Free AI API options and what free means for an agent

You search for a free AI API because you want to try something without a card on file. That is a fair goal. The trouble starts when an agent, not you, makes the calls. A free tier that is fine for ten test requests can stop an agent halfway through a job of two hundred.

This post covers what a free AI API usually gives you, the four places where "free" breaks for an agent, and what to do instead when the agent needs data and tools more than a model. It is about the data and tool side. looot does not sell model inference, so for a free language model API you will read the model vendor's own page.

What does a free AI API usually include?

A free AI API is normally a capped allowance. You get a key, a small number of requests or credits, a low rate limit, and sometimes a reduced set of endpoints. The cap is the product: it lets you test, and it tells you where the paid plan starts.

The caps come in a few forms. A fixed request count per month. A credit balance that does not renew. A rate limit per minute. A restriction to non-commercial use. Each vendor mixes these differently, and the terms change, so read the vendor's current page and note the date you read it.

There are also two different things people mean by the phrase. One is a model API, where you send text and get text back. The other is a data or tool API, where you ask for a search result, a company record or a page of text. Free options exist for both, and the limits differ.

Where does a free tier break for an agent?

A person notices when a quota ends. An agent does not unless you tell it to look. Four things go wrong most often.

The quota ends mid-run. The agent has done eighty of two hundred lookups and the next call returns an error. If the agent was not told what to do, it may retry in a loop or report success with partial data.

The rate limit stalls the job. A cap of a few calls per minute turns a short job into a long one, and long jobs hit the time limits of the agent client.

One key per provider multiplies. Free tiers are per vendor. An agent that needs search, scraping and enrichment ends up with three or four keys, three or four signups and three or four sets of terms.

You cannot see cost, so you cannot cap it. When a free tier ends and a paid one begins silently, the first sign is a bill. An agent can only stop at a limit if it can read the price before it calls.

None of this makes free tiers bad. It means they suit a human trying a tool and suit an agent badly.

What is the alternative to a free tier for an agent?

The alternative is pay per call with the price visible first. You do not get a free allowance, but you also do not get a cliff. The agent reads the price, runs the call, and the balance moves by exactly that amount.

That is how looot works. Searching the catalog and inspecting an endpoint are free. Running an endpoint costs its listed price. There is no trial credit, so you top up before the first run. The agent holds one token, and the gateway holds the provider keys, so you do not collect a key from each vendor.

For a first test, cost stays small because many calls are cheap. The live table below reads the catalog at request time, so the numbers are today's and not a figure typed into a post.

Read the row labels as jobs. The same job can have several providers at different prices, and the agent can choose the cheapest one that returns the fields it needs.

Are there free tools that use these APIs?

Yes, a few. The free tools section runs real lookups that looot pays for, such as a backlink checker and a YouTube transcript tool. They are rate limited, because someone pays for each call. They are meant to let you see the result before you connect an agent, not to replace an API for a job of any size.

If a free tool answers your one-off question, use it and stop there. If you need the same answer for two hundred rows, an agent with pay per call is the better fit, since a free tool will not take a list that size.

How do I test with almost no spend?

Keep the first run to a handful of rows and ask the agent to report the price first. A small test shows the shape of the response, the fields you get and the real cost per row. Then you multiply with a known number, and what 1,000 lookups cost shows how to read a price table that way.

Three habits keep a test cheap. Name a hard limit on the number of calls. Tell the agent to stop and ask when the next call would pass a spend limit. Ask it to print the cost of every call it ran, so you can check the total against the balance.

How do I read a free tier's terms?

Read four lines on the vendor's pricing page and write them down with the date. The request or credit cap and whether it renews. The rate limit per minute. Whether commercial use is allowed. And what happens at the cap: a hard stop, a slower lane or an automatic move to a paid plan.

The last one matters most for an agent. A hard stop is safe, because the agent gets an error and can report it. An automatic paid plan is the risky one. If you cannot find the answer, assume the worst and keep the test small. For how looot prices calls, see pricing, which links to the live catalog.

Try it: a prompt for your agent

This prompt runs one cheap lookup and prints the price before and after. Replace the domain with your own.

Steps 2 and 3 cost nothing. The run in step 4 costs the price you saw in step 3.

What it cannot do

• It does not give free calls. Connecting and searching the catalog are free, and each run is paid.
• It does not sell model inference. For a language model API, use the model vendor.
• It does not remove vendor terms. Public data endpoints still follow each provider's rules, and the catalog entry links the docs.
• It does not run jobs on a timer. A job that must repeat needs scheduled runs, not available yet.

Last checked 2026-10-03 against the live catalog and the looot skill file. Author: the looot team.