ECP Python SDK

Python SDK for building agents that comply with the Evaluation Context Protocol (ECP).

ECP is a vendor-neutral protocol for testing agent outputs, tool calls, and evaluator-visible audit context across frameworks, models, eval platforms, and CI systems.

Install

pip install "ecp-sdk==0.9.0"

Framework extras:

pip install "ecp-sdk[langchain]==0.9.0"
pip install "ecp-sdk[crewai]==0.9.0"
pip install "ecp-sdk[llamaindex]==0.9.0"
pip install "ecp-sdk[pydanticai]==0.9.0"

Usage

from ecp import Result, agent, on_step, serve


@agent(name="MyAgent")
class MyAgent:
    @on_step
    def step(self, user_input: str):
        return Result(
            public_output=f"Echo: {user_input}",
            evaluation_context="Echoed the input for evaluation.",
            tool_calls=[{"name": "echo", "arguments": {"text": user_input}}],
        )


if __name__ == "__main__":
    serve(MyAgent())

evaluation_context is the preferred field for evaluator-safe audit evidence. private_thought is still accepted as a deprecated compatibility alias.

Async lifecycle hooks

@on_step and @on_reset may be synchronous functions or async def coroutines. The SDK awaits async hooks on one persistent event loop, so loop-bound clients and other async resources can be reused across requests.

import asyncio

from ecp import Result, agent, on_reset, on_step, serve


@agent(name="AsyncAgent")
class AsyncAgent:
    @on_step
    async def step(self, user_input: str) -> Result:
        await asyncio.sleep(0.01)
        return Result(public_output=f"Processed: {user_input}")

    @on_reset
    async def reset(self) -> None:
        await asyncio.sleep(0)


if __name__ == "__main__":
    serve(AsyncAgent())

The same hooks work with serve_http(...). See examples/async_python_demo for a runnable agent that supports either transport.

Streamable HTTP

Agents can also run as an ECP Streamable HTTP server:

if __name__ == "__main__":
    ecp.serve_http(MyAgent(), host="127.0.0.1", port=8765, path="/ecp")

The endpoint accepts JSON-RPC POST requests at /ecp. It returns JSON for requests, 202 Accepted for notifications, and 405 Method Not Allowed for GET SSE streams until ECP defines server-initiated messages.

Testing Your Own Adapter

The bundled adaptors are a convenience, not a requirement - any framework works if you return a Result. If you write your own, ecp.testing is the same harness the built-in adaptors use, so you can pin its behaviour without installing your framework in CI or holding API keys:

from ecp.testing import record_fixture, replay

# Once, with the framework installed and credentials set:
fixture = record_fixture("crewai", crew, "What is 15 multiplied by 8?")
fixture.save("tests/fixtures/crew_calculator.json")

# Then in CI, offline:
from ecp.testing import AdapterFixture, diff_expected
assert diff_expected(AdapterFixture.load("tests/fixtures/crew_calculator.json")) == []

record_fixture captures the framework's real response object and the Result your adapter produced from it. replay re-runs the adapter against that capture. When a framework reshapes its internals, your tool_calls would otherwise go quietly empty and every tool_usage grader would fail as if the agent had regressed - this turns that into a clear test failure.

Links

• Documentation: https://evaluationcontextprotocol.io/
• Repository: https://github.com/evaluation-context-protocol/ecp
• Issues: https://github.com/evaluation-context-protocol/ecp/issues