5 min read

In AI tooling, "free" almost always means "runs on your computer"

We counted the pricing model of all 701 tools we list. 84 are free outright, and 55% of those publish their source against 13% of the catalogue. Nothing sold outright is open source at all.

Search for free AI tools and you will get lists of hundreds. Open any of them and most entries have a pricing page.

The word is doing a lot of work. So we counted.

What the catalogue actually contains

Every listing here records how the maker charges, taken from the maker's own pricing page rather than guessed at. Across the 701 tools published on 9 August 2026:

ModelCount
Freemium — free tier, paid plans above it399
Contact us for a price107
Free — the maker charges nothing for the product84
Subscription78
Paid outright29
One-time purchase4

84 free. And 399 more that the internet will cheerfully file under "free".

You can check this yourself: the free ones are on one page, and every listing states its model.

What the free ones have in common

Pricing modelToolsOpen source
Free8455%
Contact us10710%
Freemium3998%
Subscription781%
Paid outright290
One-time purchase40
Whole catalogue70113%

Free tools are four times more likely to publish their source than the catalogue average, and nothing sold outright is open source at all — 33 listings across paid and one-time purchase, not one of them.

The names at the free end are the ones you would guess: Aider, Cline, ComfyUI, Ollama, LibreChat, EasyInvoicePDF. They run on hardware you already pay for.

That is the mechanism. A tool running on somebody else's servers costs that somebody money every time you use it, and inference is not cheap. There are only three ways that ends: you pay, an investor pays for a while, or the free tier is small enough to work as an advert. A tool that runs on your machine has none of that pressure, because the hosting bill is your electricity bill.

The other 38, which is where our first version was wrong

Not every free tool is a local open-source one, and an earlier version of this article said otherwise on the strength of six examples.

38 of the 84 are free and closed: LM Studio and Pieces run locally without publishing source; Namelix, Colormind, Birme and TextFX are small browser utilities where the compute per use is cheap enough not to matter. Free-because-local is the biggest group, not the only one.

Open source is not the same as free, either

Thirty-two of our listings are open source and freemium. The source is published, and the hosted version is the product — Zed, Stability AI, Black Forest Labs, LlamaIndex, CrewAI among them.

That is a legitimate and increasingly common model, and it is worth understanding before you pick: you can self-host and pay nothing but your own time, or you can pay them to run it. What you should not assume is that an open-source badge means the convenient version is free.

How to read a free tier before you commit

Find the limit, not the price. Every freemium tool has a number that ends the free part: messages per month, images per day, seats, projects. That number is the actual offer. If the pricing page does not state one, that is itself the answer.

Check what happens at the wall. Does it stop, queue you, or start charging? The last one is worth knowing before you build a workflow on it.

Ask whether your data comes back out. Free tiers are where export gets restricted. A tool you cannot leave is not free, it is deferred.

For anything you will run daily, look local first. The names above are not compromises — Ollama and ComfyUI are what a lot of paid products are wrapping. Local costs you setup time and a machine that gets warm, and nothing else.

Why we counted

We are a directory, and directories have an obvious incentive to call as much as possible free, because "500 free AI tools" is a better headline than "84".

We would rather the number be useful. Every listing here states its pricing model on the page, taken from the maker's own site, and the free page says plainly that 399 others run a free tier alongside paid plans rather than folding them all into one count.

Rewritten after publishing. The first version of this counted a 155-tool catalogue and found six free tools, five of them open source, and concluded that free "almost always" meant local. The catalogue is now 701 and the free bucket is 84, so the six-tool claim was both stale and too strong — that is what a sample of six buys you. Checking the flag while rewriting turned up two free tools wrongly marked closed, Gitingest (MIT, 15,275 stars) and Skills (MIT, 28,447), which is why the free bucket reads 55% and not 52%. The pattern survived the larger sample and got sharper: 55% against a 13% baseline, and zero across everything sold outright.

If you make one of these and the model on your listing is wrong, tell us — or submit the tool and set it yourself. Listing is free and asks for nothing in return.