Hearthstone AI Assistant: A Read-Only, Privacy-First Desktop Intelligent Analysis Tool

Hearthstone Agent Assistant is a desktop AI assistant designed specifically for Hearthstone players. It constructs structured game state snapshots by reading game logs and screenshots, and invokes multiple large language models (LLMs) for strategic analysis. The tool uses a read-only architecture—no game process injection, no anti-cheat bypassing—and provides complete privacy protection. It supports multiple AI service providers and offers various UI modes, delivering legal and secure intelligent assistance to players.

With the development of AI technology, LLMs are widely used in game assistance. However, traditional cheats often involve illegal operations (memory injection, process modification), violating game terms and easily leading to account bans. This project aims to provide a legal, secure, and privacy-first AI assistance tool. Its core concept is 'read-only': no interface clicks, no card plays, no process injection, no memory reading, no anti-cheat bypassing. All analysis is based on game logs and screenshots to ensure compliance.

Multi-AI Service Provider Support

Compatible with 9 mainstream AI service providers including OpenAI, DeepSeek, Zhipu GLM, Qwen, Kimi, etc., supporting parallel query and comparison.

Power.log Parser

A handwritten state machine parses Hearthstone debug logs (events like TAG_CHANGE, SHOW_ENTITY), processes multi-GB logs, and constructs complete game state snapshots.

Visual Verification System

Captures game window screenshots, uses dHash algorithm to verify card positions, ensures log state matches the screen, and screenshots are only used locally.

Privacy Protection

Data desensitization (removing entity tags, hiding player names, etc.), sensitive information never leaves the local device.

Adopts a three-window design:

• Full Dashboard: Borderless and resizable, displays detailed game state and AI suggestions (sorted by confidence).
• Transparent Floating Window: Always on top, provides real-time AI suggestions without blocking the screen, toggle display with shortcut Ctrl+Shift+O.
• Minimized Drag Ball: Simple background running mode.

Adoption Tracking

Compares pre- and post-turn states, automatically identifies AI suggestions executed by the player, and tracks the historical adoption rate of each model.

Multi-Agent Comparison

Queries multiple LLMs in parallel, fuses results by confidence, visually compares suggestion differences, and improves decision reliability.

Structured Output and Graceful Degradation

Three layers of protection: json_schema output as the first choice, degrade to json_object, local validation ensures legality.

Layered Architecture

Code is divided into core (pure logic), main (Electron main process), renderer (React UI), and shared (type definitions), making it easy to maintain and extend.

Encrypted Credential Storage

API keys are encrypted and stored via Electron safeStorage, not saved in plain text, protecting sensitive information.

• New Player Learning: Quickly master strategies like card combinations and board control through AI suggestions.
• Decision Assistance: Provides additional perspectives in complex situations to find optimal solutions.
• Model Comparison: Compare suggestions from multiple AI models to understand strategic differences.
• Privacy Security: As a read-only local tool, it protects privacy better than cloud-based tools.

This project demonstrates a new paradigm for AI assistance: using legal technical means (log parsing, screenshots) to provide intelligent analysis, which can be extended to other games with log output. The architectural design (multi-provider support, privacy protection, layered structure) is worth learning from. Future improvements in LLM capabilities will make the tool more intelligent, and such compliant projects set a benchmark for the healthy development of the field.