lpalbou/AbstractCore

Provides unified sync/async API with first-class support for streaming, universal tool calling (native + prompted syntax), and Pydantic structured output across local (Ollama, LMStudio, vLLM) and cloud providers. Media input uses policy-driven fallbacks—if a model lacks vision support, AbstractCore can automatically invoke a vision model to generate visual observations and inject them into text requests. Designed as composable layer: by default returns tool calls without executing them, allowing optional integration with AbstractRuntime for durable workflows and state persistence.

Available on PyPI.