GameMindby JE Horizon
Features Adapters Download

Live coaching — while you play.

GameMind watches your screen in real time, narrates what's happening, surfaces the right knowledge, and (when you opt in) emits a single user-confirmed action via OS input. No game memory. No injection. No packets. No anti-cheat interaction.

The architecture in one sentence

GameMind is a Windows desktop app that captures your screen read-only via the standard OS APIs (Windows.Graphics.Capture / DXGI Desktop Duplication), runs a fast perception lane on your CPU to track HUD / minimap / inventory state continuously, and reasons over that structured state with a heavy lane only at stable decision points — not every frame.

By default the app is in OBSERVE mode: it coaches you in overlay and voice, retrieves the right wiki page or patch note when relevant, never touches the game. ASSIST mode is opt-in per title, available only where the title's risk profile is green, and emits single-shot user-confirmed actions through the OS-level SendInput path — same way any accessibility tool sends keys.

Four design choices

🛡️

Anti-cheat safe by construction

No DLL injection. No memory read or write. No packet interception. No kernel driver. No client patching. No multi-input broadcasting.

⚡

Sparse heavy lane

Cheap perception runs continuously on your CPU. The heavy reasoning lane only thinks at stable decision points — keeps latency tight and cost low.

🔒

Local-first, fenced

Capture lives in a short-lived encrypted ring buffer. Your match never leaves the device. Cloud is opt-in for bug reports only.

⚖️

The app owns risk; the model owns judgment

The model proposes. The app enforces policy / rate limit / cooldown / patch-staleness / user confirmation. Same discipline as the trading stack we ship.

Two modes: OBSERVE (default) and ASSIST (opt-in, gated)

OBSERVE mode ships on by default in every adapter. It reads your screen, structures the game state, retrieves grounded knowledge (cited to official sources with patch-version anchors), and surfaces coaching through overlay and voice. It never touches the game.

ASSIST mode is opt-in per title, available only where the title's policy tier is green. When you enable it, GameMind may propose a single concrete action — for example, "sort inventory" or "confirm this dialogue choice" — and you confirm or cancel it. The action then goes out as one SendInput call, the same channel an on-screen keyboard uses.

An ASSIST action only fires when all of these are true at once:

No action chains. No "approve all." No standing authorization. Atomic, single-shot, user-driven.

Per-title risk tiers

Adapters ship with a hard-coded risk profile. The user does not get a toggle that overrides it.

TierExamplesOBSERVEASSIST
Green Single-player RPGs (e.g., Baldur's Gate 3, Elden Ring), sandbox / single-player Minecraft Yes — coaching, knowledge, overlay, voice Yes (opt-in) — inventory sort, dialogue confirm, journal nav
Yellow Co-op shooters, friendly PvE, modded contexts Yes Default off. Coaching only, no in-match action.
Red Ranked competitive shooters (CS2, Valorant ranked), MMOs with anti-bot rules Yes Force-disabled. No keystrokes are ever sent. Coaching is the entire surface.
The rule The red tier is non-negotiable. There is no setting that re-enables in-match actions on a ranked competitive title. The app is incapable of sending input under those profiles, by design.

What we don't do — by construction, not policy

✗
No DLL injectionThe app never injects code into the game process. Capture is the same OS API a screen recorder uses.
✗
No memory reading or writingWe never read the game's RAM. The state we know is what's visible on screen.
✗
No network interception or packet manipulationWe don't intercept, modify, or replay game traffic. Ever.
✗
No kernel driverGameMind is a normal user-mode Windows app. Nothing runs as a driver.
✗
No client patching or game file modificationWe do not touch game executables, configs, save files, or asset bundles.
✗
No multi-client input broadcasting, macros, or farming loopsActions are atomic, user-confirmed, rate-limited, cooldown-gated. There is no "run this 100 times" surface.
✗
No CAPTCHA-solving or anti-cheat evasionAny feature whose purpose is to look human to evade detection is not a feature we will build.

What a session looks like

You launch GameMind, then your gameThe desktop app starts in OBSERVE mode. The screen capture begins; nothing is sent anywhere.
Perception runs continuously, the heavy lane only at stable pointsPer-game adapter detects HUD, minimap, inventory, encounter phase. The fast lane is always on; the heavy reasoning model only fires when the screen settles.
Coaching surfaces in overlay and voice"This boss switches phase at 60% HP — your current build won't out-DPS phase 2." Cited to the official wiki and the current patch.
(Green tier only, ASSIST opt-in) GameMind proposes a single action"Sort inventory by weight?" with the keystroke it would send. You hit Confirm. One SendInput call goes out. Done.
End of session — local memory writes a sparse summaryWhat you did, what you missed, what to drill next. Stays on your machine. User-deletable.

What's under the hood (Qovaryx-native)

GameMind is structurally identical to the Qovaryx trading stack — same proven pattern with the domain swapped:

Same base substrate, same audit discipline as the Qovaryx public scratch bases. Apache-2.0 weights; the recipe is proprietary. Devlog: github.com/thron-j/qovaryx-ai-research.

Ready to install the beta?

Sign in, grab the desktop app, install the adapter for your main game, play.

Browse adapters → Get the app