API

Programs constantly need to talk to each other. A weather app on your phone doesn’t have the weather data itself; it fetches it from a remote computer on the internet. For this to work, there needs to be a fixed agreement: What must the request look like, and what does the response look like? This exact agreement is called an API, spelled out as Application Programming Interface, or in German Programmierschnittstelle. So an API isn’t a program you click on, but an access point that one program opens up for other programs. It can be compared to a restaurant’s menu: it lists what you can order, and you don’t need to know how the kitchen behind it works.

Why hardly any app gets by without third-party interfaces

APIs save an enormous amount of work. Anyone building a delivery app doesn’t need to draw their own map or invent their own payment processing. Ready-made services exist for both, which can be integrated via their interface. The result is apps made up of dozens of third-party building blocks, without users ever noticing.

For companies, APIs are also a business model. Providers of AI models usually don’t sell their software as a file to download. Instead, they sell access via an interface and charge per request. Anyone reading news about OpenAI, Google, or Anthropic’s revenues is almost always reading about API business. The advantage for the provider: the model itself stays on its own servers and isn’t copied.

This dependency has a downside, however. If a provider doubles its prices, changes its terms, or shuts down an interface, all products built on top of it grind to a halt. In the tech industry, this is a recurring drama, for instance when Twitter ended its free access in 2023 and many small services vanished instantly.

Request, response, and the key in between

The most common case on the internet works like this: a program sends a request to a specific internet address, the so-called endpoint. This request contains what is wanted, for example the weather for Hamburg. The server responds with structured text, usually in JSON format. This is a notation in which every value has a name, for example temperature with the value 14. This allows the requesting program to reliably parse the response.

To prevent just anyone from accessing it freely, there is usually an API key. This is a long, random string of characters that functions like a password. The provider uses it to identify who is making the request and bills accordingly. That’s why such a key is a genuine secret that has no business being in public code.

In addition, providers limit the number of requests per minute; this is called a rate limit. It protects the servers from being overloaded. A common misconception, by the way, is that an API is the same thing as a website. A website delivers images and text for humans, while an API delivers raw data for machines.

APIs in everyday life and in the AI business

You encounter APIs practically every day without seeing them. The embedded map on a restaurant’s website, logging in with your Google account, the current stock price in a finance app: behind each of these is an interface to a third-party provider.

In the AI field, the API is the central sales channel. Almost every tool that uses a language model in the background sends the text via an interface to a large provider and gets the response back. The alternative would be to run an open model on your own hardware. That is more effort, but it makes you independent and keeps the data in-house.

What’s new is that AI systems are now using APIs themselves. An assistant that schedules an appointment or books a ticket calls on the interface of the respective service to do so. In technical texts, this is called function calling or tool use. This turns the API into the link between the language model and the rest of the software.