Introduction

GraphAI is a minimal "AI framework" that tries hard not to be one. It gives you a small, flexible graph runtime — nodes, edges, routers, state, streaming — and then gets out of the way. You build the agent, the LLM wrapper, the memory, exactly how you want them.

What you get

GraphAI is built around a computational graph. It provides:

1. A graph-based architecture for wiring components into a workflow.
2. An async-first design, so waiting on API calls doesn't waste compute.
3. Minimal abstractions — no baked-in notion of "LLM" or "Agent" to fight against.
4. Callbacks for streaming and communication between components.

Other libraries ship their own idea of what an LLM, an agent, or a tool should look like. GraphAI doesn't. It hands you the primitives and lets you define those concepts yourself.

Why

Most AI frameworks impose a shape on your application. That shape becomes a local minimum: fine right up until you need something the framework didn't anticipate. GraphAI takes the opposite stance.

• Bring your own components. Any LLM provider, any agent design, any telemetry.
• Build the workflow you actually need, not the one the framework assumes.
• Think in flow, not framework. Focus on how data moves through your app.

The core ideas

Async-first. AI apps spend most of their time waiting on APIs. GraphAI is async from the ground up, so your code stays busy while responses are in flight.

A graph of nodes. Nodes do the work. Edges say where data flows next. Routers are nodes that choose the next path. Complex workflows stay readable because the structure is the diagram.

Only what you need. A Graph to orchestrate, @node and @router to define steps, a callback for streaming, and a state dict for context. That's the whole surface.

When to reach for it

GraphAI fits when:

• You want full control over your architecture.
• Existing frameworks feel too opinionated.
• You're mixing components from different ecosystems.
• You prefer explicit code over magic.
• You're building something that doesn't fit the usual patterns.

Ready? The quickstart builds a working agent in a few minutes.