rayai 0.1.3

Scalable runtime for Agents, MCP Servers, and coding sandboxes, orchestrated with Ray.

Features

• Distributed Runtime - Resource-aware tool execution on Ray clusters with automatic scaling
• Framework Agnostic - Works with LangChain, Pydantic AI, Agno, or pure Python
• Secure Sandboxing - gVisor-sandboxed environments for AI-generated code execution
• Simple CLI - Initialize projects, create agents, and serve with single commands
• Production Ready - Built on Ray Serve for reliable, scalable deployments

Quick Start

# Install
pip install rayai

# Create a new project
rayai init my_project
cd my_project

# Create your first agent
rayai create-agent my_agent --framework pydantic

# Run locally
rayai up

Your agent is now available at http://localhost:8000/agents/my_agent/

Installation

pip install rayai

With optional sandbox support (for code execution):

pip install rayai[sandbox]

Requirements: Python 3.12+

Usage

CLI Commands

Initialize a Project

rayai init my_project

Creates a project structure with configuration files and an agents/ directory.

Create an Agent

rayai create-agent my-agent
rayai create-agent my-agent --framework langchain
rayai create-agent my-agent --framework pydantic

Supported frameworks: python (default), langchain, pydantic

Run Agents

rayai up                             # Run all agents on port 8000
rayai up --port 9000                 # Custom port
rayai up --agents agent1,agent2      # Run specific agents

Creating Your First Agent

After running rayai create-agent my_agent --framework pydantic, edit agents/my_agent/agent.py:

import rayai
from pydantic_ai import Agent

@rayai.tool(num_cpus=1)
def search(query: str) -> str:
    """Search for information."""
    return f"Results for: {query}"

# Create Pydantic AI agent with Ray-distributed tools
def make_agent():
    return Agent(
        "openai:gpt-4o-mini",
        system_prompt="You are a helpful assistant.",
        tools=[search],
    )

# Serve the agent
rayai.serve(make_agent, name="my_agent", num_cpus=1, memory="2GB")

Run with:

rayai up

Test your agent:

curl -X POST http://localhost:8000/agents/my_agent/ \
  -H "Content-Type: application/json" \
  -d '{"query": "Hello!"}'

API Reference

@rayai.tool Decorator

Creates a Ray-distributed async tool from a function. Works as both a decorator and wrapper for framework tools.

import rayai

# As decorator (uses defaults: num_cpus=1, num_gpus=0)
@rayai.tool
def search(query: str) -> str:
    """Search for information."""
    return f"Results for: {query}"

# As decorator with explicit resources
@rayai.tool(num_cpus=2, memory="4GB")
def expensive_task(data: str) -> str:
    return process(data)

# As wrapper for framework tools
from langchain_community.tools import DuckDuckGoSearchRun
lc_search = rayai.tool(DuckDuckGoSearchRun())

rayai.serve()

Serve an agent via HTTP with Ray Serve. Auto-detects framework (Pydantic AI, LangChain, custom).

import rayai
from pydantic_ai import Agent

def make_agent():
    return Agent("openai:gpt-4", tools=[search])

rayai.serve(make_agent, name="myagent", num_cpus=1, memory="2GB")

rayai.Agent Base Class

For custom agents without a framework:

from rayai import Agent

class MyAgent(Agent):
    tools = [search, analyze]

    async def run(self, query: str) -> str:
        result = await self.call_tool("search", query=query)
        return f"Found: {result}"

rayai.serve(MyAgent, name="myagent")

execute_tools

Execute multiple tools in parallel on Ray:

import rayai
from rayai import execute_tools

@rayai.tool(num_cpus=1)
def tool_1(x: str) -> str:
    """Process input with tool 1."""
    return process_1(x)

@rayai.tool(num_cpus=1)
def tool_2(x: str) -> dict:
    """Process input with tool 2."""
    return process_2(x)

# Execute both tools in parallel on Ray
results = execute_tools([
    (tool_1, {"x": "input_1"}),
    (tool_2, {"x": "input_2"})
], parallel=True)

Examples

See the examples/ directory for complete implementations:

• Token-Efficient Agent - Autonomous code execution in sandboxed environments
• Finance Agent - Multi-step financial analysis with external APIs

Telemetry

RayAI collects anonymous usage data to help improve the CLI:

• Commands run (init, create-agent, up)
• Framework choices (python, langchain, pydantic)

No PII data, project information, or code is collected by telemetry without your explicit approval.

To opt out:

rayai analytics off

Contributing

Contributions are welcome! See CONTRIBUTING.md for guidelines.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

If you find this project helpful, please consider giving it a star!