and when

At 20:15:42 UTC we asked the local copy for wikipedia.org, type A. The relay answered in 0.33 seconds (322 ms inside the engine), uncached, after 25 DNS questions and no errors. Three Wikimedia nameservers answered with authority — the aa flag set, TTL 180, serial 2026060420 — giving 208.80.154.224. All fourteen resolvers answered; ten matched the authority. Yandex and 114DNS returned 185.15.59.224, DNS.SB 103.102.166.224 and Baidu 31.13.68.169. Remaining lifetimes ran from 2 seconds at Cloudflare to the full 180.

The verdict it drew was not “not propagated”. It read four different answers from fourteen resolvers as geo-DNS suspected, and said a different address here is expected and is not a stale record; its notes named the filtering resolvers and the two mainland-China resolvers separately, each reason marked probable. That caution is earned: three of the odd answers end in .224, as the authority’s does, but Baidu’s 31.13.68.169 does not, so geo-DNS alone may not explain it. The desktop capture is from 20:14:19 UTC.

390 × 844 20:14:24Z

Callouts · phone capture, 390 × 844

1. Callout 1: At 390 pixels the ten record types wrap into two rows.
2. Callout 2: The Ask launcher covers the instrument picker in this first-screen capture.

Segment 4 · Other channels

Beside DNSChecker.org and whatsmydns.net

Each comparator was checked on the date shown. We say what each does better before we say where DNS propagation goes further.

DNSChecker.org

Checked 2026-09-27 20:20 UTC · refused our fetch; verified from its pages via search

What it is

A free propagation checker that queries DNS servers in several regions and shows a map and a list, with a tick or a red cross per server.

What it does better

Answers from servers located in many regions, on a map, and you can add a custom DNS server to the check.

Where DNS propagation goes further

The labs instrument asks the authoritative nameservers directly and judges each resolver against them, shows each resolver’s remaining TTL, and explains a difference — geo-DNS, filtering, the firewall — rather than printing a bare cross.

whatsmydns.net

Checked 2026-09-27 20:20 UTC · refused our fetch; verified from its pages via search

What it is

A propagation checker that looks a name up against the servers of large providers around the world and plots the results on a world map.

What it does better

A world map of answers from servers in many countries; quick and familiar, with a separate all-records lookup.

Where DNS propagation goes further

The labs instrument names the anycast node that answered when the resolver reports it, admits that each answer comes from one node, and never counts a failed resolver as agreement.

Segment 5 · Dropouts

Where it falls short

• All fourteen resolvers are asked from one place — the labs.llc server — so each anycast answer comes from the node nearest that server, not from your city. The page says so; the two map-based checkers present answers from servers in many regions.
• Ten record types; no SRV, PTR or DNSKEY.
• Twenty fresh readings a minute and 600 a day per visitor, and no way to add your own resolver.

Segment 6 · End card

The verdict

This instrument settles the authoritative answer before it compares anyone with it, and tells you why a resolver differs instead of simply marking it wrong. It sees the world from one vantage point and says so; when you need many vantage points, use a map-based checker alongside it.

Open DNS propagation on labs.llc How we tested

Channel 07 · History of the Internet · labs.llc/history/

Sixty years as a dated record with a strict rule about what a date means — counted, captured, and set beside the Internet Society’s essay and the Computer History Museum’s timeline.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:14:30Z Phone capture 20:14:36Z Publisher LABS, which also makes this tool

CH 07 labs.llc/history/ captured 20:14:30Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow: “The internet · a dated record”.
2. Callout 2: The rule, set out in the first screen: every date here is the date the thing WORKED.
3. Callout 3: “What is checked” — the method section, one click from the top.
4. Callout 4: “People”: fourteen of them, drawn from the same record as the timeline.

Segment 1 · On screen

What it is, and who it is for

History of the Internet sets down sixty years as a dated record: 28 entries from Shannon’s bit in 1948 to IPv4 running out in 2011, filterable by six eras, with five drawn plates of the core ideas, fourteen people, and three disputed claims printed as disputed.

It is for students, teachers and curious readers who want the order of events and how each piece worked without a textbook — and for anyone settling an argument about who did what first.

Segment 2 · Behind the picture

How it works

No outside data is fetched while you read. The whole record is one file, history-data.js — 28 events, 14 people, 6 eras — and history.js draws the timeline, the era filter, the people and the plates from it (lines 1–7). Because there is one record, a date corrected once is corrected everywhere, and the era counts cannot drift away from the entries.

The page’s discipline is in what counts as a date. Every entry is dated to when the thing worked, not when it was proposed, funded or announced (history-data.js lines 4–7): ARPANET to its first message, TCP/IP to the flag day, the web to the first page served from another machine. Where historians genuinely disagree, the entry carries the dispute instead of naming a winner — packet switching reached independently by Baran and Davies, the first ARPANET message, the first email between two machines (history.js line 270). And it refuses photographs: the five plates are drawings of mechanisms, such as packets crossing a mesh or intelligence at the ends of a simple network (lines 8–14).

• None fetched at run time: the record is written into history/assets/js/history-data.js
• No bibliography or source links in the record (none found when we searched it)

Segment 3 · The test

What we measured, and when

At 20:16:11 UTC the page came back from the local copy at 41,311 bytes in 0.0013 seconds, and its record file at 14,217 bytes. We ran the record through JavaScriptCore, the engine built into macOS, and counted: 28 events, 14 people, 6 eras (ideas, ARPANET, protocols, names, web, after), 3 disputed entries, 9 dated to an exact day, the last in 2011 — “IPv4 runs out”.

Then we searched the HTML itself for “Shannon” and “Kleinrock” and found neither. The timeline, the people and the plates are written into the page by script; the HTML holds the headings and empty containers. Our capture (20:14:30 UTC) shows the first screen, which is plain HTML and reads well; everything below it depends on JavaScript.

390 × 844 20:14:36Z

Callouts · phone capture, 390 × 844

1. Callout 1: The dating rule stays in the first screen at 390 pixels.
2. Callout 2: The three buttons wrap onto two rows.

Segment 4 · Other channels

Beside Internet Society and Computer History Museum

Each comparator was checked on the date shown. We say what each does better before we say where History of the Internet goes further.

Internet Society — A Brief History of the Internet

Checked 2026-09-27 20:20 UTC

What it is

A 1997 essay by people who built the internet — Leiner, Cerf, Clark, Kahn, Kleinrock, Lynch, Postel, Roberts and Wolff — from Licklider’s “Galactic Network” to commercialisation in the mid-1990s.

What it does better

First-hand authority, and ten numbered footnotes with a references section a reader can follow.

Where History of the Internet goes further

The labs page carries the story on to 2011, dates each step to when it worked, flags disputes in place, draws each mechanism, and filters by era.

Computer History Museum — Internet History

Checked 2026-09-27 20:20 UTC

What it is

The museum’s decade-by-decade timeline from 1962 to 1992, with a poster of the period.

What it does better

A museum’s collection and imagery behind it, and more entries per decade between 1962 and 1992.

Where History of the Internet goes further

The labs page begins with the ideas — Shannon in 1948, Kleinrock in 1961 — runs past 1992 to the web, DNS and IPv4 exhaustion, and marks its three contested claims as contested.

Segment 5 · Dropouts

Where it falls short

• No sources are cited. The record carries no links or references, so a date cannot be checked against a primary document; the Internet Society’s essay can be.
• With JavaScript off, the timeline, the people and the plates are empty.
• It stops in 2011: nothing on mobile, the cloud or anything since.
• Twenty-eight entries choose the jumps; this is not a year-by-year chronology.

Segment 6 · End card

The verdict

A strict, well-drawn record with a clear rule about dates and an even hand with disputes — let down by citing nothing. Read it for the order and the mechanisms; read the Internet Society’s essay for the footnotes.

Open History of the Internet on labs.llc How we tested

Channel 08 · EZ Score · labs.llc/ezscore/

A scoreboard that reads the betting line as information — captured, measured against the live NFL board, and set beside ESPN’s own scoreboard and Flashscore.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:14:41Z Phone capture 20:14:47Z Publisher LABS, which also makes this tool

CH 08 labs.llc/ezscore/ captured 20:14:41Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow: “Live scores · with the line”.
2. Callout 2: “Where each line opened and where it closed. Ten leagues — UFC fight cards included.” The code now holds 26 competitions (see the test).
3. Callout 3: When a game has no posted line the card says so; when the wire is down, the board says that instead of pretending.
4. Callout 4: “Where the numbers come from” — the source, one click from the first screen.
5. Callout 5: “Refresh the wire” in the header.

Segment 1 · On screen

What it is, and who it is for

EZ Score is a scoreboard that keeps the betting line in the same glance as the game: score, clock and situation beside the spread, the moneyline and the total, with where each line opened against where it sits now. It covers UFC, MLB, NFL, college football, NBA, WNBA, NHL, college basketball, the Premier League and MLS — and, in the current code, fourteen European football competitions and ATP and WTA tennis.

It is for fans following several leagues at once, and for people who read lines as information about how the market sees a game. It takes no bets and links to no sportsbook.

Segment 2 · Behind the picture

How it works

Everything is read by your browser from ESPN’s public, keyless site API: scoreboards from one path, standings from another (ezscore.js lines 1–8, 106–115). Live games are polled every 30 seconds, and the league, the view and the day sit in the address, so a pasted link reproduces the screen.

The care is in the odds. The team found that its own first version struck through the opening price — the vig — instead of the opening handicap, which hid real moves: on one college football slate, 82 of 169 games moved the handicap and not one showed it. Both are now carried, so a handicap move shows even when the price stays the same (lines 255–296). ESPN returns the literal word “OFF” when a book pulls a market; EZ Score treats it as no value instead of a quote (lines 269–275) — on one board, 10 of 25 events had read “Moneyline OFF / OFF”. The book behind the numbers is named on every card from the wire’s own provider field, an answer as long as the wire’s page cap is flagged as possibly cut, and each league was checked before it was added, European football only where every fixture carried a line.

• site.api.espn.com scoreboards — scores, clocks, odds with the book named — from your browser
• site.api.espn.com standings, from your browser
• Odds as ESPN carries them from its provider; DraftKings on the NFL board we measured

Segment 3 · The test

What we measured, and when

The page came back from the local copy at 20:16:32 UTC at 70,793 bytes. In the same second we read the NFL scoreboard the page reads: 175,988 bytes in 0.10 seconds, 16 events — 4 in progress, 8 final, 4 not started. Four carried odds, all credited to DraftKings, with spread, moneyline, over/under and team odds fields, and they were the four not yet started. On a live Sunday, then, three-quarters of the board had no line to show, and EZ Score is built to say so on the card rather than fill the space.

Counting the league table in the code gave 26 entries: the ten named leagues, fourteen European football competitions, and ATP and WTA tennis marked as carrying no odds. The first screen (captured 20:14:41 UTC) still says “Ten leagues”, so the page undersells what its own code now covers.

390 × 844 20:14:47Z

Callouts · phone capture, 390 × 844

1. Callout 1: At 390 pixels the name “EZ SCORE” wraps onto two lines in the header.
2. Callout 2: The lede still says ten leagues at phone width.

Segment 4 · Other channels

Beside ESPN scoreboard and Flashscore

Each comparator was checked on the date shown. We say what each does better before we say where EZ Score goes further.

ESPN scoreboard

Checked 2026-09-27 20:20 UTC

What it is

ESPN’s own scoreboards, with quarter-by-quarter scores, DraftKings spread and total, records, statistical leaders, and links to Gamecast, box score, play-by-play and highlights.

What it does better

The source itself, with box scores, play-by-play, video, editorial and player-level statistics.

Where EZ Score goes further

EZ Score shows the opening line against the current one for spread, moneyline and total, keeps the handicap move apart from the price move, puts 26 competitions on one board, and carries no betting links or ticket offers.

Flashscore

Checked 2026-09-27 20:20 UTC

What it is

Live scores in more than 30 sports and over 2,000 football competitions, with match statistics, odds comparison, line-ups, commentary and apps.

What it does better

Vastly wider coverage and in-match detail, odds compared across books, and native apps with notifications.

Where EZ Score goes further

EZ Score shows line movement in place on each card with the book named from the wire’s own field, and says plainly when a line is missing or a market is off.

Segment 5 · Dropouts

Where it falls short

• One source. Scores, lines and standings all come from ESPN’s wire and the one book it carries, so there is no comparison across sportsbooks.
• Lines appear only where ESPN carries them: 4 of 16 NFL events on our board.
• No box scores, play-by-play, player leaders or highlights.
• It depends on an undocumented ESPN endpoint and has no fallback source.

Segment 6 · End card

The verdict

EZ Score reads one wire with unusual care — the handicap move kept apart from the price, a pulled market never shown as a number. It is neither an odds comparison nor a stats site; it is a scoreboard that treats the market’s view as data and labels it honestly.

Open EZ Score on labs.llc How we tested

Channel 09 · Network · labs.llc/network/

A page of network instruments that labels its own limits — captured mid self-test, counted from its code, and set beside Globalping and Bitwarden’s password generator.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:14:53Z Phone capture 20:14:58Z Publisher LABS, which also makes this tool

CH 09 labs.llc/network/ captured 20:14:53Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow: “Network: probe any domain or IP from your own browser”.
2. Callout 2: “A browser cannot send ICMP, so the ping here is an HTTPS round trip and says so.”
3. Callout 3: The self-test panel, stamped 20:14:53 UTC — the same second our capture file was written.
4. Callout 4: A live reading: cloudflare.com resolved to 104.16.133.229 in 67 ms; an HTTPS round trip to netflix.com took 53 ms.
5. Callout 5: “Global test” — the sweep across six continents.

Segment 1 · On screen

What it is, and who it is for

Network is a console of browser instruments on one page: a probe for any domain or address (DNS, reachability, your public address, the registration record), an “is it you or the internet?” sweep that times HTTPS round trips to sites on six continents, a short-link maker, a password bench with seven recipes, a domain lab, and a live wire of the Hacker News front page.

It is for anyone troubleshooting a connection or needing a strong password or a short link.

Segment 2 · Behind the picture

How it works

Everything runs in your browser (app.js lines 1–15). The probe asks Google’s DNS-over-HTTPS service, times an HTTPS HEAD request, reads your public address from api.ipify.org and its reverse name from dns.google, and fetches the registration record from rdap.org (lines 198–237). The sweep resolves each site and times a request to it from your own connection — 75 sites grouped by continent, or 50 drawn from a pool of 87 — and rates each continent (lines 488–508). The only server endpoints are the site’s own short-link store, which checks that a code is free before saving and lengthens codes only when the short ones are taken (lines 606–630).

Its signature is candour about the medium. The code’s header calls honesty the whole point: a browser cannot send ICMP, trace a route or read another site’s response, so the console measures what it can and says where it cannot — down to labelling its one drawn network-path panel a sketch. The password bench draws from crypto.getRandomValues with rejection sampling, so each character is chosen uniformly, never from Math.random (lines 29–38), and nothing it makes leaves the page. A probe or a sweep can be exported as a PDF.

• dns.google (DNS-over-HTTPS), api.ipify.org, rdap.org and hn.algolia.com — all from your browser
• The target sites themselves, timed from your connection
• labs.llc’s own short-link store: labluxa/check.php and save.php

Segment 3 · The test

What we measured, and when

The page came back from the local copy at 20:17:04 UTC at 126,883 bytes. We asked the short-link store whether the code “zq7x” was free — a read-only check — and it answered {"available":true} in 18 bytes; we saved nothing. Counting the lists in the code: the continental sweep holds 75 sites (25 North America, 10 Europe, 13 Asia, 11 South America, 6 Africa, 10 Oceania) and the random pool 87, of which 50 are drawn per run. The sweep runs only in a browser, so we did not run it.

The capture caught what our command line could not. At 20:14:53 UTC the page’s own self-test ran in the capture browser and printed its results on the first screen — callouts 3 and 4. The phone capture, five seconds later, shows a different reading, openai.com resolving in 49 ms, which is what a live test ought to do.

390 × 844 20:14:58Z

Callouts · phone capture, 390 × 844

1. Callout 1: “What it will not do is dress a guess up as a measurement” — kept in the first screen at 390 pixels.
2. Callout 2: The Ask launcher covers the self-test’s time stamp in this capture.

Segment 4 · Other channels

Beside Globalping and Bitwarden Password Generator

Each comparator was checked on the date shown. We say what each does better before we say where Network goes further.

Globalping (jsDelivr)

Checked 2026-09-27 20:21 UTC

What it is

A free, open-source platform that runs ping, traceroute, DNS, MTR and HTTP measurements from a worldwide network of community-hosted probes.

What it does better

Real measurements from many locations, including the ICMP ping and traceroute a browser cannot send, and you choose where the test runs.

Where Network goes further

Network answers the everyday question — is it my connection? — by timing sites on six continents from your own browser, rated per continent, with no account; and it keeps the probe, password bench, short links and live wire on one page with a PDF export.

Bitwarden Password Generator

Checked 2026-09-27 20:21 UTC

What it is

A free generator for random passwords of 5 to 128 characters and for passphrases, made on your device, with a strength tester.

What it does better

A crack-time strength tester beside the generator, and a password manager to save the result into.

Where Network goes further

Network’s bench offers seven recipes — classic, three words, mnemonic, Markov, pronounceable, leet and grown from your own root word — and its code states the rejection sampling behind each character.

Segment 5 · Dropouts

Where it falls short

• “Global” means sites hosted around the world, not measurements taken from around the world: every timing is from your own browser.
• A no-cors HEAD request succeeds on any HTTP answer, so a site serving an error page still counts as reachable, and the time includes a fresh connection’s DNS and TLS set-up.
• A labelling slip: the PDF report calls the continental sweep “~85 sites”; the list holds 75.
• Short links need the labs.llc server, and cannot be edited or given analytics.

Segment 6 · End card

The verdict

Network labels its own limits in the interface, which is rarer than it should be. It cannot see the internet from anywhere but your desk — and for the question “is it me?”, your desk is exactly where you want to look from.

Open Network on labs.llc How we tested

Channel 10 · OmniFile · labs.llc/omnifile/

A converter that tests your browser before it offers anything — captured, weighed, and set beside CloudConvert and Squoosh.

Tested 2026-09-27 Build 537, local copy Desktop capture 20:15:04Z Phone capture 20:15:09Z Publisher LABS, which also makes this tool

CH 10 labs.llc/omnifile/ captured 20:15:04Z

Callouts · desktop capture, 1280 × 800

1. Callout 1: The eyebrow: “Batch conversion · in the tab · nothing uploaded”.
2. Callout 2: “Reads the first bytes of each one and tells you what it actually is rather than what it is called.”
3. Callout 3: “There is no endpoint on this property to send a file to.”
4. Callout 4: “About ten megabytes of compiled FFmpeg” — the core file we measured is 32,129,114 bytes (see the test).
5. Callout 5: “What this will not do” — the limits, one click from the top.

Segment 1 · On screen

What it is, and who it is for

OmniFile is a batch converter that runs in the browser tab. Drop files, a folder or a ZIP, or paste a direct link; it reads what each file really is from its bytes, then converts images and text through the browser’s Canvas and audio and video through FFmpeg compiled to WebAssembly. Nothing is uploaded.

It is for anyone converting a batch of photos, clips or data files who would rather not send them to an online service.

Segment 2 · Behind the picture

How it works

The code’s first lines set the rule: no file you pick is ever sent anywhere, and there is no endpoint on the property to send one to (app.js lines 1–10). A sniffer names each file from its first bytes — about 45 identities, HEIC variants told apart by their ftyp brand — and says so when a name lies (lines 186–349). Readers parse headers for measured facts, among them EXIF orientation and whether a photo carries GPS tags; the GPS values themselves are never read or printed, and the page warns that a Canvas re-encode destroys them (lines 509–517).

The most unusual step comes before any conversion is offered: OmniFile tests the browser in front of it. It encodes a one-pixel canvas as PNG, JPEG, WebP and AVIF and sniffs what actually came out; it decodes tiny PNG, WebP, AVIF, HEIC and TIFF samples with both of the browser’s decoders, because they have been known to disagree; and it finds the largest canvas edge that really paints (lines 940–1035). FFmpeg — @ffmpeg/ffmpeg 0.12.10 with core 0.12.6 — comes from unpkg.com only when you ask for it, its addresses named on the page first.

• No data sources: files are read from your own disk
• unpkg.com: the FFmpeg class, worker and WebAssembly core, fetched on demand — code coming in, never files going out
• Bundled: jszip.min.js
• A direct link you paste, fetched by the tab from the host that serves it

Segment 3 · The test

What we measured, and when

At 20:17:30 UTC the page came back from the local copy at 85,877 bytes, its script at 275,289 bytes and JSZip at 97,630 bytes. We then fetched the FFmpeg core the page asks for from unpkg: 32,129,114 bytes in 0.40 seconds on our connection, which matches the code’s own warning that “a 32 MB fetch should never be a surprise”.

The first screen (captured 20:15:04 UTC) describes the same engine as “about ten megabytes of compiled FFmpeg” — callout 4. We measured the uncompressed core, not what crosses a compressed connection, so we cannot say which figure a visitor will see on the wire; we can say the page and its code give two different sizes. We did not run a conversion, which needs a real browser.

390 × 844 20:15:09Z

Callouts · phone capture, 390 × 844

1. Callout 1: The “ten megabytes” line again, at 390 pixels.
2. Callout 2: The Ask launcher covers the end of the second paragraph in this capture.

Segment 4 · Other channels

Beside CloudConvert and Squoosh

Each comparator was checked on the date shown. We say what each does better before we say where OmniFile goes further.

CloudConvert

Checked 2026-09-27 20:21 UTC

What it is

An online converter for 212 formats across 11 categories — documents, images, video, audio, spreadsheets, slides, e-books, archives, vector, CAD and fonts — processed on its servers, with an API.

What it does better

Far more formats, documents, e-books and CAD among them; processing that does not depend on your device; and an API.

Where OmniFile goes further

OmniFile never uploads a file — there is nowhere to upload one to — and it identifies each file from its bytes and tests what your browser can read and write before offering a conversion.

Squoosh (Google Chrome Labs)

Checked 2026-09-27 20:21 UTC · from its GitHub repository page

What it is

An open-source image compression app that works on one image at a time in the browser, across several codecs.

What it does better

Fine-grained codec controls and a side-by-side visual comparison for a single image; open source.

Where OmniFile goes further

OmniFile works in batches — files, folders and ZIPs of up to 20,000 entries — with a ledger of size in, size out and saving per file, handles audio, video and text as well as images, and flags GPS data in photos.

Segment 5 · Dropouts

Where it falls short

• Documents are named but not converted: PDF, Office, OpenDocument, EPUB, RAR, 7-Zip, gzip, fonts and Photoshop files are ones the desk “can name but not convert”. CloudConvert handles 212 formats.
• Audio and video need a 32 MB WebAssembly download from a third-party CDN on first use; if unpkg.com is unreachable, so is every audio or video conversion.
• Heavy video work runs on your machine inside a tab, so speed and size limits depend on your device.
• What it can read depends on your browser: HEIC, AVIF and TIFF decoding is tested per browser and may be missing.

Segment 6 · End card

The verdict

OmniFile measures before it promises: the file’s real type, your browser’s real abilities, every byte in the ledger. It is narrower than CloudConvert and leans on a CDN for its heaviest work, but your files stay exactly where they are.

Open OmniFile on labs.llc How we tested