Agents

Agents have emerged as the go-to framework for coordinating diverse, specialized LLM applications to collaborate effectively. Numerous real-world scenarios demand external tools to construct intricate workflows for these LLM-driven agents.

Model Context Protocol (MCP) has swiftly risen as the open standard for developing Agentic systems. It enables straightforward integration of standard tool services and ensures seamless interoperability among models throughout the AI ecosystem.

What is MCP?

Model Context Protocol (MCP) is an open protocol crafted to standardize interactions between AI models—particularly large language models (LLMs)—and external tools, data sources, and context providers in a secure, modular, and composable manner. MCP delivers a unified framework for transmitting structured requests from an agent or application to various “tool services,” such as databases, APIs, or custom logic modules. By implementing MCP, developers can:

• Decouple agent logic from tool implementations: Agents can invoke tools (like a database or search service) via a standard protocol, avoiding hardcoded integrations.

• Enforce consistent security and governance: MCP specifies authentication, authorization, and data boundary controls between the model and external resources.

• Support modular, reusable agent architectures: Tools can be exchanged, updated, or expanded without altering the agent code, simplifying the evolution of complex workflows.

• Run tools locally or remotely: The same protocol functions whether a tool operates in the customer’s environment or the cloud, accommodating privacy and data residency needs.

MCP serves as the “middleware” connecting AI models to the external world, facilitating secure, adaptable, and sustainable integration of real-world context and capabilities into conversational or autonomous agents.

Agents in the enterprise

In the modern enterprise environment, conversational agents—particularly those powered by voice—are rapidly establishing themselves as essentials for customer support, internal helpdesks, and task automation. However, developing robust, scalable voice agents poses challenges due to disjointed tooling, integration hurdles, and the necessity for dependable orchestration of backend systems. A prevalent pattern in enterprises involves creating agents supported by knowledge bases (both structured and unstructured). These bots fall into categories such as:

• copilots for internal use, and

• customer-facing assistants.

The latter, customer-facing assistants, demand superior accuracy, usability, and design. Moreover, a frequent need for customer-facing chatbots is incorporating voice as a user interface modality (e.g., for phone call automation).

These Q&A chatbots serve a broad array of industries: healthcare, government, legal, and others requiring straightforward knowledge retrieval at users' fingertips.

One prime example is the insurance industry, where we've observed substantial value for our customers in this sector. Insurance policies are intricate, and system navigation can be challenging for policyholders.

What's in this Cookbook?

This cookbook offers a complete, modular recipe utilizing MCP to build voice-enabled agents with the OpenAI Agents SDK. We highlight its application for dynamic context management and agentic tool-calling. We illustrate the system's potential through the insurance use case mentioned earlier. In this example, we showcase MCP integration with various tools relevant to your applications. In particular, we feature custom MCP servers (for text retrieval and web search) alongside predefined MCP servers (for SQLite).