Documentation versions
Engine Atlas MCP
MCP connects your AI to this site. The model reads cards, studies the code in Ghidra and writes straight in edit mode: its edits appear on the cards and in My edits, and it sees your changes right away. Nothing is deployed locally: you only need the server itself and Ghidra.
Strong models are strongly recommended, such as Sonnet 5.5, GPT-6 Sol or newer. There are no hard requirements, but weaker models more often retell code instead of explaining it.
Connecting
Installation is in the README, and setup for Claude Code, Codex and Antigravity is in the client guide. Ghidra with the Ghidra MCP plugin and libasphalt8.so open is required, see Researching the engine.
To watch the model's work, open the edits link it gives you, or turn on MCP link in the site settings. The browser asks once for permission to reach your local network: allow it. Chrome, Edge and Firefox work.
How to work
You do not need to explain the rules: the model gets them from the server. Just say what to study.
- Research. "Study the class gin::Widget in Engine Atlas." One class or a small group of related classes works best.
- Questions. If the model asks about in-game meaning, answer with what you know. Its guesses must not turn into facts.
- Critic review. Open a new session, ideally with another model: "You are the critic. Check the cards for gin::Widget by the critic role in the guide."
- Fixes. Pass the findings to the author and repeat the review.
- Sending. Review My edits and press Send, or tell the model: "Sign the edits with your name and send them for review."
What the critic is
The critic is a second AI session that looks for errors in someone else's cards. It must be a new session because an author is poor at seeing its own guesses. The critic does not trust the author's report, reads the database and game code itself, tests each statement and returns the card with specific findings. It does not write the text for the author.
A returned card is normal work, not a failure. If the critic invents nitpicks or demands facts the code does not contain, that is the critic's own error and you can ignore it. The worrying case is the opposite: a critic that praises everything has probably not read the code. Ask for a new review.
A good result and a bad one
Open the card in reading mode and imagine seeing the class for the first time.
| Good | Bad |
|---|---|
| "Fits children into a box of a given shape and centres it" | "A container class for working with aspect ratio" |
| Explains what a value means and what changes in the game | "Updates state", "performs logic", a translated method name |
| An argument is described with its limits and modes | "Value according to the signature" |
| "Ownership is unknown: the creator was not found" | A confident ownership claim with no check |
| The condition is stated: "only in mode 1" | One code branch presented as a general rule |
| An example solves a task and says honestly whether it was tested | An invented API call, or "tested" when nothing ran |
Warning signs:
- The article contains addresses, hashes, file paths or a paragraph about the checking process. These belong in evidence, not in text for readers.
- The text reads smoothly, but afterwards you cannot say why the class exists.
- Every property is confirmed, yet there are no notes or evidence.
- One method's description is copied into its neighbours.
- It says "tested in game" although nobody ran the game. Reading code in Ghidra is a static check, not a test in the game.
An honest "unknown" beats a pretty guess. Such cards are valued: the next researcher sees what is left to find out.
What to keep
Do not delete notes or research history. They travel with your edits, and they are why AI research is valued above manual work: you can see how each conclusion was reached.
Never send other people's keys, personal data or game files. Published cards change only through review: the MCP publishes nothing by itself.