Making KataGo Teach Instead of Just Crush You

What it does

KaTrain is a desktop Go client built around KataGo. It reviews finished games to flag the most expensive mistakes in points, and lets you play teaching games where the engine can automatically undo your sufficiently bad moves. The app packages its own KataGo binary for Windows, Linux, and older Intel Macs, so it runs locally without needing a separate engine install.

The interesting bit

The project’s real work is pedagogical translation: it turns KataGo’s raw analysis into visual feedback—colored and sized dots on the board that show not just how much a move cost, but whether the opponent actually punished the error. Beyond full-strength play, it offers a zoo of calibrated and stylized opponents, from a rank-calibrated bot tested against GnuGo and Pachi to experimental personalities like “KataJigo,” which tries to win by exactly half a point.

Key highlights

• Visual feedback encodes both mistake severity (color) and whether it was punished (dot size), so you know which blunders actually mattered.
• “Calibrated Rank Bot” is tuned against weaker engines to provide a balanced game without serious blunders across all phases.
• Includes style variants like Simple, Local, Tenuki, Influential, and Territory that bias move selection to simulate different strategic approaches.
• Teaching mode can auto-undo moves that cross a quality threshold, turning review into real-time correction.
• Ships with a working OpenCL KataGo binary and a default model, though you can swap in newer models or CPU/GPU backends.

Caveats

• The bundled KataGo binary and model are described as “rather old,” so out-of-the-box strength is not state-of-the-art.
• Pre-packaged macOS support explicitly excludes Apple Silicon; M1/M2 users will need to configure their own binary.
• The “Undo” button in AI games reverts both your move and the AI’s response, which may feel counterintuitive.

Verdict

Worth a look if you’re a Go player who wants structured feedback from superhuman analysis without drowning in raw engine output. Pure KataGo command-line users already comfortable with SGF parsing and GTP won’t find much new here.