Timbal is an open-source Python framework for building reliable AI applications with a small, transparent architecture. It provides autonomous agents and explicit workflows through the same calling interface and event stream. Developers define behavior with familiar async functions, Pydantic validation, and composable tools instead of relying on hidden orchestration. The framework supports multiple model providers and can automatically fail over between configured models. Human approval gates can pause runs, preserve their state, and resume them after a process restart. Built-in capabilities cover persistent memory, context compaction, structured outputs, reusable skills, MCP servers, tracing, and evaluations. Applications can run locally, be self-hosted, or be deployed through the Timbal platform with an accompanying API and user interface.
Features
- Unified agents and workflow interface
- Asynchronous and event-driven execution
- Provider-agnostic models with automatic fallback
- Human approval gates with resumable state
- Memory, structured outputs, tools, and MCP support
- Tracing, evaluations, local execution, and deployment