Any screen, one paste
Chess sites, apps, streams, chat threads. Dark or light boards, any piece set, arrows and highlights ignored.
8/5pk1/6p1/8/3P4/6P1/5PK1/8 w - - 0 1
Open the appa1 — Position recognitionEvery board,
Every board,
read exactly.
Screenshots, 1920s book scans, YouTube frames. Drop in an image and get the exact position in a fraction of a second, on your device.
100%
lichess boards99.3%
board libraries92.3%
noisy screenshotsboard · 0.98
Resultwhite to move1 square unsure
r1bq1rk1/pp2bppp/2n1pn2/3p4/2PP4/2N2N2/PP2BPPP/R2QKB1R w KQ - 4 9
b1 — How it works
Three steps. About a second.
The model looks at each of the 64 squares, repairs anything that breaks the rules of chess, and tells you what it is unsure about.
- 1
Drop in an image
Paste a screenshot, drag a photo of a book page, or open a PDF. Boards are found automatically, several per image if needed.
⌘Vpaste to scan - 2
It reads all 64 squares
Two small neural nets vote on every square. Chess rules clean up the result: one king per side, no pawns on the back rank.
rbqrkppbpppnpnpPPNNPPBPPPRQKBR - 3
Check it, then use it
Doubtful squares are flagged red instead of guessed. Fix them in the board editor, then copy the FEN or open the position in lichess.
d6 · 41%legal positionr1bq1rk1/pp2bppp/2n1pn2/3p4/2PP4/2N2N2/PP2BPPP/R2QKB1R w KQ - 4 9
c1 — What people read
Four places a board hides
The same recognizer behind all of them. It is tuned on the hardest case, old printed books, so screens are easy.
Hatched squares, read
Outline pieces, dense hatching, yellowed pages. The hardest case: 95–98% exact on familiar printing styles, 65% across 844 held-out diagrams from 11 books, and unsure squares are flagged for a person. Pro turns a whole PDF into PGN.
87654321
abcdefgh
Videos become drills
Paste a link. Stable frames become positions, positions become a PGN with timestamps, and the moves you miss come back as spaced-repetition cards.
frames→positions→PGN→drills
Included in ProOne HTTP call
Send an image, get the FEN, a confidence for every square and every board found. 500 requests a month are free.
curl https://api.chessglance.com/v1/recognize \ -H "X-API-Key: $KEY" \ -F image=@board.pngRead the API docs
d1 — Benchmark
Exact boards, not close ones
A board counts only when all 64 squares are right. We run the same crops through ten tools, ours included.
100%
lichess, 1,000 boards99.3%
board libraries92.3%
noise benchmark84.1%
held-out 1920s books (107)PreliminaryFull table and method
OursNext best tool
e1 — Privacy
Runs on your device.
The recognizer is a small model your browser downloads once. Free scans never upload the image, so there is nothing for us to store, sell or lose.
Images reach our servers only for features that need one: the public API, PDF books and video to trainer. API images are deleted as soon as the request finishes. Scans of live games are the exception: they are flagged and logged for platform fair-play teams. Read the policy.
- 8 MBThe whole model: two small CNNs, downloaded once and cached.
- ~200 msPer board on a laptop CPU. No GPU, no queue, no round trip.
- 0 uploadsFor free scans. Your screenshot stays in the tab.
f1 — Pricing
Scanning is free. Books and video are Pro.
Free
$0
Unlimited on-device scans
Screenshots, book photos and video frames, with uncertain squares flagged. No account needed.
Open the appPro
Books + video$39per year
or $4.99 a month
PDF books to PGN, YouTube to trainer, spaced repetition and cloud sync.
See ProAPI
from$29per month
500 requests a month free
One endpoint, FEN plus per-square confidence. Plans for apps, platforms and publishers.
See API plansg1 — Questions
Answers, with numbers
How accurate is it?
On our test sets it reads 100% of lichess boards exactly (1,227 boards), 99.3% of boards drawn by common board libraries (600), and 92.3% of boards in a noise benchmark (220). Old printed books are the hardest case: across 844 held-out diagrams from 11 books (1846–1929) it reads 65% of boards exactly and 98% of squares, from 95–98% on familiar printing styles down to near zero on unusual piece glyphs. Exact means all 64 squares are right, and unsure squares are flagged. The benchmark page has the full table and the method.
Do my images leave my device?
Not for free scans. The recognizer is an 8 MB model that runs in your browser, in about 200 ms per board on a laptop CPU. Images go to our servers only for features that need them: the public API, PDF book conversion and video to trainer.
What does it not read from an image?
Side to move, castling rights, en-passant square and move counters cannot be read from pixels. You set them yourself; the default is "w - - 0 1". Orientation (which side is at the bottom) is detected automatically, 97.9% correct on held-out real games, and flagged when the margin is thin so you can flip it in one tap.
What kinds of boards work?
Top-down 2D boards: app and website screenshots, printed or scanned book diagrams, and video frames, including images that contain several boards. Photos of 3D boards at an angle are out of scope. Boards under about 20 pixels per square are the weakest case.
Can I use it during a live game?
No. Our terms forbid it. Recognition works on any page, but scans of live games are flagged and logged, and the log may be shared with the platform, so their fair-play teams can act on it. Engine analysis of a flagged position is withheld until the game is over.
Is it free?
Scanning on your device is free and unlimited. Pro adds PDF book conversion, video to trainer, spaced repetition and cloud sync. The API has a free tier of 500 requests a month; paid plans start at the price on the pricing page.
Can I use it in my own app?
Yes. The API takes an image and returns the FEN, a confidence for every square and every board it found, in one HTTP call. See the API docs for the request, the response and the limits.
h1 — Start
readyRead your
Read your
first board.
No account, no upload, no card. Paste a screenshot and see the FEN.