Skip to main content

​
Introduction

​
In one sentence

Agentium is the toolkit for building AI assistants that actually work in the real world — ones that remember your customers, take real actions, ask a human before doing anything risky, and don’t blow your budget.
New here? Read these three sections (“In one sentence”, “What problem it solves”, “How to think about it”) and you’ll understand what Agentium is — no technical background needed. Everything after that is for the people who’ll build with it.

​
What problem it solves

Today’s AI models (GPT, Claude, Gemini) are incredibly smart but have three real-world gaps:
  1. They forget. Every conversation starts from zero — the model doesn’t remember who you are.
  2. They can’t act. Out of the box, a model can only write text. It can’t look up an order, issue a refund, or send an email.
  3. They’re hard to trust in production. No spending limits, no human approval for risky actions, no record of what they did.
Agentium fills all three gaps. It’s the layer between the raw AI model and a real product — the part that turns “a clever chatbot” into “a reliable digital employee.”
You are a…What Agentium means for you
Business / CEOShip AI features customers trust — with memory, spending controls, and human oversight built in. Less time and money than building it yourself.
Product / CTOA production-grade foundation: memory, multi-agent teams, cost tracking, observability, voice, and browser automation — one TypeScript codebase, no paid add-ons.
DeveloperPure TypeScript. Swap models, storage, and transport with one line. Type-safe tools, batteries included, no Python runtime.

​
How to think about it

Imagine hiring a new employee. To be useful, they need more than intelligence:
  • a memory of who they’ve talked to,
  • tools to actually do their job (your systems, your APIs),
  • rules about what they can and can’t do without checking with a manager,
  • and a manager watching costs and quality.
A raw AI model is just the intelligence. Agentium gives it the memory, the tools, the rules, and the oversight — everything that turns raw intelligence into a dependable team member.

​
What is Agentium? (the technical version)

Agentium is a TypeScript-native agent orchestration framework for Node.js. It provides a complete toolkit for building AI-powered applications—from simple chatbots to complex multi-agent systems—with zero meta-framework dependency. Write pure TypeScript, plug in your preferred LLM provider, and ship production-ready agent applications. Agentium is designed for developers who want:
  • Full control over their agent architecture
  • Native TypeScript without Python runtime or transpilation layers
  • Flexibility to swap models, storage, and transport layers without rewriting code
  • Production readiness with built-in session management, memory, RAG, and background job queues

​
Key Features

Model Agnostic

Use OpenAI, Anthropic, Google Gemini, or Ollama with a unified interface. Switch providers with a single line change.

Multi-Agent Teams

Coordinate, route, broadcast, and collaborate across multiple agents. Build sophisticated agent hierarchies.

Stateful Workflows

Define agent steps, conditions, and parallel execution. Orchestrate complex multi-step pipelines.

Tools & Function Calling

Zod-validated tool definitions with type-safe execution. Full support for OpenAI-compatible function calling.

RAG & Knowledge Base

Vector stores, embeddings, and retrieval. Build context-aware agents with your own documents.

Multi-Modal

Handle images, audio, and files. Pass rich content to vision and audio-capable models.

Transport Layer

Express REST API and Socket.IO WebSocket gateway. Expose agents as HTTP or real-time endpoints.

Background Jobs

BullMQ queue and worker support. Process long-running tasks asynchronously.

Storage Drivers

In-memory, SQLite, PostgreSQL, and MongoDB. Choose the right persistence for your scale.

Memory & Sessions

Session history, LLM-powered long-term summaries, and cross-session user memory. Three complementary layers.

Voice / Realtime Agents

Speech-to-speech conversations via OpenAI Realtime and Google Gemini Live. Same tools, same memory, real-time audio.

Browser Agents

Vision-based autonomous browser automation. An agent sees screenshots, decides actions, and operates web pages via Playwright.

Sandbox Execution

Run tools in isolated subprocesses with timeout and memory limits. Fully optional — off by default.

Human-in-the-Loop

Pause the agent loop for human approval before executing sensitive tools. CLI, event-driven, or Socket.IO.

​
Why Agentium?

LangGraph.js is a powerful low-level orchestration framework, but it requires you to model everything as a directed state graph — defining nodes, edges, reducers, and conditional routing. It’s explicitly described as “very low-level, focused entirely on agent orchestration.” Agentium takes a higher-level, declarative approach: define an Agent with tools, a Team with members, or a Workflow with steps — no graph theory required.Both frameworks support human-in-the-loop, streaming, and state persistence. Where they diverge:
  • Voice agents: LangGraph has no built-in voice support — you’d need to integrate LiveKit or another real-time layer yourself. Agentium ships VoiceAgent with OpenAI Realtime and Google Live providers out of the box.
  • Browser automation: Not part of LangGraph. Agentium includes BrowserAgent with Playwright, stealth mode, credential vaulting, and video recording.
  • Multi-agent teams: LangGraph supports multi-agent via handoffs and supervisor patterns, but you wire the graph manually. Agentium provides Team with four built-in modes (coordinate, route, broadcast, collaborate) — one config object.
  • Transport: LangGraph relies on LangGraph Platform (paid) for deployment. Agentium ships free Express, Socket.IO, and A2A transport out of the box.
The Vercel AI SDK (v6) is a solid choice for single-agent tool-calling loops, especially if you’re already on Vercel. It has good model provider support, tool validation via Zod, and a ToolLoopAgent class. However, it’s designed primarily for individual agents inside Vercel’s ecosystem.Where Agentium goes further:
  • Multi-agent teams: The AI SDK has no built-in team coordination. Agentium provides Team with coordinator, router, broadcast, and collaborate modes.
  • Workflows: The AI SDK offers workflow “patterns” (docs examples), but no first-class Workflow class with typed state, parallel steps, retry policies, and conditions. Agentium does.
  • Voice agents: The AI SDK has speech and transcription utilities, but no real-time bidirectional voice agent. Agentium has VoiceAgent with OpenAI Realtime and Google Live.
  • Browser agents: Not part of the AI SDK. Agentium includes BrowserAgent.
  • Background jobs: The AI SDK relies on Vercel’s Fluid Compute (waitUntil). Agentium provides a standalone @agentium/queue package with BullMQ — runs anywhere, not just Vercel.
  • Framework lock-in: The AI SDK is optimized for Vercel + React. Agentium is framework-agnostic — Express, Fastify, Socket.IO, or headless.
LangChain.js is a large ecosystem with integrations for nearly every vector store, memory backend, and LLM provider. It’s great if you need breadth. However, the chain-based abstraction layer sits between you and the LLM, and the framework is actively steering users toward LangGraph for anything beyond simple chains (AgentExecutor is deprecated, sunset December 2026).Agentium takes a different philosophy:
  • Direct control: You work with Agent, Team, and Workflow directly — no chain composition or AgentExecutor to reason about.
  • Batteries-included: Session management, user memory, hybrid RAG (BM25 + vector), voice agents, browser automation, sandbox execution, and human-in-the-loop are all first-class — not spread across langchain, @langchain/community, and langgraph.
  • Simpler mental model: One framework, one set of docs. LangChain.js users currently need to decide between LangChain, LangGraph, LangSmith, and LangServe — each with its own API surface.
CrewAI is Python-only (98% Python on GitHub). Autogen supports Python and .NET, with more languages “coming.” Neither has native TypeScript support. If your stack is Node.js, you’d need to run a separate Python service, manage two runtimes, and deal with inter-process communication.Agentium keeps everything in TypeScript — same language, same process, same deployment. It also provides capabilities that these frameworks don’t:
  • Real-time voice agents with OpenAI Realtime and Google Live
  • Autonomous browser agents with Playwright and stealth mode
  • Transport gateways (Express REST, Socket.IO, Voice Gateway, Browser Gateway)
  • Background job queues with BullMQ
  • Sandbox execution for isolated tool running

​
Get Started

Ready to build your first agent? Head to the Quickstart to install Agentium and run a simple example in under five minutes.

Quickstart

Install, configure, and run your first Agentium agent in minutes.