Architecture

Overview

Viber is a lightweight, modular framework for building multi-agent AI applications. It provides the infrastructure for teams of AI agents to collaborate on complex tasks.

Core Concepts

Agent

An Agent is a specialized entity designed to perform specific tasks. Each agent has:

• Role Definition: Clear purpose defined by system prompt
• Tool Access: Set of tools it can use
• LLM Connection: Interface to language models
• Stateless Design: No memory between invocations (context provided at runtime)

Space

A Space is an isolated workspace that encapsulates:

• Artifacts: Generated files and outputs
• History: Conversation and task logs
• Configuration: Space-specific settings

Tool

A Tool extends agent capabilities with:

• Schema Validation: Zod-based parameter validation
• Type Safety: Full TypeScript support
• Sandboxed Execution: Secure runtime environment

Layered Architecture

┌─────────────────────────────────────────┐
│           Application Layer             │
│     (React / Svelte / Vanilla JS)       │
├─────────────────────────────────────────┤
│          Framework Adapters             │
│    (viber/react, viber/svelte)          │
├─────────────────────────────────────────┤
│              Core Engine                │
│   (Agent, Space, Tool, Streaming)       │
├─────────────────────────────────────────┤
│           LLM Providers                 │
│   (OpenAI, Anthropic, Google, etc.)     │
└─────────────────────────────────────────┘

Design Principles

Note

Viber follows a “batteries included but removable” philosophy.

1. Framework Agnostic: Core logic works with any UI framework
2. Type-First: Full TypeScript support with Zod schemas
3. Streaming Native: Real-time responses as first-class citizens
4. Composable: Small, focused modules that combine flexibly
5. Observable: Built-in event system for debugging and monitoring

Data Flow

User Input
    ↓
Agent.streamText()
    ↓
┌─────────────┐
│  LLM Call   │ ←→ Tool Execution (if needed)
└─────────────┘
    ↓
Stream Events → UI Updates
    ↓
Space Persistence (artifacts, history)

Multi-Agent Collaboration

Agents can collaborate through:

1. Task Handoff: Natural language delegation
2. Shared Spaces: Common artifact storage
3. Event Bus: Cross-agent communication

const team = {
  lead: new Agent({ name: 'Lead', ... }),
  writer: new Agent({ name: 'Writer', ... }),
  reviewer: new Agent({ name: 'Reviewer', ... }),
};