MCP is trending

Introduced by Anthropic in late 2024, Model Context Protocol (MCP) is a standardized syntax to inform AI agents / AI assistants on how to use external tools and services, like databases and APIs.

MCP has quickly become the talk of the town, with the CEOs of OpenAI, Microsoft and Google rushing to announce their support.

If you’re an entrepreneur or leader embedding AI capabilities into your products and systems, here are 4 main things to know about MCP.

1. Assistant tools are crucial to integrate AI into business applications.

2. Until today, creators of AI business agents had to write dozens of code lines to instruct the agent on using each tool. Given that many agents use the same tools (e.g., Google Calendar, Airtable), there must be a better way.

3. MCP enables a new, standardized syntax for telling an AI agent how to use an external tool or service. It accelerates tool integration for AI agent creators, as they can reuse code published by others to obtain tool specifications .

4. While MCP can speed up software development, it doesn’t fundamentally change what agents can do. If your company is developing agents, let the devs decide if and when to use MCP. But if you are a vendor wanting AI agents to use your service, ensure that there are publicly available MCP resources to facilitate the integration of your offerings.

Let’s review each of these insights in more detail.

Assistant tools are essential to integrate AI into business applications

Let’s consider an AI assistant for a sales development representative (SDR).

This AI assistant handles the following tasks:

• Collect customer form submissions from various sources (e.g., information requests, events) by reading internal databases and spreadsheets.

• Populate these contacts with additional information (e.g., existing or new customer, industry segment, number of employees) from internal databases (e.g., Google Contacts, CRM) and external services (e.g., Apollo, CoreSignal, Crunchbase).

• Prioritize these contacts based on business criteria.

• Email high-priority contacts to propose a follow-up call with an SDR.

• Check the SDRs’ calendar availabilities to finalize the appointment.

• After the call, follow up with the SDR to collect feedback and route the potential lead to the appropriate channel (e.g., self-service, account manager).

Each task requires the assistant to send queries to internal or external services, called “tools” in the AI assistant context.

Most tools involve the AI assistant calling an API with credentials and parameters and interpreting the response.

Without these tools, AI assistants would be ineffective for businesses.

If you are developing AI assistants for business, you must absolutely integrate tools.

Without MCP, tool integration takes a bit of time

Let’s consider the SDR AI assistant described above. To schedule meetings, it needs to call the Google Calendar API.

To achieve this, most LLM models (e.g., OpenAI, Anthropic, Gemini, Cohere) require the assistant developer to define 2 tools, “list_events” and “create_event”, and to write a “tools” object for the LLM to understand their functionalities and arguments.

The “tools” object looks like this (for an Anthropic LLM).

In addition to declaring the tools, the developer must code each of the functions that translate tool calls from the AI assistant into API calls using Google Calendar’s specifications (e.g., authentication, parameters).

Honestly, when using AI coding assistants like Cursor, most of the above code can be generated by starting sentences and pressing “tab” multiple times. Most tools can be integrated in minutes. But if you’re doing this work for the first time, your tool definitions may not be perfect.

MCP enables a standardized syntax to tell an AI agent how to use tools

With MCP, the developer of the AI assistant (also called the MCP client) can leverage pre-defined services (also called MCP servers) created by SaaS product vendors or the community.

Instead of declaring and coding the tools, the developer can instruct the AI assistant to call the MCP server, which will:

• Provide the list of available tools and their declaration.

• Translate the AI assistant’s tool calls into appropriate API calls.

The list of MCP servers is growing every day, so MCP is a great resource for AI assistant developers who want to integrate many tools quickly.

For a demo of MCP, let’s use a free MCP server provided by Gumloop. First, create a free Gumloop account and an integration with your Google Calendar account. Then, under “Credentials,” create a guMCP API Key.

In our demo, you can see that the MCP server provides the full declaration of Google Calendar tools. This is comparable to the tools declarations required without MCP, except you did not have to write them down yourself.

The full demo is shown in this Google Colab notebook.

While MCP can speed up software development, it does not fundamentally change what agents can do

From the standpoint of the AI assistant developer, MCP is mainly a time saver, as you can see from the above demo. If someone has already created a MCP server for the tool that you want to integrate, that’s great news! You don’t need to declare the tool’s specifications yourself.

However, MCP is still new, and not all tools have robust MCP servers.

Also, to be honest, the hardest part about tool definition is not the declaration of the tool’s functionalities and arguments; it’s the authentication. For example, if you want the users of your AI-powered product to connect to their own Google Calendar, you need to develop a workflow so that they can log in with their Google email, and then you need to store their Google credentials securely.

Authentication workflows can be complex, and MCP does not make them easier.

So, whether you declare the tool yourself, or use an existing MCP server, is mostly a tech team decision. You can let the devs decide if and when to use MCP.

In many scenarios, you may be better off coding the tools yourself, or using the MCP server to query the tools declaration but handling tool calls yourself so that you can manage authentication appropriately.

However, if you are a vendor and want AI agent developers to use your service as a tool, it is definitely good practice to publish an MCP server yourself, or to work with MCP server operators like Gumloop. This will ensure that other developers can easily integrate your offerings in their AI-powered products.

Exciting times!