AmbiA small, composable AI Agent Runtime for Rust

Compose capabilities, not Agent hierarchies. Ambi provides execution mechanisms; you define the behavior.

Get started

Design philosophy

GitHub

🧩

Composition over inheritance

Replace engines, pipelines, tools, parsers and formatters. No mandatory Agent class hierarchy or graph architecture.

🌳

Agent · Pipeline · Leaf

Reusable capabilities form the tree, execution strategies are branches, and optional disposable task executions are leaves.

🔧

Small runtime, open extension points

Built-in ReAct and optional pipeline/memory/macro crates. Other developers define what grows around the core.

⚙️

Cloud, local and custom backends

OpenAI-compatible APIs, native llama.cpp inference and user-defined engines behind a common runtime contract.

Begin with the design, then the API

Ambi does not prescribe how all agents should think or cooperate. It supplies the small mechanisms for running them. See Design philosophy and Architecture before choosing a workflow.

New in 0.4.1: opt-in Leaf task delegation gives ChatRunner a temporary, isolated subtask execution without creating a persistent second Agent. A Leaf inherits the parent's original tools and shares its engine, so it is not a security sandbox.

Continue with Getting started or Custom Pipeline.