CEGM — AI skill for Claude Code

CheatEngineGM (CEGM)

A live, observable LLM-driven layer over Cheat Engine. Open CE → port `27077` is up → talk to the model from a browser tab or any MCP client (Claude Desktop, Cursor, Claude Code, Codex) → watch tool calls and memory diffs land in real time. Single-player offline use only.

**Status:** alpha (`v0.1.0a1`). Closed-loop MVP + RE Workbench shipped. Roadmap: [docs/ROADMAP.md](docs/ROADMAP.md).

What it is

CEGM is a thin **experience layer** built on top of the excellent [miscusi-peek/cheatengine-mcp-bridge](https://github.com/miscusi-peek/cheatengine-mcp-bridge) (MIT, vendored as a git submodule). When Cheat Engine starts, our Lua autorun spawns a small Python broker that:

• Exposes an MCP Streamable HTTP endpoint at http://127.0.0.1:27077/mcp so any MCP client (Claude Desktop, Cursor, Claude Code, Codex, …) gets the full tool list with zero per-host configuration — as long as CE is running, the URL is live.
• Serves a built-in browser dashboard at http://127.0.0.1:27077/ with chat input, a live tool-call timeline, before/after diffs on memory writes, and a settings drawer for your LLM key.
• Adds a small set of CEGM-namespaced tools layered on top: preview-before-commit for writes, named snapshots, restore points, and (Phase 3) a recipe library of common workflows.

The differentiating bet: **observability and safety**. Every surveyed competitor pipes tool calls headlessly to whatever client the user has open. CEGM shows them, lets you preview destructive ones, and lets you roll back.

Why

Manually finding base addresses, offsets, and pointer chains for single-player game memory is repetitive. CEGM offloads the grind to a model while keeping every step visible and reversible.

**Scope:** single-player offline games only. CEGM is not for online or multiplayer titles, and is not designed to evade anti-cheat.

How it works

                    ┌─ External MCP clients ─────────────┐
                    │  Claude D