A next-generation open-source AI agent development framework and runtime platform featuring an event-driven architecture, container-like concept models (Prototype → Image → Agent), and three deployment modes (local embedded, WebSocket server, remote client) with multi-LLM provider support via Vercel AI SDK.
AgentX is an AI agent development framework and runtime platform developed by the Deepractice organization, primarily written in TypeScript (98.1%), running on Node.js with Bun support, under the MIT license. It is actively developed (871+ commits, 60 releases, latest v2.9.0).
Core Architecture The project adopts a layered event-driven architecture built on an RxJS Pub/Sub EventBus for decoupled inter-module communication, defining a four-layer event system: Stream / State / Message / Turn. All input (Client → WebSocket → BUS → LLM Driver) and output (Driver → BUS → AgentEngine → BUS → Client) flow through the event bus.
Container-like Concept Model Drawing from container runtime paradigms, AgentX defines three levels of abstraction: Prototype (Agent template, analogous to Dockerfile) → Image (persisted configuration, analogous to Docker image) → Agent (runtime instance, analogous to Container). Images can declaratively specify mcpServers, RoleX configurations, etc.
Three Deployment Modes
- Local Mode: In-process embedded Agent creation for single-process development
- Server Mode: One-line
ax.serve({ port })to expose Agents as WebSocket servers - Remote Mode:
ax.connect("ws://...")to connect to remote servers with API parity to local mode
Multi-LLM Support
Unified via Vercel AI SDK, supporting Anthropic (Claude), OpenAI (GPT), Google (Gemini), DeepSeek, Mistral, xAI (Grok), and OpenAI-compatible providers. Per-Request Overrides allow dynamic switching of model, thinking depth, and other parameters per send call.
Tools & Extensions
Supports declarative MCP Servers integration at the Image configuration level; integrates with the sister-project RoleX for identity cognition and growth cycle management; provides @agentxjs/devtools BDD testing toolkit (MockDriver, RecordingDriver, Fixtures).
Package Structure
npm monorepo built with Turbo. Core packages include agentxjs (Client SDK), @agentxjs/core (core abstractions), @agentxjs/node-platform (Node.js platform adapter with SQLite persistence), @agentxjs/mono-driver (multi-provider driver), @agentxjs/claude-driver (Claude extended driver), @agentxjs/devtools (testing tools).
API Layers
ax.chat.*— Chat management (create, list, get → AgentHandle)ax.provider.*— LLM Provider configuration and switchingax.runtime.*— Low-level subsystems (image, session, container)
Ecosystem Part of the Deepractice AI infrastructure alongside RoleX (AI role management), ResourceX (unified resource manager), and IssueX (AI-collaborative structured issue tracking). Repository links for sibling projects are pending confirmation.
Pending Confirmation: Official website/docs URL, npm publication status (agentxjs etc.), Hugging Face presence, accompanying academic papers.