How AI Agents Use APIs: A Plain-Language Guide

Guides · Updated 2026-11-03

A useful analogy: the agent as a very fast employee with a phone

Imagine hiring an assistant who does everything by calling businesses on your behalf. The assistant can only call numbers that are published, must identify themselves at the start of every call, can only call so many times an hour, and pays a small fee per call to some businesses. That is almost exactly an agent using APIs:

• The published phone numbers are API endpoints. A company decides which actions outsiders may request — "search flights," "create invoice," "transcribe this audio" — and publishes them. Anything not published cannot be requested, no matter how clever the agent is. This is why agents sometimes say "I can't do that": there is no doorway for it, or the doorway is closed to agents.
• Identifying yourself is the API key. Every request carries a key — a long secret string tied to an account. The key is how the company knows who is asking, what plan they pay for, and what they are allowed to do. Whoever holds the key holds the account's powers, which is why keys are treated like passwords and never posted in public.
• "So many calls an hour" is the rate limit. Companies cap how fast one account may ask, to keep their systems stable and fair. Exa's published search limit is 10 queries per second on starter use (up to 25 on pay-as-you-go); Tavily publishes 100 requests per minute on development and 1,000 on production. Hit the cap and the company replies "slow down" (an error called 429) and the agent must wait and retry.
• The per-call fee is usage pricing. Many agent tools charge per request: a web search might cost a fraction of a cent, a minute of transcription a few tenths of a cent at the cheapest published rates. Tiny per call, but agents make many calls, which is why every serious comparison (including this site's dataset) records the meter, not just the monthly sticker.

What an API call actually looks like

Strip away the code and a call has five parts. The agent chooses a doorway (for example, a search API's "search" endpoint). It writes a structured request: the question, how many results, how fresh they must be. It attaches its key. The company's computers run the request and return a structured answer — not a web page designed for human eyes, but tidy data (JSON) with fields like title, link, and text snippet. Finally, the agent reads the answer and decides what to do next: quote it, follow a link, call another API, or report back to you.

That last step is what makes an agent different from a script. A script follows fixed steps; an agent loops — ask, read, decide, ask again — until the job is done or it hits a limit. A question like "find out why sales dipped last week" might become forty API calls across a database, a spreadsheet, and a search engine. Forty calls is why rate limits and per-call pricing are not fine print; they are the physics of the system.

Keys, permissions, and why agents get blocked

Because a key equals power, good setups scope it: a key that can read a calendar should not be able to send email or spend money, and agent platforms increasingly issue keys that act as a specific user with only that user's permissions. Two failure modes show up constantly in real agent work:

The front door refuses agents. Some companies block automated callers at their normal website (agents then need the official API, if one exists — this is why "does it have an API?" is a scored field in our dataset). The key is missing or revoked. Accounts lapse, free tiers end, cards expire; the agent fails not because it is incapable but because its badge stopped working. When an agent says it cannot reach a service, one of these two is usually the true story.

Rate limits, quotas, and the monthly bill

Three different governors get confused, so separate them. A rate limit is speed: requests per second or minute, protecting the company's servers moment to moment. A quota is volume: how many calls your plan includes this month (750 tasks on a Zapier plan, 10,000 operations on a Make plan). Pricing is what happens beyond the quota: pay-as-you-go tools like search and transcription APIs simply keep metering, while subscription tools may pause you or upsell you. Well-built agents are taught to respect all three — backing off when told to slow down, and surfacing costs instead of silently running up a meter. When you read that an agent "ran up a bill," this is the mechanism: thousands of small, individually cheap calls, compounding.

MCP: the standard plug

Until recently, every tool needed a custom integration for every agent framework — the same "twenty chargers in a drawer" problem phones had before USB-C. MCP (Model Context Protocol) is the shared standard: a tool publishes an MCP server describing what it can do, and any MCP-compatible agent can plug in and use it, with the tool list readable by the agent itself. That is why this site tracks MCP support as a first-class field: as of this dataset's verification, all six major search APIs offer official MCP servers, most transcription APIs do, and platforms like Zapier, Make, and n8n expose their entire app libraries through MCP. The practical effect for you: agents get more capable without you hiring anyone to wire integrations — the tools come pre-plugged.

What this means if you are buying tools for an agent

Five questions, in the order our dataset records them. Does it have an API? If not, an agent mostly cannot use it. What is the entry price, honestly metered? Convert units to your workload before comparing. Is there a free tier? Five of the six search APIs and most transcription APIs offer one big enough to test for real. Are rate limits published? Published limits mean the vendor expects programmatic use and lets you design around reality. Does it support MCP? If yes, integration is an afternoon, not a project. Ask those five of any tool — generator of invoices or generator of electricity — and you will know more than most comparison pages will tell you.

A worked example: "book the cheapest flight Friday"

Watch the pieces move together. The agent receives your request in plain words and plans: it needs flight data, so it calls a travel API's search doorway with dates and cities, key attached. The API returns structured options — times, prices, flight numbers — not a web page. The agent compares them (that reasoning is the AI part), picks the cheapest that fits your stated rules, and calls a booking doorway with the flight ID and traveler details from another system. Each call was small, metered, and rate-limited; the search alone might have fanned out into a dozen calls for nearby airports and return options. If the airline's site had no API, the agent would be stuck at the front door — forced to drive a browser like a person, slowly and brittlely, or to report that it cannot complete the booking. That single difference — API or no API — is most of what separates an agent that feels magical from one that feels like a macro. It is also why directories of API-ready tools (including this site's dataset) exist: builders choose tools their agents can actually reach.

FAQ

What is an API, in plain English?

A published set of doorways into a company's software that other programs — including AI agents — are allowed to use: send a structured request through a doorway, get a structured answer back. If a service has no API, an agent generally cannot work with it directly.

What is an API key?

A secret string that identifies an account on every request — part ID badge, part password. It determines what the caller may do and who pays. Anyone holding the key can act as the account, so keys are kept secret and scoped to the smallest necessary permissions.

What happens when an agent hits a rate limit?

The service replies with a "slow down" error (HTTP 429) and refuses further requests until the window resets. Well-behaved agents wait and retry; the fix for chronic limits is a higher plan tier or a less chatty design — Exa publishes 10 queries/second for search at starter level, Tavily 100 requests/minute on development (1,000 on production), as examples of stated caps.

How much does it cost to run an agent?

It is a stack of small meters: the AI model itself, plus every tool it calls — searches at fractions of a cent to a cent each, transcription from roughly $0.12-$0.46 per published hour rates in our dataset, automation steps by plan quotas. A light personal agent might cost a few dollars a month in tool calls; an autonomous research agent can cost hundreds. The meters, not the headline subscriptions, decide.

What is MCP and why does everyone mention it?

Model Context Protocol: a standard way for tools to describe themselves to agents, so one plug fits many tools. Instead of custom code per integration, an MCP-compatible agent reads a server's tool list and starts using it. It is the current standard this site's dataset tracks for every tool.