OpenAI Agents SDK

The OpenAI Agents SDK is built on the OpenAI client, so pointing it at Mesh is a base-URL change. Mesh serves both the Responses API and chat completions, so either path works — unlike most OpenAI-compatible gateways, you do not have to force the SDK off its default.

Install

pip install openai-agents

Connect

Default (Responses API)
Chat completions

import os

from agents import (
    Agent,
    set_default_openai_api,
    set_default_openai_client,
    set_tracing_disabled,
)
from openai import AsyncOpenAI

set_tracing_disabled(True)

client = AsyncOpenAI(
    base_url="https://api.meshapi.ai/v1",
    api_key=os.environ["MESH_API_KEY"],    # rsk_...
)
set_default_openai_client(client)
set_default_openai_api("responses")        # the SDK's own default

agent = Agent(name="support", instructions="Be concise.", model="auto")

Mesh serves POST /v1/responses, so the SDK’s native path works. A model without Responses support returns 400 model_capability_not_supported; auto resolves to one that supports it.

import os

from agents import Agent, OpenAIChatCompletionsModel, set_tracing_disabled
from openai import AsyncOpenAI

set_tracing_disabled(True)

client = AsyncOpenAI(
    base_url="https://api.meshapi.ai/v1",
    api_key=os.environ["MESH_API_KEY"],    # rsk_...
)

agent = Agent(
    name="support",
    instructions="Be concise.",
    model=OpenAIChatCompletionsModel(model="auto", openai_client=client),
)

Pin this wrapper when you want POST /v1/chat/completions specifically — for example to use a model that has no Responses support.

Call set_tracing_disabled(True), or set OPENAI_AGENTS_DISABLE_TRACING=1.The SDK uploads traces to OpenAI’s own backend by default, authenticated with OPENAI_API_KEY. Pointing the SDK at Mesh does not redirect tracing. Left on without an OpenAI key it logs export errors on every run; left on with one, it ships your agents’ prompts and tool calls to OpenAI.

Tool calling

from agents import Runner, function_tool


@function_tool
def get_shipment_status(order_id: str) -> dict:
    """Look up the shipment status of an order."""
    book = {"A-1042": {"state": "in transit", "eta_days": 2}}
    if order_id in book:
        return {"status": "success", "order_id": order_id, **book[order_id]}
    return {"status": "error", "error_message": f"No order {order_id}."}


agent = Agent(
    name="support",
    instructions="Use the tool for order questions. Be concise.",
    model="auto",
    tools=[get_shipment_status],
)

result = await Runner.run(agent, "What's the shipment status of order A-1042?")
print(result.final_output)

result.new_items carries the ToolCallItem and ToolCallOutputItem entries if you need to inspect the tool exchange. Handoffs and guardrails work unchanged — they are SDK-side features and never reach the model API.

Caveats

Everything in Caveats that apply to every framework applies here.

Agent frameworks — the shared pattern and compatibility matrix
Auto-routing — how auto picks a model
Authentication — rsk_ keys versus mak_ admin keys

MCP

Frameworks

OpenAI Agents SDK

Install

Connect

Tool calling

Caveats

​Install

​Connect

​Tool calling

​Caveats

​Related

Install

Connect

Tool calling

Caveats

Related