Introduction

Anvia is a small, explicit runtime layer for AI behavior inside a TypeScript application. It accepts the dependencies your application creates—models, tools, memory stores, vector indexes, observers, services, and transports—and coordinates them without taking over the rest of your architecture.

The boundary

Anvia owns:

• Normalized model calls and streaming.
• Agent turns and tool execution.
• Runtime messages, usage, and events.
• Connections to memory, context, observers, and transports.

Your application still owns:

• Authentication and authorization.
• Product permissions and data access.
• Durable storage and migrations.
• Deployment and infrastructure.
• The API and interface presented to users.

This boundary keeps application dependencies explicit. Create them in application code, inject them into Anvia, and replace them in tests.

Runtime layers

@anvia/core provides provider-neutral completions, agents, tools, memory interfaces, streaming events, and context.

Provider packages such as @anvia/openai, @anvia/anthropic, @anvia/gemini, and @anvia/mistral create models that implement the core interfaces.

Integration packages connect the runtime to application surfaces:

• @anvia/server exposes runtime streams from HTTP routes.
• @anvia/react consumes streams in React applications.
• @anvia/logger sends runtime events to application logs.
• @anvia/sandbox provides isolated command and file tools.
• @anvia/studio provides a local browser runtime for development.

Start small

Use a direct completion when your application already owns the control flow. Introduce an agent when behavior must be reusable, use tools, remember sessions, or span multiple turns.

Build your first agent