Think of a restaurant menu. You do not walk into the kitchen; you order from a fixed list, in a format the kitchen understands, and food comes back. An API works the same way for software. A weather app does not run its own satellites. It sends a request like “forecast for Seattle” to a weather service’s API and gets back structured data it can display.
The U.S. National Institute of Standards and Technology describes an API as an access point with “well-defined syntax” that programs use to get “well-defined functionality” (NIST). The key idea is the contract: the provider promises that a request in a certain shape gets a response in a certain shape, so the two sides can be built by different teams without knowing each other’s internals.
This is how most AI products reach a model. A customer-support app, a coding assistant or a spreadsheet plugin usually does not run the model itself. Its code sends an HTTP request to a provider such as Anthropic or OpenAI with an API key, the model name and the conversation so far, and the provider returns the model’s reply (Claude API documentation). Settings like temperature and limits on output tokens travel in the same request, and the provider usually bills per token. Agents use APIs in the other direction too: through tool use, a model asks the surrounding program to call an API, such as a calendar or a database, on its behalf. The Model Context Protocol is one standard way to describe those tools to a model.
Two episodes show why the concept matters on the show. In episode 5 at 2:00, Galileo CEO Vikram Chatterji says some 30 million developers “can just invoke these LLMs with just an API call,” which is why software engineers suddenly need evaluation habits that used to belong to data scientists. Galileo sponsored Season 1, when Conor worked there; the episode’s Galileo read is at 20:00. In episode 63 at 2:20, Dhruv Batra argues the opposite limit: most websites, like a city’s tennis-court booking page, will not add APIs for agents any time soon, because “infrastructure takes extremely long to evolve and change.” That is the case for computer use, where an agent operates the human interface instead.
People often say “API” loosely to mean “the paid way to use a model from code,” as opposed to the chat app. That usage is fine; just remember the API is the interface, not the model behind it.
Hear it from the guest
“There are 30 million developers out there that now can just. Invoke these LLMs with just an API call.”
“Infrastructure takes extremely long to evolve and change.”
Quotes lightly edited to remove filler words.
Sources
- NIST CSRC glossary: application programming interface — Defines an API as a system access point or library function with well-defined syntax, accessible from application programs or user code to provide well-defined functionality; checked October 24, 2026.
- Claude API documentation: Messages API — Official reference showing a model call as an HTTP request with an API key, a model name and a list of messages, returning the model's response; checked October 24, 2026.
- Chain of Thought, episode 5: Vikram Chatterji on developers calling LLMs — Chatterji says millions of developers can now invoke LLMs with an API call, which brings data-science experimentation to software engineers.
- Chain of Thought, episode 63: Dhruv Batra on the web without APIs — Batra argues most websites will not add APIs for agents soon, so agents need to operate interfaces built for people.