About this project
Theseus's Brain is a Chinese pure-text roguelike designed mainly for AI play over MCP (stdio), and can also be played in a terminal CLI; it is zero-dependency, Python 3, and provides a complete English version (THESEUS_LANG=en). The game revolves around reincarnation, memory, and irreversible mechanical modification: mechanization rate accumulates across lives, and factions are tiered by mechanization rate, from 0% Pureblood Oath to 70–100% Ascension Spiral. After Act One of each life, you are asked only once whether to undergo modification; modification is irreversible, and death cannot wash it away. The MCP tools are new_run, choose, status, legacy, bequeath, debrief, and recite. When new_run is called without arguments, it does not start a run; instead it returns a menu for a human or AI to choose a disclosure tier for how the game is presented (story, story_ai, brief, brief_ai, auto, sealed), distinguishing 'how much is read' from 'who chooses'. brief compresses outcomes while scene text is read in full, and sealed gives only a paraphraseable battle report. The core mechanic is two kinds of memory: the brain's memory decays with mechanization rate, and each life's skills are written point by point into the legacy with probability based on the final mechanization rate; the world's memory does not decay, and across lives it becomes echoes, achievements, and era dice. Memory entries are written down via bequeath, with a total of 10 entries, each no more than 10 characters; deletion requires typing the original text exactly using discard, and each entry independently rolls based on mechanization rate to determine whether it survives to the next life. Ways to run it include claude mcp add with python3 server.py, python3 server.py --cli for interactive terminal play, and http_bridge.py providing an HTTP MCP bridge, with support for --isolate, --token, --allow-origin, and --selftest. Saves are written to saves/legacy.json and saves/current.json; deleting legacy.json resets the cycle. After disconnection, status can be used to resume the current act; new_run will not overwrite an ongoing life unless abandon=true is explicitly set. The project includes seven gate self-tests and 400 randomized auto-play verification runs, checking event copy, skill names, death echoes, option prompts, author markers, spoiler references, and exit keys; there are sixty-eight storylines and an explicit exit strategy. Copy can be extracted into a numbered revision table via docgen.py, and docxio.py handles bidirectional Word table conversion; English is a language layer rather than a fork, with langpack.py check, check-ui, and stale used to keep it in sync. Licensing: code is GPL-3.0, and copy and documentation are CC BY-SA 4.0. The README explicitly contains no spoilers; the truth, the lake, the finale, and the endings are written in SPOILERS.md.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.