●TYPESCRIPT AI AGENT FRAMEWORK

Build AI agents.
Not scaffolding.

Decorator-based agents with tools, behaviors, and multi-agent composition.
Works with any LLM — local or cloud.

agent.ts

@agent({
  name: "researcher",
  llmClient: new OllamaClient("qwen2.5:7b"),
  tools: [new WebTools()],
  behavior: "./researcher.md",
  maxSteps: 6,
})

class Researcher {}

// That's it. Run it.
const result = await runAgent(Researcher, "What's new in AI this week?");

Core Features

Decorator-first API

Define agents and tools with @agent and @tool. Zero boilerplate, full type safety.

⊕

Multi-agent composition

Use compose() to nest agents as tools. Build orchestrator-executor patterns without custom routing code.

✎

Behavior engine

Define personality in plain Markdown files. Hot-reload in dev. Extend and merge behaviors across agents.

⟲

Resilient runtime

Auto-retry with backoff, state persistence, middleware hooks, and streaming support — out of the box.

How it works

SINGLE AGENT

User prompt

LLM decides action

Tool executes

LLM synthesizes output

MULTI-AGENT VIA COMPOSE()

Orchestrator agent

SubAgent A (tool call)

SubAgent B (tool call)

Aggregated result

Up in 60 seconds.

Works with Ollama locally.

# install

$ npm install typed-agents reflect-metadata zod

# pull a local model

$ ollama pull qwen2.5:7b

# run your first agent

$ npx tsx agent.ts

Build AI agents.Not scaffolding.