# OpenAI Agents SDK Integration

> Expose MCP tools natively to the OpenAI Agents SDK via MCPServerStdio. Real (not example) integration shipped in v0.3.0; v0.4.0-rc.2 release candidate.

The OpenAI Agents SDK consumes Attestix as a native MCP server. MCP tools are discovered automatically, with no per-tool wrapping.

## Install

```bash
pip install attestix[openai-agents]
```

Verified against `openai-agents` SDK v0.14.x.

## Minimal wiring

`MCPServerStdio` is an async context manager. Wrap it in `async with` so the server handle closes cleanly when the agent run ends.

```python
import asyncio
from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async def main():
    async with MCPServerStdio(
        name="Attestix",
        params={"command": "attestix", "args": ["mcp"]},
    ) as attestix:
        agent = Agent(
            name="quarterly-analyst",
            instructions="You are a financial analyst.",
            mcp_servers=[attestix],
        )
        result = await Runner.run(agent, input="Run Q4 analysis")
        print(result.final_output)

asyncio.run(main())
```

## Restricting which tools are exposed

Use `tool_filter` with `create_static_tool_filter` to scope the Attestix tool surface for a given agent.

```python
from agents.mcp import MCPServerStdio, create_static_tool_filter

attestix_filter = create_static_tool_filter(
    allowed_tool_names=[
        "create_agent_identity",
        "issue_credential",
        "record_training_data",
    ],
)

async with MCPServerStdio(
    name="Attestix",
    params={"command": "attestix", "args": ["mcp"]},
    tool_filter=attestix_filter,
    cache_tools_list=True,
) as attestix:
    agent = Agent(
        name="compliance-agent",
        mcp_servers=[attestix],
    )
```

## Auto-issued audit entries

Every tool call made through the MCP server is appended to the Attestix hash-chained audit trail. No extra code needed on the SDK side.

## Streaming and tool-use events

`Runner.run_streamed` returns a `RunResultStreaming`. Iterate its `stream_events` async iterator.

```python
result = Runner.run_streamed(agent, input="Run Q4 analysis")
async for event in result.stream_events():
    if event.type == "run_item_stream_event" and event.name == "tool_called":
        print(event.item)
```

Audit-trail entries can be inspected after the run via the Attestix CLI or by calling the audit service directly in your own code.

## HTTP transport

If the Attestix MCP server runs as a separate service, use `MCPServerStreamableHttp` (SDK v0.14 name). Wrap it in `async with`.

```python
from agents.mcp import MCPServerStreamableHttp

async with MCPServerStreamableHttp(
    name="Attestix",
    params={"url": "http://localhost:8501/mcp"},
) as attestix:
    agent = Agent(
        name="quarterly-analyst",
        mcp_servers=[attestix],
    )
```

Start the Attestix MCP server in HTTP mode with:

```bash
attestix mcp --transport http --port 8501
```

## Multi-agent orchestration

Wire Attestix into every agent in an orchestration so the entire graph is attestable. Role hand-offs are recorded as first-class audit events.

```python
async with MCPServerStdio(
    name="Attestix",
    params={"command": "attestix", "args": ["mcp"]},
) as attestix:
    analyst = Agent(name="analyst", mcp_servers=[attestix])
    summarizer = Agent(name="summarizer", mcp_servers=[attestix])
    supervisor = Agent(
        name="supervisor",
        mcp_servers=[attestix],
        handoffs=[analyst, summarizer],
    )
    await Runner.run(supervisor, input="Prepare the Q4 board pack.")
```

UCAN delegations make the hand-off chain verifiable.

## Sample repository

Full runnable example at [github.com/VibeTensor/attestix/tree/main/examples](https://github.com/VibeTensor/attestix/tree/main/examples).