What is diplomacy-ai?

This project runs a full game of the board game Diplomacy where all seven countries are played by AI language models instead of people. It builds on an existing open source Diplomacy game engine and adds AI agents that negotiate with each other and submit moves each turn, just like human players would. Each round of the game has three parts. First, during movement phases, every AI power reads messages sent to it and can send its own messages back to a specific other power privately or to everyone at once, across a set number of negotiation rounds. Second, each power submits its moves as structured data, which get checked against the game's official rules, if a move is illegal the system asks that power's AI once to fix it, and if it is still illegal the unit simply does nothing that turn rather than breaking the game. Third, the engine applies everyone's moves and saves the full record, including private messages and each AI's internal reasoning, so a person can review the whole game afterward, including any scheming that never became public. To run it, someone needs Python 3.11 or newer and a tool called just to run setup commands. The quickest path uses Google's Gemini AI model, which has a free tier: a person gets an API key from Google AI Studio, puts it into a configuration file, and starts a game with one command. Any AI model supported by a library called LiteLLM can be used instead, including running a model entirely on your own computer for free through a program called LM Studio, so an API key is not strictly required. Each of the seven AI powers can even be assigned a different underlying model or a different personality, or persona, described in plain text. After a game runs, its full record is saved to a folder including a game file that can be loaded into Diplomacy's official web interface to watch the board and public messages, plus separate files showing each AI power's private messages and reasoning per turn, and a running log of progress. There is also a test suite that runs without any network access, plus an optional live test that actually calls the AI model and costs a small amount of money to run.