spire-mentor reimplements Slay the Spire 2 from scratch, checks itself against the game's own code after every action, and runs fast enough to search thousands of futures per decision, with a run-level AI on top. Two front ends show that AI thinking: a browser workbench and a mod for the real game.
Six pieces, each usable on its own:
Rust, deterministic, the whole game: five characters, ~580 cards, ~300 relics, three acts. Built for speed (1.3 M actions a second on one core, 31 M on a 24-core server; about 1,500× the shipped game's throughput per core), checked against the game's own code after every action, and usable from Python.
Runs the game's own assembly headless and checks the engine against it after every single action, plus a fuzzer that hunts for divergences.
Chance-node MCTS for fights (it never sees the draw order); a PPO-trained transformer per character for every other decision. Ironclad wins 90% at A0 and 39% at A9; Defect 22% at A10.
The game in your browser, on any platform: rewind and branch, an AI coach, autobattle, post-run blunder review, the MCTS tree, state import from the real game, and SimLab.
A mod that puts the AI inside the real game as a coach, an autopilot and a rewind button.
just import-game decompiles your copy and extracts data, art and sound; a generator builds a static wiki.
sts2-rs re-implements the game's rules in Rust, pinned to game
build v0.107.1. It is not an approximation: the random number
generator is a hand port of .NET's System.Random (the 56-element
subtractive generator the game uses), and every RNG stream is consumed in the
game's order. A seed yields the same map, the same card rewards, the same
shuffles and the same monster rolls as the real game.
| Kind | Count | Notes |
|---|---|---|
| Characters | 5 | Ironclad, Silent, Defect, Regent, Necrobinder |
| Cards | 580 | card ids incl. colorless, status and curse cards |
| Relics | 297 | |
| Potions | 63 | |
| Powers | 272 | |
| Monsters | 121 | with their intent state machines |
| Events | 69 | incl. the Ancients |
| Acts | 3 | 4 act models (act 1 rolls Overgrowth or Underdocks) |
Full benchmarks → one core against the whole machine, the shipped game compared five ways, search and training speed, the hardware, and two other Slay the Spire 2 simulators.
| Measurement | Result |
|---|---|
| Random-play fights per second, one core | 59,400 |
| Engine actions per second, same fights, one core | 1.30 M (766 ns each) |
| The same on the whole machine (24 cores, 48 threads) | 1.41 M fights/s 31.0 M actions/s |
| Against the shipped game run headless, best measured throughput | 1,500× per core 2,430× whole machine |
| Against the shipped game, CPU time per action | 2,970× (2,250× its game thread) |
| Against the previous engine (12 Oct 2026), same fights | 2.25× |
| Against two other simulators, same random-play fights, one core | 3.13× faster than sansb sts2core 1.56× faster |
| Browser hint's search, 1,000 simulations | 13.7 ms native 18.4 ms WASM in Chromium |
| Browser hint at its default 2,000 simulations, WASM in Chromium | 39.9 ms |
| Advisor's search on an Act-1 boss, 8,000 simulations | 275 ms on 1 thread 39.6 ms on 8 |
| A typical PPO training iteration (Ironclad, 960 games) | 208 s (609 s before) 150 in ~9 h |
Measured on 14 October 2026 on one AMD EPYC 7443P server (24 cores, 48 threads, Linux) on a quiet host; plain release builds, except the training row, which uses the profile-guided build training runs on; medians of interleaved runs. The fights are random play: Ironclad's starter deck and Burning Blood against each of the 22 Act-1 Overgrowth encounters in turn, 22.0 actions per fight on average (an action is a card play or an end of turn). One core means one process on one physical core with its second hardware thread idle. The shipped game is its own assembly run headless by the project's harness on the same fights. One game process spends most of its time asleep in the harness's 1 ms polling, so the fair comparison runs many game processes at once, the best counts measured: 16 on one core, 192 on the machine. Earlier versions of this page said 23,000×: that was the wall time of one cold game process per fight, a different measurement, explained on the benchmarks page. Both other simulators are combat only. sansb/sts2-sim (game build v0.111.0) is certified against checksums the real game writes, and its review solver searches the true state by design; the 3.13× has our relic switched off, as Burning Blood is inert there. lain-wood/sts2core is effectively Ironclad only, with partial content and its own RNG, and its live solver is given the true draw-pile order. The comparison is random play, with no search involved (details). Method: 2026-10-14-benchmarks.md.
A cargo workspace of twelve crates: sts2-core (state, hooks,
damage, powers, RNG), sts2-content (every card, relic, potion,
monster and event), sts2-mcts, sts2-runai (run
arena, seed panels, policy/value nets with ONNX inference),
sts2-cli (REPL, replays, sweeps), sts2-wire,
sts2-py (PyO3 bindings), sts2-bench,
sts2-workbench + sts2-workbench-wasm, and helpers.
7,000+ Rust tests, plus a cargo-fuzz target on the combat step.
Claims of accuracy are cheap, so the project checks them mechanically. The game's own C# assembly, taken from your copy, is IL-patched and run headless with Godot stubbed out. The engine and that oracle then play the same seed with the same actions, in lockstep:
Details: parity-suite.md.
Every combat decision is an MCTS search over the engine: 8,000 simulations per decision in the real game and the published win rates, 1,000 (4,000 at bosses) in training gates. Draws and monster rolls are chance nodes, not clairvoyance; leaves are scored by a 216-feature evaluation.
Paths, card rewards, removals, upgrades, transforms, enchantments, events, shops, rest sites and relic choices come from a transformer policy per character, trained with PPO by self-play over full runs in the simulator and served from Rust through ONNX.
Read more: How the AI plays (search, hidden information, the network) and How it's trained (PPO, compute, the promotion gate, what didn't work).
| Character | Win rate | Promotion gate (vs previous champion) |
|---|---|---|
| Defect | 76.1% | +5.5 pp, p = 2e-6 |
| Ironclad | 74.9% | +5.2 pp, p = 1e-6 |
| Silent | 59.9% | +5.6 pp, p = 7e-7, all 11 ascension levels up |
| Regent | 53.8% | +4.6 pp, p = 1e-4 |
| Necrobinder | 51.7% | +6.8 pp, p = 6e-10, all 11 levels up |
| Character | A0 | A1 | A2 | A3 | A4 | A5 | A6 | A7 | A8 | A9 | A10 | Mean |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Defect | 93 | 90 | 92 | 89 | 90.5 | 85 | 87.5 | 79 | 70 | 39.2 | 22 | 76.1% |
| Ironclad | 89.5 | 95 | 92.5 | 90 | 86.5 | 85 | 86 | 79 | 70 | 39 | 11 | 74.9% |
| Silent | 78.5 | 80 | 77 | 74 | 67.5 | 64 | 65.5 | 68 | 55.5 | 20 | 9 | 59.9% |
| Regent | 67.5 | 77.5 | 73 | 72 | 66 | 59.5 | 60 | 51 | 43.5 | 15.6 | 6.2 | 53.8% |
| Necrobinder | 66.5 | 73 | 73 | 68.5 | 62.5 | 59.5 | 52.5 | 49 | 44.5 | 13.8 | 5.6 | 51.7% |
One fresh seed panel for all five champions, measured
October 2026: 200 runs per cell at A0–A8 and 500 at A9–A10, the same seeds
for every character, Act 1 rolled between Overgrowth and Underdocks. Combat
search is honest (it does not see the draw pile order) at 8,000 simulations
per decision, boss fights included, the advisor's budget. A cell's 95%
interval is roughly ±6 pp in the middle of the range and ±2–4 pp at A9–A10.
The cliff lines up with the game's own steps: A9 (Deadly Enemies) makes
every enemy hit harder, and A10 (Double Boss) adds a second boss fight at
the end of the run. Full write-up:
docs/experiments/2026-10-11-release-winrates.md.
Every champion is promoted on a pre-registered gate: 200
seeds × 11 ascension levels = 2,200 paired runs against the incumbent,
McNemar test, seed-clustered error check, and a guard that no single level
may regress by more than 5 pp. Promotion records and provenance live next
to the nets in sts2-pycli/data/ppo/.
Driven through the Spire Mentor mod against the actual
SlayTheSpire2.exe, random seeds, no cheats, 40 runs each at
Ascension 8, with 8,000 simulations per combat decision.
These runs used the champions deployed in August 2026; the current ones,
above, have not been measured in the real game yet.
| Character | Wins | Win rate | 95% CI |
|---|---|---|---|
| Ironclad | 13 / 40 | 32.5% | 20–48% |
| Silent | 11 / 40 | 27.5% | 16–43% |
The engine compiled to WebAssembly, running in a Web Worker behind a
React UI styled after the game. No server: saves live in IndexedDB and the
site installs as an offline app. Without the game it runs with the real
card text and placeholder art; just import-game fills in the
real art and sound from your copy.
Run it: just workbench-local builds the site into
sts2-pycli/web/dist and serves it on http://127.0.0.1:4173.
With real art the build contains the game's images and sound, extracted
from your copy for your own use, so please run it locally and don't host it
publicly. Architecture: browser-workbench.md.
A C# mod for Slay the Spire 2 that connects the running game to the AI stack over a local TCP socket. Installed but idle, it leaves the game vanilla. The advisor backend is launched by the mod and ships as a self-contained executable.
Play with it: subscribe on the Steam Workshop (Windows; Steam installs BaseLib with it). The workbench comes bundled and follows your run in the browser.
You play. Ranked moves with MCTS statistics and the principal line, card-reward and shop scores, map routes, a win estimate.
Let the AI take one decision or play until you stop it; autopilot plays whole runs unattended and never falls back to random play.
Every room is snapshotted with its RNG state. Undo a turn, restart a room, or go back to any earlier room, bit-exact.
Open the current state in the workbench, browse the live game's tree, and rewind the real game from the page.
| Key | Action |
|---|---|
| F1 | Help panel: mode, keys, file locations |
| F3 | Action log |
| F4 | Advisor overlay on/off |
| F6 | AI takes this one decision |
| F7 | AI plays until you stop it |
| F8 | Advisor mode on/off |
| F9 / F10 | Game speed − / + |
| Ctrl+Z | Undo to the start of the combat turn |
| Ctrl+Shift+Z | Restart the room |
Every key is also on the in-game toolbar, and the settings live on a mod-configuration page in the main menu.
Players: subscribe on the Steam Workshop.
To build it yourself: Windows, the game at v0.107.1, the .NET 9 SDK and
BaseLib in the game's
mods/ folder. Then cd spire-mentor-mod && dotnet build
builds the mod and installs it into mods/SpireMentor/. Configuration,
the console commands and the wire protocol are in
spire-mentor-mod/README.md
and PROTOCOL.md.
Three tiers; each needs a little more than the last.
No game needed. Rust 1.96, cargo-nextest, just 1.38+.
just rs-check
just rs-test # cargo nextest run --workspace
just rs-doc-testAdds your game install, the .NET 9 SDK and uv.
just import-game "/path/to/Slay the Spire 2" --tier oracle
cd sts2-pycli && uv sync --dev
just build-cs && uv run pytest testsAdds ilspycmd, Node 20+, the wasm32-unknown-unknown target and wasm-bindgen-cli 0.2.120.
just import-game "/path/to/Slay the Spire 2"
just workbench-local # → http://127.0.0.1:4173Use cargo nextest, never bare cargo test: tests
rely on per-process isolation. Full prerequisites:
bring-your-own-game.md;
Windows: the engine tier runs from PowerShell, the rest under WSL2
(BUILDING-WINDOWS.md).
| Path | What it is |
|---|---|
sts2-rs/ | Rust: engine, content, MCTS, run-AI, CLI, PyO3 bindings, workbench engine (WASM) |
sts2-pycli/ | Python, C# and TypeScript: the headless C# oracle, parity suite, PPO training, the workbench web app, the real-game advisor, an MCP server so an LLM can play |
spire-mentor-mod/ | C#: the Spire Mentor mod |
sts2-wiki/ | Python / Astro: decompile, extract, generate a static wiki |
scripts/import_game.py | just import-game: every game-derived input from your install |
docs/ | Reference docs, ADRs, this site (docs/site/) |
Start with ARCHITECTURE.md, then CONTRIBUTING.md. Training a run policy? Read the champion training playbook first.
Slay the Spire 2 is © Mega Crit. This project is not affiliated with or endorsed by Mega Crit. The repository carries the game's text and extracted data so the engine, the AI and the workbench work out of the box, but no art, audio, game assemblies or decompiled code. The tooling extracts those from a copy of the game you own, into git-ignored paths. The screenshots on this page show the game's art.
The original work is MIT-licensed. sts2-pycli began as a
fork of wuhao21/sts2-cli;
Spire Mentor started as a port of zhiyue's STS2BridgeMod from
zhiyue/sts2-rl-agent.
See NOTICE and
PROVENANCE.md.