The first line of OTBSync code was written on July 8, 2026. Open beta launched ten days later, and improvements have shipped near-daily since. This page tracks what's live, what's in testing, and what's coming. Found a bug or want something prioritized? Email hello@OTBSync.com.
OTBSync improves fastest from real games. Opting in to anonymized detection reports (in settings, or on the FAQ page) is what turns a misread at your board into a fix on this page — many of the shipped items below started as a single player's report, and several were fixed the same day.
Today the engine waits for the whole board to go quiet before reading a move, so hovering hands stall detection in fast games. The new per-square approach reads the squares that matter while ignoring motion elsewhere. Bench results are strong (zero wrong commits); now being validated in real blitz games before it's on for everyone.
Rethinking detection to consider the whole board at once, making it more forgiving of hands and shadows lingering in frame — a stepping stone to smoother blitz and, someday, over-the-board bullet.
Support for cameras that move. Staged: first automatic recovery when your camera rig gets bumped, then a steadily-held phone, and eventually continuous tracking for glasses-style cameras.
Training tools on a real board, starting with repertoire drills — physically play your prep lines and get drilled on the branches.
Let stream viewers vote on moves, with legal-move filtering, moderation, and explicit streamer approval before anything touches the physical game.
These are designed and ready, but we only add engine complexity when real games prove it's needed. Your opt-in reports are the evidence.
Automatically re-fit the board when the camera gets knocked, instead of asking you to recalibrate. The measurement side shipped Aug 16; if real sessions show bumps are common, recovery gets built.
Driving the phone camera's exposure directly when overhead lighting washes out a glossy board, recovering detail that filters can't.
Telling a drifting shadow apart from a hand so the engine spends less time waiting for a "covered" board that's actually clear.
Low camera angles make castling hard to corroborate; a targeted loosening is scoped but waits for a careful safety pass so it can never confirm a wrong move.
Phone pairing works across normal networks today; if beta shows real-world networks that block direct connections, a managed relay gets added.
Dates are when each item went live. Zero to open beta took ten days; everything since has shipped into a live product without breaking play.
The core of OTBSync, playable on day one: point a camera at your board, move the pieces, and your moves play on Lichess through the official Board API — calibration, per-square detection, and legal-move matching included.
Use your phone as the board camera, streamed directly to your computer over an encrypted peer-to-peer connection — no app install.
Spoken move announcements and prompts, a wooden "thock" on each commit, and optional downloadable neural voices that run entirely on your device — the backbone of phone-only, hands-free play.
Bot games at eight skill levels — and since Jul 14 they run on Stockfish inside your own browser, so bot games work offline and stay completely private.
Direct peer-to-peer games over a share link — no Lichess account needed on either side.
Analyze on a real board with an eval bar and best-move arrows, no account required. Finished games open straight into analysis, with PGN/FEN import.
Texture and lean-aware detection from day one, then dynamic color separation — so low-contrast, decorative, and non-black/white sets track, not just tournament boards.
Public launch ten days after the first commit: FAQ, terms and privacy, opt-in anonymized detection reports, and a per-game detection accuracy dashboard with a "flawless games" stat.
A settled hand shadow could commit a move nobody played. Every auto-commit now checks that the board actually changed — if the origin still shows your piece or the destination still reads empty, it asks instead of guessing.
Settling shadows on pale boards could impersonate a move convincingly enough to fool the appearance readers. Texture polarity — a real move brightens its destination — now catches it, and the engine asks instead of committing.
A castle caught mid-move is pixel-identical to a rook move (and vice versa). Both directions of the confusion now produce a quick question instead of a wrong commit.
Captured pieces or a clock at the board edge could keep detection waiting for the board to "clear" forever. The engine now recognizes an obstruction that never moves and plays on, with an escape hatch if the board stays covered.
Your arm crossing the board could fake a pawn double-push. The engine now uses a law of chess against it: a pawn can't pass through an occupied square, so the phantom is refuted and the real move is offered instead.
Built from open beta's very first day of reports: three guards for frames where the evidence isn't where the move is — a drifted calibration, steep camera lean displacing a move onto neighboring squares, and a tall piece's own image confirming a move into its shadow. All three now ask instead of committing.
A retreating arm could leave evidence on a square that made a wrong move look right. When a second square contradicts the read, the engine now waits for a clean frame rather than committing through the mess.
With auto-queen enabled, a noisy frame could still misread a promotion as ambiguous and stall the game. Promotions now resolve cleanly.
Castling caught mid-move could still be misread as the king merely stepping aside — at that instant the two are nearly indistinguishable on camera, and the squares that could tell them apart are exactly the ones your hand is over. Like July's rook-side guard, the engine now asks the quick question instead of committing. Caught and fixed from a single player report, same day.
A zoom slider on the calibrate and phone-camera screens, for cameras that support it.
Exposure, contrast, and color balance are auto-tuned at game start and adopted only when measurement shows they help — low-contrast and white-on-white boards benefit most.
Correctly placed corners could visibly jump when you hit confirm, and play-time drift correction could slowly walk the grid off the board. Every automatic adjustment is now strictly bounded and must prove it fits better than your placement.
Starting a game with a stale calibration (rig moved between games) taught the engine the tabletop as "the start position." The first frame is now verified against the start position — and the board is re-found automatically if it doesn't match.
A well-tracking grid used to suppress all warnings — including glare, which destroys square readability without touching geometry. Readability warnings now always surface once, while a tracking board is still never blocked.
A settings button that wipes what the engine has learned about your board and pieces — the recovery path if it ever learned through a bad calibration.
A calibration set up backwards or mirrored used to silently poison move reading (the start position looks fine either way). The engine now proves the board's orientation from piece colors and square colors at game start, refuses to learn through a bad map, and — as of Aug 14 — automatically flips a mirrored calibration to the correct one when it can prove it.
The engine can now notice mid-session camera disturbance in its diagnostics — the groundwork (and evidence gate) for automatic bump recovery.
A camera aimed square-on to a board edge could silently lose its 3D understanding of piece lean during calibration — and then need to ask about nearly every move for the whole session. The geometry solve now handles that angle, and in the rare case it genuinely can't be recovered, the app says so at calibration instead of staying quiet.
A brief Wi-Fi blip no longer leaves a stale "reconnecting" banner up after the camera self-heals.
A blocked network used to show "Phone connected!" over a board that never appeared, then blame camera permissions. Pairing now waits for a real connection and tells you exactly which side — network path or camera — actually failed.
Stream your board to Twitch/YouTube: a studio with scene presets, synchronized clean outputs, clocks, player plates, optional face cam, and simultaneous multi-platform setup. Free during open beta. Your video stays peer-to-peer — it never passes through our servers.
Four base presets with an optional face cam (eight scenes total), color-blocked player plates, mascot moment animations, and opt-in eval bar / best-move hints for bot games and analysis — never engine help in live PvP.
The recommended capture path: your camera and game state flow peer-to-peer straight into OBS, immune to the browser-window blackouts that plague window capture. Video never touches our servers — only the connection handshake is relayed.
Browser Source links work in streaming apps' web widgets, bringing broadcasting to iPad setups.
A silent board-cam tile now says what's actually wrong — no camera, no publisher, or an unreachable connection — instead of one generic message for all three.
Switching layouts or scene presets in the studio now updates live outputs — including OBS Browser Sources, which live in an isolated browser and previously never heard about changes — without reloading or flashing black on stream.
Ready-made color-and-piece presets plus a custom editor for board, interface, and accent colors — live preview, accessible contrast, and your look carried through play, OTB Analysis, and broadcasts. Free during open beta.
A silent engine failure used to be able to stall a bot game forever; it now recovers with a fallback instead.
Layout passes across iPad and iPad Mini portrait/landscape so the full experience — landing, pairing, live game controls — works on tablet.
An honest accuracy score after every game: auto-detected moves, assists, and resyncs counted so a misread can never raise the score.
The mascot reacts to your game — a nod on a move, a chomp on a capture, a barrel roll on castling, a gold spin on promotion — respectful of reduced-motion settings.
The "what move did you make?" voice line now speaks once per position instead of stuttering while you think.
The correction prompt now asks what the pieces physically show — not what you meant to play — so reports diagnose the engine instead of confusing it.
Leaving analysis could strand a half-blank lobby, and recalibrating mid-session quietly threw away your move list. The lobby now restores properly, and recalibration returns you to your analysis exactly where you left off — moves, review, and side lines intact.
An app-wide pass on edge cases where two things happen at once — a game ending at the exact moment a move or takeback is in flight, or an undo while the promotion chooser is open — so they all resolve gracefully instead of erroring.