MCP.md

AI agents

An AI model can drive Cheat Engine for you: find a value, narrow it down as it changes in the game, freeze it, follow pointers, find the code that changes it, write an Auto Assembler script that changes that code, and save a cheat table. Use it from the app (the AI Assistant) or from another AI app over MCP; both use the same tools.

In the app: the AI Assistant

Open Tools > AI Assistant (Ctrl+Shift+I) and describe what you want, for example "My health shows 100. Find it and keep it at 100." The assistant works on the process you have open.

• Your model, your key. In Settings > AI Assistant, pick Anthropic (Claude) or any OpenAI-compatible API: OpenAI, or a model on your own computer with Ollama (http://localhost:11434/v1) or LM Studio. Paste your API key there; local servers usually need none.
• Keys stay in your keyring. The key is saved in the system keyring (GNOME Keyring, KWallet or any Secret Service), never in Cheat Engine's settings file. ANTHROPIC_API_KEY or OPENAI_API_KEY in the environment takes precedence, but only for that provider's own API: it is never sent to another server you enter as the API address. A saved key belongs to the API address it was entered for: change the address and it is not sent until you enter it again, and it is never sent over plain http except to this machine. Without a keyring the key lasts until Cheat Engine closes.
• You approve every change. Scanning, reading and disassembling run on their own; writing a value, freezing, enabling or disabling a script, watching which addresses an instruction accesses (a breakpoint in the game's code), saving a file and opening any process other than yours wait for your Allow in the panel. Allow for this chat covers only the process it was given for, and never opening another process or overwriting a file. Stop ends a turn at any time.
• What leaves your computer. Your messages and what the tools find in the game (process names, addresses, values, code) are sent to the provider you chose, and nothing else. With a local model nothing leaves your computer.
• Keep the results. Add to address list copies the values, pointer paths and scripts the assistant saved into your table. Scripts it enabled are undone when you close Cheat Engine or open another process.

From another app: MCP

cescan mcp speaks the Model Context Protocol over stdin/stdout, so any MCP client can drive Cheat Engine with the model of your choice.

Connect

Claude Code:

claude mcp add cheat-engine -- cescan mcp --allow-writes

Claude Desktop and other clients take the same command in their MCP server settings, for example:

{ "mcpServers": { "cheat-engine": { "command": "cescan", "args": ["mcp", "--allow-writes"] } } }

Without --allow-writes the agent can look but not touch: scanning, reading and disassembling work, while writing values, freezing, enabling and disabling scripts, and find_accessed_addresses (which puts a breakpoint in the game's code) are refused. Turning writes off ends any freeze. Scripts checked or enabled by the agent cannot load libraries or binaries, include local files, start threads or compile C code; both their [ENABLE] and [DISABLE] sections are dry-run first. When the session ends or the agent attaches elsewhere, its scripts are undone by restoring the saved bytes, not by running their [DISABLE] sections. The server only talks to the client that started it; it opens no network port.

Tools

Tool	What it does
list_processes, attach, target_info	Find and open the game
first_scan, next_scan, undo_scan, scan_results	Scan for a value and narrow the results (including increased or decreased by an amount)
read, write	Read or change a typed value
disassemble	Show the code at an address, with symbols
find_what_writes	Watch an address and report the instructions that write to it, with the pointer each one followed
find_accessed_addresses	Watch one instruction and list every address it touches (code shared by player and enemies)
compare_structures	Dissect structures side by side and find the field that tells two groups apart, such as a team
pointer_scan, pointer_rescan	Find pointer paths to a value the game moves, and keep the ones that still work after it moves
aa_template	Build a code-injection script for an instruction
aa_check	Check a script by assembling it against the game without changing anything
aa_enable, aa_disable, aa_list	Apply a script to the game and undo it
add_record, list_records, set_frozen, save_table	Build, freeze and save a cheat table, including pointer paths (followed on every freeze) and scripts

Results are short summaries (counts and the first candidates), never memory dumps. Scripts are checked by assembling them against the live game with every change kept aside, so a misspelled instruction or an assert that no longer matches is reported before anything is patched. Agent scripts cannot run embedded Lua ({$lua}), and when the session ends every script the agent enabled is disabled, restoring the original code. Addresses in saved tables are module-relative (game+0x301C), so the table keeps working after the game restarts.

Example

You: My health is 100 in the game window. Find it and keep it at 100.

The agent attaches, scans for 100, asks you to take some damage, scans for the new value, confirms the address, freezes it at 100, and offers to save game.CT.

You: Make health never go down instead.

The agent watches what writes health while you take damage, builds an injection at that instruction, removes the store, checks and enables the script, asks you to confirm, and saves it to the table as "Infinite health".

Try it on the tutorials

cetutorial and cegtutorial (also in the app's Help menu) are games made to be changed. Ask the agent to finish a step, for example "Do step 9 of the tutorial". test/tutorial_gate.py plays every step and level through these tools the way an agent would, and runs in CI.

Use it on single-player games and software you are allowed to analyze.