We opened 722 free AI tool demos. 270 of them were dead.

Not a link roundup — every tool was opened and checked. As of 2026-09-07. All figures are computed from our live catalogue, so this page is never out of date.

1. One in three "try it free" demos is already dead

We opened 722 hosted demos. 270 of them (37%) failed to load: 130 crashed at runtime, 102 had been paused by their authors, 35 could not start at all.

The surprise is the 102 that were paused deliberately. They did not break — the authors shut them down. Free compute quotas are finite, so popularity itself can end a demo. "Published" does not mean "still works." The older a roundup article is, the higher this rate climbs.

Alongside the 442 that loaded and the 270 that did not, 10 are held as undecided: 8 were asleep and may wake up when a person opens them, and 1 refused our automated request but loads fine in a browser. We do not fold undecided cases into whichever bucket flatters the number.

2. Most of them need no special skills

Put the other way round: only 26% run entirely on your own machine. Everything else ships your input to somebody's server. That matters the moment you feed it a recorded meeting, a contract, or a customer list.

3. You cannot judge commercial use from the licence name

We read 536 licence texts and found restrictions in 42 of them. A further 28stated no licence at all — which does not mean the work is free to use; it means the author retains every right by default. For hosted demos the answer depends on the model running inside, not on the page's own licence, so we read those terms too.

The trap: a licence can say non-commercialand still allow commercial use of what you make with it. FLUX is exactly that case — the restriction covers the model itself, while the current licence (revised 25 Nov 2025) states outputs may be used for any purpose including commercial ones, with one exception: you may not use them to train a competing model. The reverse also happens — licences without "non-commercial" in the name that still restrict the outputs.

4. Active, but not all of them

Median time since the last update is 80 days. 35% were updated within the last 30 days, while 7% have not moved in over two years. We have delisted 20 tools for being broken, non-compliant or abandoned, and we publish the reason for each.

5. It is not all famous projects

Median GitHub star count is 1,062; 559 tools (49%) sit below 1,000. A low star count usually means a narrow purpose, not poor quality — and narrow-purpose tools are exactly the ones nobody has written about yet.

What we checked, and how far

We do not check every one of the 1,142listed tools against every item. Here is the actual count per item, with the date it was last confirmed. We do not blend checked and unchecked into a single "verified" claim.

Method and limits

Scope is the 1,142 tools listed on Mirondo as of 2026-09-07. Hosted demos were requested and their state classified automatically. Licences were judged by reading the GitHub LICENSE text and the model distributor's terms. Setup requirements were judged by reading official documentation. Anything we could not determine is recorded as not verified rather than guessed.

The catalogue is weighted towards tools that are free to use, and does not cover paid services exhaustively. Read this as the shape of the free AI tool landscape, not of the market as a whole. States change daily, so cite with the date.

Using these numbers

Please do. CC BY 4.0 — attribution is the only requirement. Machine-readable source data lives at /data/tools.json and /data/verification.json.

Suggested citation: "Mirondo (https://mirondo.net), surveyed 2026-09-07". A Japanese version of this page is at /report.