Jev plays Vampire Survivors

The character, the stage, every level-up choice and the walking direction, four times a second, in the real Steam game. Character and stage picks are sampled from Jev's probability distribution so consecutive runs vary, while level-up and direction decisions always take its top choice.

A thin BepInEx 5 plugin reads raw game state, applies the brain's answers and drives the menus. It holds no strategy of its own and connects to the Python brain over newline-delimited JSON on TCP 127.0.0.1:48231, reconnecting automatically if the connection drops.

One static page the brain serves over a WebSocket at http://127.0.0.1:48232: a card per question with a probability bar per option, the chosen option highlighted, confidence, latency and a radar view of what the model was shown.

Vampire Survivors on Steam on native Linux - it does not work on Windows or macOS as shipped - access to Jev with an API key, and some command-line comfort: setup involves a few scripts and one manual Steam setting. It is not affiliated with poncle or TypeSafe.

The brain still runs, but every decision falls back to a small heuristic instead of the model, as documented in .env.example. The design uses two local processes only; there is no cloud infrastructure of your own to run.

It turns raw coordinates into language a language model can reason about: compass sectors, distance buckets and HP buckets, never raw numbers. All judgment lives in the brain - question definitions, Jev calls, safe fallbacks when the model or network is unavailable, run logs and the dashboard.

Python 3.12+ with uv and typesafe-sdk for the brain, C# for the BepInEx plugin, MIT licensed with CI on GitHub Actions. The README says neither process has any game-specific strategy hard-coded beyond asking the model, and the repo includes none of the game's assets or code.