AI Tools and Protocols: How AI Connects to the Outside World

An AI model on its own is like a very well-read person locked in a room with no phone and no internet. It knows a lot, but only what it learned before the door shut. It can’t check today’s weather, look up your order, or open a file on your computer.

The Tools & Protocols group on the Periodic Table of AI is about opening that door. It’s the six cyan tiles, and each one is a way for AI to reach past its training and connect to fresh facts and real software.

A glowing cyan sphere linked by cables of light to a magnifying glass over a globe, a cabinet of index cards, a wrench, and a plug clicking into a socket
Search, documents, tools and a universal plug: the ways AI reaches past its training.

The last post covered the model families and who makes them. This one is about how those models plug into everything else.

The connections

API

An API, short for application programming interface, is a way for one program to talk to another. When an app shows you a weather forecast, it’s usually asking a weather service’s API for the numbers.

In AI, the API is how most software uses a model. A developer’s program sends a prompt to the AI company’s servers and gets the answer back, without anyone opening a chat window. A lot of the AI features you see inside other apps run this way, through an API, behind the scenes.

Function Calling

Function calling lets an AI ask for a tool to be used. The developer describes the tools that are available, like “check the weather” or “look up an order.” When a question needs one, the model doesn’t just guess. It replies with a neat, structured request saying which tool to use and with what details. The developer’s code runs the tool and hands the result back, and the model uses it to finish its answer.

OpenAI added function calling to its API in June 2023, and the idea has since spread across the industry. The model never pushes the buttons itself. It asks, and the software around it decides whether to act. When a tool can do something real, like sending an email or making a purchase, it’s smart to have a person confirm it first.

MCP

MCP stands for Model Context Protocol. Anthropic introduced it in November 2024 as an open standard for connecting AI assistants to the places where data lives, like business tools, file storage and coding environments.

Here’s the problem it solves. Before MCP, every AI app needed custom code for every tool it wanted to connect to. Ten apps and ten tools meant up to a hundred separate connections. With a shared standard, a tool builds one MCP connection and works with any AI app that speaks MCP. People often compare it to a universal plug for AI.

In December 2025, Anthropic donated MCP to the Linux Foundation as a founding project of the new Agentic AI Foundation, which Anthropic co-founded with Block and OpenAI. By then, Anthropic reported more than 10,000 active MCP servers, with support in ChatGPT, Claude, Gemini, Microsoft Copilot and other major platforms.

Fresh information

Web Search

Web search gives an AI a way to look things up online while it answers. Without it, a model only knows what was in its training data, which stops at a certain date. With it, the AI can find today’s news, current prices, or a page published last week, and it can point you to its sources so you can check them.

It’s worth clicking those sources. A search can turn up a bad page as easily as a good one, and an AI can still misread what it found.

RAG

RAG stands for retrieval-augmented generation. The idea is simple: before the AI answers, find the most relevant documents and give them to the AI along with the question. The AI then answers from those documents instead of from memory alone.

The name comes from a 2020 paper by researchers at Facebook AI Research, University College London and New York University. They found that models built this way gave more specific and more factual answers than models relying only on what they’d learned. Another big plus is that you can update what the AI knows just by changing the documents, with no retraining needed.

RAG is how many companies build chatbots that answer from their own help articles, policies and manuals. It’s a close cousin of Grounding from the Prompting post, and one of the main tools for cutting down on hallucination.

Vector Database

A vector database is the filing cabinet that makes RAG fast. It stores documents as embeddings, the long lists of numbers that capture meaning, which we covered in the Core Concepts post.

Because similar meanings get similar numbers, a vector database can search by meaning instead of by exact words. Ask “How do I get my money back?” and it can find a help article titled “Refund policy,” even though the two share no words. Regular keyword search would miss that.

How they fit together

Picture a company’s support chatbot. You ask about a late order. The chatbot’s software reaches the model through an API. The model uses function calling to request an order lookup, and the order system might be connected through MCP. To explain the shipping policy, the system uses RAG, pulling the right policy page from a vector database. If you ask about a carrier delay in the news, web search fills in what the company’s files don’t know.

Each piece does one job. Together, they turn a model that only knows its training data into an assistant that can work with real, current information.

Where to go next

Once AI can use tools, the next step is letting it use them on its own, one step after another, to finish a whole task. That’s the next group on the chart, Agents & Automation, and it’s the next post.

For now, click any tile above and read the full page. If you only read one, make it RAG, since it’s behind so many of the AI assistants you’ll run into at work.

1 thought on “AI Tools and Protocols: How AI Connects to the Outside World”

Comments are closed.