SmythOS - The Linux of AI Agents

Reliable Agent Engineering starts with great, open source infrastructure. This repository contains the Smyth Runtime Environment Kernel (SRE), the Software Development Kit (SDK) and Command Line Interface (CLI) for running agents and creating them with code. If you prefer visual drag & drop agent interfaces instead, check out our open source SmythOS Visual Agent Studio! Great community, support, tutorials. Start in minutes!

Why SmythOS exists

1. Shipping production-ready AI agents shouldn’t feel like rocket science.
2. Autonomy and control can, and must, coexist.
3. Security isn’t an add-on; it’s built-in.
4. The coming Internet of Agents must stay open and accessible to everyone.

Design Principles

SmythOS provides a complete Operating System for Agentic AI. Just as traditional operating systems manage resources and provide APIs for applications, SmythOS manages AI resources and provides a unified SDK that works from development to production.

Unified Resource Abstraction

SmythOS provides a unified interface for all resources, ensuring consistency and simplicity across your entire AI platform. Whether you're storing a file locally, on S3, or any other storage provider, you don't need to worry about the underlying implementation details. SmythOS offers a powerful abstraction layer where all providers expose the same functions and APIs.

This principle applies to all services - not just storage. Whether you're working with VectorDBs, cache (Redis, RAM), LLMs (OpenAI, Anthropic), or any other resource, the interface remains consistent across providers.

This approach makes your AI platform easy to scale and incredibly flexible. You can seamlessly swap between different providers to test performance, optimize costs, or meet specific requirements without changing a single line of your business logic.

Key Benefits:

• Agent-First Design: Built specifically for AI agent workloads
• Developer-Friendly: Simple SDK that scales from development to production
• Modular Architecture: Extensible connector system for any infrastructure
• Production-Ready: Scalable, observable, and battle-tested
• Enterprise Security: Built-in access control and secure credential management

Quick Start

We made a great tutorial that's really worth watching:

Method 1: Using the CLI (Recommended)

Install the CLI globally and create a new project:

npm i -g @smythos/cli
sre create

The CLI will guide you step-by-step to create your SDK project with the right configuration for your needs.

Method 2: Direct SDK Installation

Add the SDK directly to your existing project:

npm install @smythos/sdk

Check the Examples, documentation and Code Templates to get started.

---

Note: If you face an issue with the CLI or with your code, set environment variable LOG_LEVEL="debug" and run your code again. Then share the logs with us, it will help diagnose the problem.

Repository Structure

This monorepo contains three main packages:

SRE (Smyth Runtime Environment) - packages/core

The SRE is the core runtime environment that powers SmythOS. Think of it as the kernel of the AI agent operating system.

Features:

• Modular Architecture: Pluggable connectors for every service (Storage, LLM, VectorDB, Cache, etc.)
• Security-First: Built-in Candidate/ACL system for secure resource access
• Resource Management: Intelligent memory, storage, and compute management
• Agent Orchestration: Complete agent lifecycle management
• 40+ Components: Production-ready components for AI, data processing, and integrations

Supported Connectors:

• Storage: Local, S3, Google Cloud, Azure
• LLM: OpenAI, Anthropic, Google AI, AWS Bedrock, Groq, Perplexity
• VectorDB: Pinecone, Milvus, RAMVec
• Cache: RAM, Redis
• Vault: JSON File, AWS Secrets Manager, HashiCorp

SDK - packages/sdk

The SDK provides a clean, developer-friendly abstraction layer over the SRE runtime. It's designed for simplicity without sacrificing power.

Why Use the SDK:

• Simple API: Clean, intuitive interface that's easy to learn
• Type-Safe: Full TypeScript support with IntelliSense
• Production-Ready: Same code works in development and production
• Configuration-Independent: Business logic stays unchanged as infrastructure scales

CLI - packages/cli

The SRE CLI helps you get started quickly with scaffolding and project management.

Code examples

The SDK allows you to build agents with code or load and run a .smyth file. .smyth is the extension of agents built with our SmythOS builder.

Example 1 : load and run an agent from .smyth file

async function main() {
    const agentPath = path.resolve(__dirname, 'my-agent.smyth');

    //Importing the agent workflow
    const agent = Agent.import(agentPath, {
        model: Model.OpenAI('gpt-4o'),
    });

    //query the agent and get the full response
    const result = await agent.prompt('Hello, how are you ?');

    console.log(result);
}

Want stream mode ? easy

Click to expand: Stream Mode Example - Real-time response streaming with events

    const events = await agent.prompt('Hello, how are you ?').stream();
    events.on('content', (text) => {
        console.log('content');
    });

    events.on('end', /*... handle end ... */)
    events.on('usage', /*... collect agent usage data ... */)
    events.on('toolCall', /*... ... */)
    events.on('toolResult', /*... ... */)
    ...

Want chat mode ? easy

Click to expand: Chat Mode Example - Conversational agent with memory

    const chat = agent.chat();

    //from there you can use the prompt or prompt.stream to handle it

    let result = await chat.prompt("Hello, I'm Smyth")
    console.log(result);

    result = await chat.prompt('Do you remember my name ?");
    console.log(result);


    //the difference between agent.prompt() and chat.prompt() is that the later remembers the conversation

Example 2 : Article Writer Agent

In this example we are coding the agent logic with the help of the SDK elements.

Click to expand: Complete Article Writer Agent - Full example using LLM + VectorDB + Storage

import { Agent, Model } from '@smythos/sdk';

async function main() {
    // Create an intelligent agent
    const agent = new Agent({
        name: 'Article Writer',
        model: 'gpt-4o',
        behavior: 'You are a copy writing assistant. The user will provide a topic and you have to write an article about it and store it.',
    });

    // Add a custom skill that combines multiple AI capabilities
    agent.addSkill({
        id: 'AgentWriter_001',
        name: 'WriteAndStoreArticle',
        description: 'Writes an article about a given topic and stores it',
        process: async ({ topic }) => {
            // VectorDB - Search for relevant context
            const vec = agent.vectordb.Pinecone({
                namespace: 'myNameSpace',
                indexName: 'demo-vec',
                pineconeApiKey: process.env.PINECONE_API_KEY,
                embeddings: Model.OpenAI('text-embedding-3-large'),
            });

            const searchResult = await vec.search(topic, {
                topK: 10,
                includeMetadata: true,
            });
            const context = searchResult.map((e) => e?.metadata?.text).join('\n');

            // LLM - Generate the article
            const llm = agent.llm.OpenAI('gpt-4o-mini');
            const result = await llm.prompt(`Write an article about ${topic} using the following context: ${context}`);

            // Storage - Save the article
            const storage = agent.storage.S3({
                /*... S3 Config ...*/
            });
            const uri = await storage.write('article.txt', result);

            return `The article has been generated and stored. Internal URI: ${uri}`;
        },
    });

    // Use the agent
    const result = await agent.prompt('Write an article about Sakura trees');
    console.log(result);
}

main().catch(console.error);

Architecture Highlights

Built-in Security

Security is a core tenant of SRE. Every operation requires proper authorization through the Candidate/ACL system, ensuring that agents only access resources they are permitted to.

const candidate = AccessCandidate.agent(agentId);
const storage = ConnectorService.getStorageConnector().user(candidate);
await storage.write('data.json', content);

Development to Production Evolution

Your business logic stays identical while infrastructure scales: When you use the SDK, SmythOS Runtime Environment will be implicitly initialized with general connectors that covers standard agent use cases.

Click to expand: Basic SRE Setup - Default development configuration

// you don't need to explicitly initialize SRE
// we are just showing you how it is initialized internally
// const sre = SRE.init({
//     Cache: { Connector: 'RAM' },
//     Storage: { Connector: 'Local' },
//     Log: { Connector: 'ConsoleLog' },
// });

async function main() {
    // your agent logic goes here
}

main();

But you can explicitly initialize SRE with other built-in connectors, or make your own Use cases :

• You want to use a custom agents store
• You want to store your API keys and other credentials in a more secure vault
• You need enterprise grade security and data isolation
• ...

Click to expand: Production SRE Setup - Enterprise-grade configuration with custom connectors

const sre = SRE.init({
    Account: { Connector: 'EnterpriseAccountConnector', Settings: { ... } },
    Vault: { Connector: 'Hashicorp', Settings: { url: 'https://vault.company.com' } },
    Cache: { Connector: 'Redis', Settings: { url: 'redis://prod-cluster' } },
    Storage: { Connector: 'S3', Settings: { bucket: 'company-ai-agents' } },
    VectorDB: { Connector: 'Pinecone', Settings: { indexName: 'company-ai-agents' } },
    Log: { Connector: 'CustomLogStore'},
});


async function main() {
    // your agent logic goes here
}

main();

Component System

40+ production-ready components for every AI use case. These components can be invoked programmatically or through the symbolic representation of the agent workflow (the .smyth file).

• AI/LLM: GenAILLM, ImageGen, LLMAssistant
• External: APICall, WebSearch, WebScrape, HuggingFace
• Data: DataSourceIndexer, DataSourceLookup JSONFilter
• Logic: LogicAND, LogicOR, Classifier, ForEach
• Storage: LocalStorage, S3
• Code: ECMAScript, ServerlessCode

Key Features

Feature	Description
Agent-Centric	Built specifically for AI agent workloads and patterns
Secure by Default	Enterprise-grade security with data isolation
High Performance	Optimized for high-throughput AI operations
Modular	Swap any component without breaking your system
Observable	Built-in monitoring, logging, and debugging tools
Cloud-Native	Runs anywhere - local, cloud, edge, or hybrid
Scalable	From development to enterprise production

Contributing

We welcome contributions! Please see our Contributing Guide and Code of Conduct.

Contributors

<i