Connect OpenAI Agents SDK to Skysay

Use OpenAI Agents SDK with hosted Skysay MCP to inspect a workspace and, when separately configured and approved, operate calls and SMS. Start with one read-only tool; no private repository or Twilio account is needed.

Prerequisites

• Python 3.13 and uv; a Skysay workspace.
• A workspace key with account:read, held in SKYSAY_API_KEY through your local environment or secret manager. Never commit it or paste it into chat.
• No OpenAI key is needed for the default transport probe. The optional model-mediated mode needs your own securely configured OPENAI_API_KEY.

Create a scoped credential. Get the public sample.

Install the pinned example

In the public sample's examples/agent-sdks directory:

uv venv --python 3.13.13
uv pip sync --require-hashes requirements.lock
uv run --no-sync python openai_agent.py

The sample pins openai-agents==0.23.1 and claude-agent-sdk==0.2.166; its checked-in hash-locked requirements.lock pins the transitive MCP transport too. If copying only this OpenAI sample, install openai-agents==0.23.1 in your own Python 3.13 virtual environment. The combined sample installs both SDKs for reproducibility.

Connect and verify a tool result

Save the following as openai_agent.py. The default run initializes MCP, limits the model-visible tools to account_overview, and calls that tool directly. It does not request model inference or print account contents.

"""Inspect Skysay through OpenAI Agents SDK. Default mode makes no model call."""
import argparse
import asyncio
import os
from agents import Agent, Runner, set_tracing_disabled
from agents.mcp import MCPServerStreamableHttp, create_static_tool_filter

URL = 'https://api.skysay.ai/mcp/read-only'

def connection_params():
    return {'url': URL, 'headers': {'Authorization': f"Bearer {os.environ['SKYSAY_API_KEY']}"}, 'timeout': 15}

async def main(ask_model=False):
    set_tracing_disabled(True)  # Account/tool content must not enter hosted traces.
    async with MCPServerStreamableHttp(name='skysay', params=connection_params(),
            tool_filter=create_static_tool_filter(allowed_tool_names=['account_overview']),
            max_retry_attempts=0) as server:
        tools = await server.list_tools()
        if [tool.name for tool in tools] != ['account_overview']:
            raise RuntimeError('Expected the account_overview tool; check the endpoint and permissions.')
        result = await server.call_tool('account_overview', {})
        if result.is_error:
            raise RuntimeError('Account read refused; check account:read and the workspace credential.')
        print('Connected: account_overview succeeded. No phone action was performed.')
        if ask_model:
            agent = Agent(name='Workspace reader',
                instructions='Use account_overview to confirm the workspace is accessible. Do not repeat private account details.',
                mcp_servers=[server])
            response = await Runner.run(agent, 'Confirm that my Skysay workspace is accessible.', max_turns=3)
            print(response.final_output)

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--ask-model', action='store_true', help='Optional model request: requires OPENAI_API_KEY and may incur model charges.')
    asyncio.run(main(parser.parse_args().ask_model))

Expected output:

Connected: account_overview succeeded. No phone action was performed.

OpenAI tracing is disabled so workspace tool inputs and results are not sent to hosted traces. A successful read demonstrates this key's account access; it does not prove that a model has used the tool.

Optional: let an agent make the read

Configure your own OpenAI API credential securely, then run:

uv run --no-sync python openai_agent.py --ask-model

This sends tool output to the model provider and may incur model charges. Review the provider's data policy and use a test workspace. Never run this against customer data merely to check installation.

Before calls or SMS

A successful read is not an active phone number. In the workspace, choose a country and number type, inspect voice/SMS capabilities and prices, and complete any required identity, business and carrier-document review. Skysay handles carrier onboarding; you still provide the required information. Number activation is not instant in every market. See regulated numbers and pricing.

Your SDK agent configures and operates Skysay. It is not automatically the voice model speaking to the person on the phone. Configure the voice agent's published flow before enabling real calls.

Keep the initial read-only endpoint. Enabling the full MCP resource is a separate application change: use the least scopes for the selected tools, an explicit destination/action/cost confirmation, and a server-side approval flow before number purchases, calls or SMS. An SDK allowlist is not an authorization or spending boundary. Quickstart explains these boundaries; MCP tools gives exact names and arguments. This guide does not dial, text or purchase a number.

Troubleshooting

• 401: confirm the workspace key is active and present in the launching process. Do not print the key while debugging.
• Tool error / missing scope: the initial key needs account:read. Connection success alone does not prove authorization to call the tool.
• 405 in your browser: MCP uses POST. Opening the endpoint as a web page is not an initialization or tool-call test.
• Timeout: check outbound HTTPS to api.skysay.ai and proxy configuration. Do not disable TLS validation.
• Unexpected tools: confirm the read-only URL and use a clean example directory. Do not inherit unrelated MCP, shell or file tools.

What was verified

The pinned client packages were exercised against the real Skysay HTTP handler in an isolated local workspace with disposable credentials. No production account, customer data, hosted model request, phone call, SMS or number purchase was used. Optional model-mediated runs above remain untested; they incur provider usage. This is transport/configuration acceptance, not a live-carrier or model-quality benchmark.

SDK reference

Configuration checked on 14 October 2026 against the official SDK MCP documentation. Recheck pinned versions before upgrading.