Skip to main content
Open source · MIT licence

The open-source
agent harness.
Built in Europe.

Build production AI agents with tools, plugins, MCP, teams, and skills. TypeScript-native. Runs on Vercel AI SDK v6.

Get StartedView on GitHub

EU-Made & Open Source

MIT licensed, built in Europe. Full data sovereignty, no vendor lock-in. Audit every line of code.

Production-Grade

Durable sessions, checkpointing, context compaction, and full observability. Ready for real workloads.

Plugin Ecosystem

MCP integration, skills following the agentskills.io specification, extensible hooks system, and middleware.

Multi-Agent Teams

Coordinate multiple agents with shared tasks, message passing, spawning, and team-level observability.

Everything you need to build agents

A complete framework for creating, orchestrating, and deploying AI agents in production.

createAgent()

Agent Creation

Define agents with createAgent() — model, system prompt, tools, plugins, and hooks in one call.

@ai-sdk/mcp

Tool Calling & MCP

Built-in tools for file ops, bash, search. Connect any MCP server for external integrations.

SKILL.md

Skills System

File-based (SKILL.md) and programmatic (defineSkill) skills following the agentskills.io spec.

PreToolUse

Hooks & Middleware

Lifecycle hooks for every stage: PreGenerate, PostToolUse, guardrails, rate limiting, logging.

stream()

Streaming & Real-time

Stream responses to clients with stream(), streamResponse(), and streamDataResponse().

createAgentTeamsPlugin()

Agent Teams

Multi-agent coordination with shared tasks, mailboxes, spawning, and team observability.

Simple, powerful API

From a single agent to coordinated teams — a few lines of TypeScript is all you need.

import { createAgent } from '@lleverage-ai/agent-sdk'
import { anthropic } from '@ai-sdk/anthropic'

const agent = createAgent({
  model: anthropic('claude-sonnet-4-5-20250514'),
  system: 'You are a helpful coding assistant.',
  tools: ['read', 'write', 'edit', 'bash', 'glob', 'grep'],
})

const result = await agent.stream(
  'Refactor the auth module to use JWT tokens'
)

How it works

The SDK composes agents from modular building blocks. Each piece is optional and replaceable.

createAgent()
Agent
Model + System Prompt + Config
Tools
read, write, bash, MCP
Plugins
bundled tool+hook sets
Skills
SKILL.md + defineSkill
Hooks
PreGenerate, PostToolUse, guardrails
Middleware
logging, context, rate limiting
AI Provider
Anthropic · OpenAI · Google · Any AI SDK provider

Get started in minutes

Three steps to your first agent. Full documentation on GitHub.

01

Install

bun add @lleverage-ai/agent-sdk ai zod
bun add @ai-sdk/anthropic  # or @ai-sdk/openai
02

Create an agent

import { createAgent } from '@lleverage-ai/agent-sdk'
import { anthropic } from '@ai-sdk/anthropic'

const agent = createAgent({
  model: anthropic('claude-sonnet-4-5-20250514'),
  system: 'You are a helpful assistant.',
  tools: ['read', 'write', 'bash'],
})
03

Run it

const result = await agent.stream(
  'Help me set up a new Next.js project'
)

// Or use streamResponse() for HTTP endpoints
return agent.streamResponse(request)
Full documentation on GitHub