SimVX AI

SimVX's LLM layer: one client for any OpenAI-compatible /chat/completions endpoint, a record-and-replay cache for tests, and in-game loops that never stall a frame. Most users want simvx, which includes it.

It works with any OpenAI-compatible server, local (vLLM, llama.cpp, OpenWebUI) or hosted: bring your own endpoint and key.

Pre-alpha. SimVX publishes development releases only (0.0.0.devN) and is under active development; the API changes between builds. Once stable versions ship to PyPI, development builds continue on pypi.simvx.com.

Install

pip install --pre simvx        # the whole engine (recommended)
pip install --pre simvx-ai     # the LLM layer on simvx-core alone

Latest development build: add --extra-index-url https://pypi.simvx.com/.

Use

Bark runs a model call off the frame thread at most every period seconds; poll returns the last good line at once, every frame.

from simvx.ai import Bark, OpenAICompatibleClient

client = OpenAICompatibleClient.from_env()   # SIMVX_LLM_BASE_URL, SIMVX_LLM_API_KEY, SIMVX_LLM_MODEL
bark = Bark(client, persona="a weary night watchman", period=6.0)

def on_update(self, dt):
    self.line = bark.poll(dt, hp=self.hp)

Planner does the same for a JSON-schema decision; CachingClient replays recorded responses. AgentSession lets a model play a running game (observe, send input, watch signals, step), and NodeGenerator turns a one-line description into a validated Node subclass.

Documentation · Source · Licence: AGPL-3.0-or-later with Additional Terms, including the Game-Distribution Exception.