AgentFrameRelay

Define agents and tools independently from the framework that executes them.

MVP

• Framework-neutral @tool
• Pydantic/JSON schemas
• Framework-neutral Agent
• Runtime adapter interface
• LangGraph and CrewAI agent runtimes
• OpenAI Agents and Google ADK agent runtimes
• LiteLLM agent runtime and OpenAI-compatible tool schemas
• MCP/FastMCP tool registration
• Native escape hatch
• Capability discovery
• Optional dependencies

Quick start

from agentframerelay import Agent, tool

@tool
def search_customer(customer_id: str) -> dict:
    """Retrieve a customer."""
    return {"customer_id": customer_id}

agent = Agent(
    name="customer-agent",
    instructions="Help with customer lookup.",
    tools=[search_customer],
    runtime="mock",
)

print(agent.run({"input": "Find customer 123"}))

Install an adapter only when needed:

pip install "agentframerelay[langgraph]"
pip install "agentframerelay[crewai]"
pip install "agentframerelay[openai]"
pip install "agentframerelay[litellm]"
pip install "agentframerelay[google-adk]"
pip install "agentframerelay[mcp]"

The core is intentionally framework-neutral. Adapters isolate framework API churn.

Core execution API

Agent.run() always returns an AgentResult with the native framework output, the runtime name, and optional metadata. RuntimeResult remains available as a backward-compatible alias.

Tools validate annotated inputs before direct execution and raise public, framework-neutral exceptions such as ToolInputError and ToolExecutionError. Asynchronous tools support await tool.ainvoke(...); calling tool.invoke(...) from synchronous code runs them safely when no event loop is active.

from agentframerelay import tool

@tool
async def fetch_customer(customer_id: int) -> dict:
    """Retrieve a customer."""
    return {"customer_id": customer_id}

customer = await fetch_customer.ainvoke("123")

Tool portability

An AgentFrameRelay tool retains its original typed Python function and can be adapted for each supported tool protocol:

@tool
def add(a: int, b: int) -> int:
    """Add two integers."""
    return a + b

langchain_tool = add.to_langchain()
crewai_tool = add.to_crewai()
litellm_schema = add.to_litellm()
google_adk_tool = add.to_google_adk()
openai_agents_tool = add.to_openai_agents()
native_mcp_function = add.to_mcp()

For FastMCP, register the original function directly through the relay tool:

add.register_mcp(mcp)

The Google ADK runtime creates an in-memory session by default. Pass user_id, session_id, state, session_service, or run_config to Agent.run() when your application needs to control ADK session handling. MCP is a tool protocol integration, not an Agent runtime.

Integration examples

Credential-backed smoke scripts live in examples/integrations/; they are kept separate from the automated tests/ suite. The LiteLLM, OpenAI Agents, and Google ADK examples load credentials from .env and assert that the original AgentFrameRelay tool function executed.