AI tools with stated limits

Directories of AI tools usually count features, which is a number nobody can check. This one counts routes, and the number is derived rather than asserted: every route on the site declares an engine family, and a single resolver decides which surface that family gets on that path.

Run across all 483 routes, the resolver yields 345 that call a language model, 83 that execute entirely in your browser, 38 that hand files to our media workflow, and 15 with no engine at all, this page among them. The 345 break down as 160 rewriters, 134 analyzers and 51 generators.

The resolver is worth understanding because it is the whole basis of the privacy and limit story. Family alone does not decide the surface. Route path participates: an analyzer aimed at watermarks or invisible characters resolves to a local scanner, while the same family on any other path resolves to a remote classifier instead.

That is why the analyzer count here, 134, exceeds the number of routes literally declared as classifiers. 64 analyzer routes are not watermark routes, so they are re-pointed to the remote classifier surface. It also means a claim like nothing leaves your device is true of some routes and false of others, and the honest version names which.

Remote, 345 routes

Rewriters, generators, classifiers, and analyzers on non-watermark paths. Metered and bounded.

Local, 83 routes

Deterministic cleaners and transformers, plus 16 scanner routes. No request, so no limit and no upload.

Media workflow, 38 routes

Image and video watermark work. Files, not text, and outside this directory.

The path through any of the 345 is the same, and each step has a failure mode worth knowing in advance.

Step 1

Pick by tool name. The server composes the model instruction from the route label, so route choice is the only prompt control you have.

Step 2

Submit 1 to 12,000 characters. The bound is enforced before the request reaches the model, so an oversized paste fails fast and costs you nothing.

Step 3

Read the output critically. It returns as finished prose with no hedging or preamble, so its confidence is a property of the format, not of its accuracy.

All 160 rewriting routes share one standing instruction: perform the transformation implied by the tool name, preserve the original meaning and factual claims, and return only the rewritten text without preamble, commentary, quotation marks or a markdown fence.

Two things follow. First, differentiation between rewriters comes from the route label alone, so the gap between a tone rewriter and a resume humanizer is real but narrower than the separate pages suggest. Second, preserving factual claims is an instruction, and instructions to a language model are followed with high but not perfect reliability. Treat fact preservation as a stated intent to verify, not a property of the output.

The 134 analysis routes carry a constraint that is unusual to publish. They are instructed never to claim certainty about authorship, plagiarism or which AI model wrote a text, and to describe observable signals and limitations instead. They also run at a markedly lower temperature than the rewriters, because variation is a defect in a description and a feature in a rewrite.

The consequence is that these tools underclaim relative to competitors advertising accuracy percentages. An analyzer here can tell you sentence length is unusually uniform. It cannot tell you a model produced the text, because that inference is not available from the text at all.

51 routes generate rather than transform. Their instruction differs in an important way: produce exactly the content the brief requests, match the format implied by the tool name, return only the finished content. There is no source text to constrain the output, so the brief is the only constraint there is.

With a five run allowance, this is where runs are most often wasted. A rewriter given poor input still produces something derived from it. A generator given three words produces three words of substance, and the retry is a second run.

Five runs per rolling 24 hour window, counted across all 345 routes together rather than per tool. The window rolls from the moment of use, so the allowance regenerates incrementally rather than at a fixed daily reset. The counter is keyed to a hashed identifier derived from the connection and holds a count and an expiry timestamp. Input text is not part of that record.

One to 12,000 characters per request, enforced at the boundary. Both limits apply only to the three remote families. The local routes never form a request, so neither limit exists for them, and quoting either limit as a site-wide fact would be wrong in both directions.

No, and the question deserves a direct answer rather than a deflection. Detection assigns a probability from surface patterns such as uniform sentence length. Because those patterns also occur in heavily edited human prose, a high score is evidence of a pattern and never proof of origin.

We hold no affiliation with Turnitin, GPTZero, Copyleaks or Originality.ai, and make no prediction about what any of them will return. A rewriter changes the text, which is all it claims. Any inference about what a third party scores afterwards would be a promise about software we do not run.