Finished: version 1, on view now

Antikythera Cosmos

The oldest geared computer we know of, rebuilt from its tooth counts. Sixty-nine bronze wheels, after the 2021 reconstruction by Tony Freeth's team at UCL, turned by one crank in your browser and held against the real sky.

Turn the crank Source on GitHub See my other projects

Fragment A, the largest piece, from a CT scan of the original (Ashkan Pakzad, CC BY 4.0).

What came out of the sea

In 1901 sponge divers working a Roman-era wreck off the island of Antikythera brought up a lump of corroded bronze. It lay unregarded until May 1902, when the archaeologist Valerios Stais saw a gearwheel in it. The ship had gone down around 70 to 60 BC.

What survives is 82 fragments, about a third of the machine, in the National Archaeological Museum in Athens. Thirty gears are still in them: 27 in Fragment A and one each in B, C and D.

The reconstruction laid inside the scan, so the wheels show through the crust.

Reading through the corrosion

In 2005 an X-ray scanner was shipped to Athens and the fragments were scanned slice by slice. Inside the crust were the gears, their teeth countable where the rims survive, and thousands of letters of Greek, many read for the first time since antiquity.

From those counts, and from the numbers inscribed on the front cover, Freeth and his colleagues worked out the rest. Thirty-nine of the 69 wheels here are their proposals, not readings, and the exhibit marks which is which.

Taken apart: every wheel spread along its arbor, still turning.

Exact in the teeth

Every gear is cut from one table of tooth counts, meshes and couplings, and every rate is an exact fraction, checked by tests against the 2021 paper. The Moon's pointer runs through 64 on 38, 48 on 24 and 127 on 32, which multiply out to exactly 254/19: in nineteen years the Moon goes round the zodiac 254 times.

The planets' period relations, as the 2021 model gears them
Planet	Synodic cycles	in years	Where the numbers come from
Mercury	1,513	480	derived by Freeth et al.
Venus	289	462	462 is read on the front cover
Mars	133	284	derived
Jupiter	315	344	derived
Saturn	427	442	442 is read on the front cover

The lunar train on its own, the pin and slot riding on the 223-tooth platform.

The pin and slot

The Moon does not move at an even pace: it hurries near perigee and lags near apogee. At the back of the machine, two 50-tooth wheels sit face to face on axes 1.1 mm apart, a pin on one riding in a slot on the other. As they turn, the slotted wheel speeds up and slows down by about six and a half degrees.

The pair rides on a platform that turns once in about nine years, so the swing follows the slowly turning line of the Moon's orbit.

The back: the Metonic spiral above, the Saros eclipse spiral below, about four years of crank.

Calendars and eclipses

The upper spiral counts the 235 months of the Metonic cycle, after which the Moon's phases fall on the same days of the year: nineteen years. Small dials inside it keep the four-year round of the games and the 76-year Callippic cycle.

The lower spiral is the Saros, 223 months after which eclipses repeat. Fifty-one of its cells carry glyphs, Σ for an eclipse of the Moon and Η for one of the Sun, with the hour: 38 lunar and 28 solar, some cells holding both. The little Exeligmos dial adds 0, 8 or 16 hours, because a Saros runs a third of a day past a whole number.

One crank, one year

A hand crank on the side turns a 48-tooth crown wheel against the 223-tooth main wheel. One turn of the main wheel is one year, and every other pointer is geared off it. Here the front dial runs through the year from 12 May 205 BC, the full moon of the eclipse dial's first month, and stops at each month whose Saros cell carries an eclipse glyph, so the glyph can be held against NASA's canon.

The gold hand with the golden ball is the Sun; the silver hand is the Moon, with its half-dark phase ball riding on it; the small stones on the inner rings are the five planets. Drag the crank below to turn it yourself.

Crank

Checked against the sky

The exhibit keeps a second column beside the machine: where the real Sun, Moon and planets stood that day, from a modern ephemeris, and which eclipses really happened, from NASA's Five Millennium Canon. When the bronze is wrong, it says so.

Eight years of the machine's sky as a long exposure: Venus draws its eight-year pentagram, Mars its loops. Bronze ticks on the rim mark where the machine puts each body, verdigris ticks where the real one stood.

28 of 28: solar glyphs in one Saros from the epoch fall in a month with a solar eclipse somewhere on Earth
27 of 38: lunar glyphs land on a real lunar eclipse; 10 more on one that only grazes the Earth's outer shadow, and one on none
3.5°: the Moon stays within about this of the sky near the epoch, then drifts about a degree every eighteen years

Every comparison with the paper is written up: the first accuracy pass and the second, which found and fixed a Moon pointer running up to 20° off and a broken second epoch.

Where this copy departs from the 2021 paper

Every tooth count, gear train, pin-and-slot offset and pin distance follows Freeth et al. 2021 (Fig. 3 and Supplementary Tables S8 and S9). These parts are this copy's own choices:

A simpler front display. In the paper the true Sun and the date are rings, the Sun's ring carries an Age of the Moon scale and eclipse-limit marks, and the planet rings carry opposition markers and proposed index letters. Here the Sun and the date are plain pointers, and none of those scales, marks or letters are drawn.
The size of the Sun's anomaly. The paper gives the true Sun's gearing (56, 52, 56 and a follower) but not its dimensions. This copy sets the pin for an eccentricity of 1/24, the ancient value that Ptolemy gives from Hipparchus.
How the pointers are set. The paper says each anomaly device was set against observations on chosen dates. This copy sets each pin's starting phase, and each pointer's zero at the epoch, by fitting a modern ephemeris (astronomy-engine), which no ancient maker could have done. The 2021 paper does not choose an epoch; the two offered here come from Carman and Evans and Freeth (both 2014) and from Voulgaris and colleagues (2023).
Geometry, not ratios. The reconstructed wheels are cut at a module of 0.5 where the paper adjusts a few (about 0.53 and 0.54 for the 51 and 44 of the Venus train, 0.521 for its 34 and 26); arbor angles and layer depths are this copy's own; and two surviving lunar wheels, e6 and k2, are cut at a module of 0.534 rather than the measured 0.536 and 0.541 so the train closes on its axis. None of this changes a tooth count or a rate.

Two editions of one exhibit

The exhibit in its default look: the bronze machine in a dark gallery, a column of readings beside it — The museum vitrine: a dark gallery at night, the bronze lit like an exhibit.

The same exhibit in its manuscript look: parchment ground, iron-gall ink and red rubrics, the machine set into the page as a numbered plate — The manuscript: the same machine as a plate in a scholar's copy of Ptolemy.

Both carry every feature: a fourteen-step narrated walkthrough, the gear trains one at a time, the Fragment A scan, two proposed epochs, and a share link for any date you crank to.

How it was made

One gear table
Every tooth count, mesh, coupling and pin offset from the 2021 paper and its supplementary tables goes into a single file. Nothing else defines a gear.
Exact arithmetic
A Python solver turns the table into rates as exact fractions and lays out the arbors; tests hold each planet's period relation to the paper.
Built in Blender
Blender 5.1 cuts every wheel, plate and pointer from the table by script, with dial faces drawn to the millimetre, and exports the machine as glTF.
Driven in the browser
three.js re-derives every rate from rules carried in the model, so the browser turns the machine with the same arithmetic. Time runs as Julian Day on the proleptic Julian calendar, never a JavaScript date.
Checked twice
Two accuracy passes against the paper, the second refereed by a second AI model whose every finding was verified against the sources before anything changed. The narrated walkthrough was then filmed and critiqued, and its staging fixed where the picture missed what the voice described.

Designed and directed by Stewart Gregerson. Built with Claude Code. Narration in a synthetic ElevenLabs library voice.

Sources and credits

Freeth, Higgon, Dacanalis, MacDonald, Georgakopoulou and Wojcik, A Model of the Cosmos in the ancient Greek Antikythera Mechanism, Scientific Reports 11:5821 (2021), CC BY 4.0. The reconstruction this exhibit follows, with the Antikythera Mechanism Research Project's papers of 2006 and 2008.
Freeth, Eclipse Prediction on the Ancient Greek Astronomical Calculating Machine Known as the Antikythera Mechanism, PLOS ONE 9 (2014): the eclipse glyphs and the epoch. Carman and Evans, Archive for History of Exact Sciences 68 (2014): the epoch of 12 May 205 BC. Voulgaris, Mouratidis and Vossinakis, Almagest 14.1 (2023): 22 December 178 BC.
Bitsakis and Jones, The Front Dial and Parapegma Inscriptions, Almagest 7.1 (2016), CC BY-NC 4.0. Iversen, Hesperia 86 (2017): the games and the calendar.
Budiselic et al. 2020 and Woan and Bayley 2024: the calendar ring's holes. Edmunds, Journal for the History of Astronomy 42 (2011), and Szigety and Arenas 2025 (a preprint): the teeth and their errors.
Eclipse Predictions by Fred Espenak and Jean Meeus (NASA's GSFC), the Five Millennium Canon. Positions from astronomy-engine by Don Cross, MIT.

Third-party material in the exhibit
What	Who	Licence
Reconstruction model	Freeth et al., Sci. Rep. 11:5821 (2021)	CC BY 4.0
Fragment A CT scan	Ashkan Pakzad	CC BY 4.0
Parapegma and front dial inscriptions	Bitsakis and Jones (2016)	CC BY-NC 4.0
Moon face	NASA CGI Moon Kit (LROC)	Public domain
Ephemeris (astronomy-engine)	Don Cross	MIT
Rooms (studio_small_09, artist_workshop)	Poly Haven	CC0
Layout reference	Thomas Weibel	CC BY
Narration voice, David (Documentary)	ElevenLabs library	ElevenLabs terms

The machine only knows turns. Give it some.