Skip to content
/ ART Public

ART is a powerful, modular, and browser-first JavaScript/TypeScript framework for building sophisticated LLM-powered intelligent agents capable of complex reasoning, planning, and tool usage.

License

Notifications You must be signed in to change notification settings

InferQ/ART

Repository files navigation

✨ ART: Agentic Runtime Framework Version 0.4.14

ART Framework Logo

ART Framework Overview

ART (Agentic Reactive Triad) is a powerful, modular, and browser-first TypeScript framework for building sophisticated LLM-powered agents capable of complex reasoning, planning, and tool usage.

It provides the missing runtime for production-ready agentic systems, emphasizing Reliability, Observability, and Resilience.

License: MIT PRs Welcome

Sponsor on Patreon

The Reactive Triad

ART is architected around three core nodes that ensure your agents are production-ready:

  1. Reactive Agent (Orchestration): The flagship PESAgent uses a Plan-Execute-Synthesize workflow to handle complex, multi-step tasks with dynamic refinement.
  2. Persistent State (Memory): Durable state management that persists across sessions, crashes, and human-in-the-loop pauses.
  3. Standardized UI (Observability): A reactive socket system that streams granular "thoughts," tool calls, and status updates directly to your frontend.

Key Features

🧠 Advanced Reasoning (PES Agent)

  • Structured Planning: Decomposes user intent into a discrete TodoList with dependency mapping.
  • TAEF (Tool-Aware Execution Framework): Strictly validates tool usage to eliminate hallucinations and ensure protocol adherence.
  • Dynamic Refinement: Self-correcting plans that adapt in real-time based on tool outputs.
  • Synthesis Engine: Aggregates task results into rich responses with UI metadata and citations.

🛡️ Production Robustness

  • HITL V2 (Human-in-the-Loop): Seamlessly pause execution for human approval with full state preservation and resumption.
  • Crash Recovery: Automatic state hydration ensures agents resume exactly where they left off.
  • A2A (Agent-to-Agent) Delegation: Coordinate complex workflows by delegating sub-tasks to specialized agents.
  • Step Output Table: Persists all historical tool results for consistent cross-step data access.

🔌 Universal Connectivity

  • Multi-Provider Support: First-class support for Gemini (Thinking), Claude (Extended Thinking), GPT (Reasoning), DeepSeek, Groq, and local models via Ollama.
  • MCP (Model Context Protocol): Dynamically discover and execute tools from any MCP-compliant server.
  • Pluggable Storage: Integrated adapters for IndexedDB (Browser), Supabase (Backend), and In-Memory.
  • OAuth & Auth Strategies: Built-in support for PKCE, Generic OAuth, and API key management.

Installation

npm install art-framework
# or
pnpm install art-framework

Quick Start

import { createArtInstance, PESAgent, OpenAIAdapter, CalculatorTool } from 'art-framework';

const art = await createArtInstance({
  storage: { type: 'indexedDB', dbName: 'AgentDB' },
  providers: {
    availableProviders: [{ name: 'openai', adapter: OpenAIAdapter }],
  },
  tools: [new CalculatorTool()],
  agentCore: PESAgent, // Default flagship agent
});

// Configure the thread (Where API keys live)
await art.stateManager.setThreadConfig('my-thread', {
  providerConfig: {
    providerName: 'openai',
    modelId: 'gpt-4o',
    adapterOptions: { apiKey: process.env.OPENAI_API_KEY },
  },
  enabledTools: ['CalculatorTool'],
});

// Process a request
const result = await art.process({
  query: 'Calculate the compound interest for $1000 at 5% over 10 years.',
  threadId: 'my-thread',
});

console.log(result.response.content);

Documentation

Contributing

Contributions are welcome! Please refer to the Contributing Guide.

License

ART Framework is released under the MIT License.

About

ART is a powerful, modular, and browser-first JavaScript/TypeScript framework for building sophisticated LLM-powered intelligent agents capable of complex reasoning, planning, and tool usage.

Resources

License

Contributing

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages