Hagrid and Snitch were written between 2001 and 2006 – from a first buggy 1800 Elo prototype to an engine playing around 2350. Twenty years later, Snitch 2.0 learned to evaluate with a neural network. Both are free and run in any UCI or WinBoard GUI – and the game on the board is a real one.
Two independent engines with very different internals. Hagrid came first and is fast; Snitch is a complete rewrite – slower, but stronger. Both speak UCI and WinBoard and ship without a GUI. Speeds are measured on an Intel Core i7-12700T.
A complete rewrite with piece lists, table-based move generation, attack-based king safety, evaluation-based pruning and book learning. Version 2.0 replaces the hand-written evaluation with an NNUE network and runs natively on Windows, Linux and ARM.
The first engine. Bitboards and attack boards, Nalimov tablebase support, UCI and WinBoard pondering. Fast – and the one that started it all.
Five years of evenings, bugs and small victories – newest first.
Twenty years after 1.6.2, Snitch evaluates positions with an NNUE network instead of hand-tuned terms. Native 64-bit builds for Windows and Linux (AVX2, plus a POPCNT fallback for very old CPUs) and for Linux on ARM64, e.g. Raspberry Pi 4/5.
This site gets its first redesign since 2002. Both engines still play: the replay at the top is Hagrid 0.7.56 against Snitch 1.6.1, played on 25 October 2026 at 10 s + 0.1 s.
The previous 64-bit compile was accidentally a debug build. No version number change.
WinBoard protocol fixes only – no evaluation or search changes.
MultiPV mode added, several minor search and evaluation tunings, general code cleanup.
Fixes a time control bug introduced in 1.2.x that could make Snitch forfeit on time.
Minor bug fixes and evaluation changes, general speedup.
Fixed a bug that sometimes made Snitch hang after the end of a game. Better move ordering, changed quiescence pruning.
Removed super-verbose UCI info output that slowed Snitch down under some UCI GUIs.
A completely new engine – slower but stronger than Hagrid. See the internals for a detailed comparison.
Minor bug fixes and evaluation tuning – the final public Hagrid.
UCI fix: Hagrid pondered only with the latest Fritz interface. Now pondering works with other UCI GUIs as well.
WinBoard pondering added, evaluation tuning, several minor bugs removed.
Nalimov tablebase support (compressed or uncompressed), pawn hashing bug fixed, evaluation tuning – plays much stronger than before.
Fixed a crash with incremental time controls under WinBoard. New or fixed WinBoard features: move now, analyze, draw claims on threefold repetition and the 50-move rule, resigning.
A big opening book for Hagrid becomes available.
Fixed a bug in the hashtable update and replaced a broken book file.
…to make Hagrid available to the public.
The UCI protocol is implemented so Hagrid runs under the Fritz GUI. By the end of March it plays somewhere between 2100 and 2200 Elo.
Mostly WinBoard features – the only real improvement is null-move pruning, worth several Elo. After months of developing a nameless engine it needs a name: Hagrid is born.
Full of bugs – playing around 1800 Elo, or crashing. Weeks of bug hunting later, by the end of 2001, a WinBoard engine plays reasonable chess at about 2000 Elo.
A first idea – keeping the complete game tree in memory – does not go well. A few days of reading about how it is really done later, the first own engine is under way.
A short overview of the techniques used in my engines, rated by their impact on playing strength – based on my own experience. Snitch 2.0.0 keeps the search of 1.x but replaces the hand-written evaluation with an NNUE network.
| Technique | Hagrid 0.7.56 | Snitch 1.0.0 | Snitch 2.0.0 | Importance |
|---|---|---|---|---|
| Search | ||||
| Alpha-beta / PVS | yes | yes | yes | |
| Hashtable | single (depth) | double (depth, always) | double (depth, always) | |
| Null move | yes (R=2) | yes (R=3,2) | yes (R=3,2) | |
| Aspiration search | yes (sophisticated) | yes | yes | |
| Quiescence pruning | yes | yes | yes | |
| Special eval-based pruning | no | yes | yes | |
| Futility pruning | yes | yes | yes | |
| Lazy eval | yes | yes | no | |
| Fail-high reductions | yes | kind of | kind of | |
| Endgame tablebases | yes | yes | yes | |
| Opening book | yes | yes | yes | |
| Book learning | no | yes | yes | |
| Checks in quiescence | no | yes | yes | |
| SEE | sometimes | no | no | |
| Pawn hash | yes (basic) | yes (basic) | no | |
| Singular extensions | no | no | no | negative |
| Move ordering | ||||
| MVV/LVA | yes | yes | yes | |
| Killer moves | yes | yes | yes | |
| History heuristic | yes | yes | yes | |
| Knowledge-based | no | yes | yes | |
| SEE-based ordering | no | no | no | |
| Evaluation | ||||
| NNUE (768 → 2×256 → 1, AVX2) | no | no | yes | |
| Material | yes | yes | yes (NNUE) | |
| Square control | no | yes | no (NNUE) | |
| Pawns | yes | yes | no (NNUE) | |
| King safety | yes (basic) | yes (attack-based) | no (NNUE; king-danger proxy for search) | |
| Piece-square tables | yes | yes | no (NNUE) | |
| Rooks | yes | yes | no (NNUE) | |
| Bishops | yes | yes | no (NNUE) | |
| Endgame knowledge | basic | basic | basic | |
| Board representation & move generation | ||||
| Piece lists | yes | yes | yes | |
| Bitboards | yes | no | no | |
| Attack tables | no | yes | yes | |
| Move generator based on | bitboards, attack boards | piece lists, tables | piece lists, tables | |
| Move undo | yes | no | no | |
| Attack boards | yes | no | no | |
| Plain 8×8 board | no | yes | yes | |
| Generated moves | pseudo-legal | legal | legal | |
| Result | ||||
| Speed (Athlon XP 1700) | 700 kn/s | 300 kn/s | – | |
| Speed (Core i7-12700T) | 2.5 Mn/s | 1.75 Mn/s (1.6.2) | 1.91 Mn/s | |
| Estimated Elo | 2250 | 2350 | +127 vs. 1.6.2 (measured) | |
Importance: no measurable gain … essential; “negative” means it did not pay off.
Engines only – no GUI included. Load them into any UCI or WinBoard interface and set the working directory to the unpacked folder. On x64, use the AVX2 build – it runs on any Intel Haswell / AMD Zen or newer, i.e. practically every PC from the last ten years. The slower POPCNT build is only meant for very old CPUs without AVX2.
NNUE. AVX2 build, plus POPCNT for very old CPUs.
NNUE. AVX2 build, plus POPCNT for very old CPUs.
NNUE. Raspberry Pi 4/5 or other ARM64 with a 64-bit OS.
The original engine.