How our AI tools work
How Client-Side AI Actually Works
"AI" is used loosely enough online that it's worth being precise about what ToolerWork's AI tools actually are: real, small machine learning models that download once to your browser and then run entirely on your own device — not a wrapper around a cloud chatbot. This guide explains the real mechanism, why the models are intentionally small, and where that honestly falls short of a large cloud AI service.
Reviewed 2026-09-13
Key takeaways
- ToolerWork's AI tools download a real, quantized machine learning model (tens to a few hundred megabytes) to your browser once, then run every request on-device from then on — nothing is sent to a server.
- The models are deliberately small (60 million to a few hundred million parameters) compared to a large cloud model like GPT-4 or Claude, which run on server hardware with tens or hundreds of billions of parameters.
- Small models are genuinely good at one narrow, well-defined task (summarizing, translating a language pair, checking grammar) but will not match a large general-purpose AI at open-ended conversation or reasoning.
- Because processing happens on-device, these tools work without an account, without a per-use cost, and without your file or text ever leaving your browser.
What "runs in your browser" actually means
Each ToolerWork AI tool uses a real neural network — the same general kind of technology behind large cloud AI models, just far smaller — converted to the ONNX format and run through ONNX Runtime Web, a version of Microsoft's ONNX Runtime compiled to WebAssembly so it can execute inside a browser tab without a plugin.
The first time you use a specific AI tool, your browser downloads that tool's model file (typically 60MB to around 250MB depending on the task) and caches it. Every use after that runs the model directly on your device's CPU (or GPU, where the browser supports it) — no network request happens for the actual processing step, only for that one-time download.
Why the models are small on purpose
Large cloud AI models are typically tens to hundreds of billions of parameters, need dedicated server GPUs, and cost real money per request to run — which is exactly why they live on a server rather than a phone or laptop. A model that size cannot realistically be downloaded and run in a browser tab.
ToolerWork's tools instead use models in the tens-to-low-hundreds-of-millions-of-parameters range, quantized (compressed from full precision down to 8-bit integers) to shrink the download further without destroying accuracy for their specific task. This trade-off is deliberate: it is what makes running real AI, for free, entirely on your own device, technically possible at all.
What ToolerWork's AI tools actually use
For transparency, here is what each on-device AI tool is actually built on:
| Tool | Model | Approx. size | Task |
|---|---|---|---|
| Grammar Checker | grammar-synthesis-small | ~95MB | Sentence-level grammar correction |
| Text Summarizer / PDF Summarizer | distilbart-cnn-6-6 | ~247MB | Real abstractive summarization |
| Hindi ↔ English Translator | opus-mt (en-hi / hi-en) | ~110MB per direction | Machine translation, one model per direction |
| Sentiment Analyzer | distilbert-sst2 | ~67MB | Binary positive/negative text classification |
| Handwriting to Text | trocr-small-handwritten | ~64MB | English handwriting recognition from a photo |
Where small on-device models fall short — honestly
A small task-specific model is not a substitute for a large general-purpose AI, and it is worth being direct about the difference instead of overselling it. These models are trained and tuned for one narrow job; asking them to do open-ended reasoning, hold a multi-turn conversation, or handle a task outside their training produces noticeably weaker results than a large cloud model would.
Concretely: the Grammar Checker catches common mistakes but won't replace a careful human edit on anything important. The Sentiment Analyzer is a binary classifier — positive or negative only, with no separate neutral category. The Handwriting to Text tool is trained on English handwriting specifically and does not read Hindi, Urdu, or other non-Latin scripts. Always treat the output as a fast first pass, not a final authority.
Client-side AI vs cloud AI, side by side
The practical trade-offs, laid out directly:
| ToolerWork's on-device AI | A large cloud AI service | |
|---|---|---|
| Where it runs | Your own browser, after a one-time download | A remote data center |
| Your data | Never leaves your device | Sent to a server for processing |
| Cost per use | Free, no account | Often free-tier limited or paid per use |
| Capability | Strong at one specific, narrow task | Broad, general-purpose reasoning and conversation |
| Speed after first load | Instant, no network round-trip | Depends on server load and your connection |
Related tools
Put the guide into practice
FAQ
Questions about this guide
Is ToolerWork's AI the same as ChatGPT running in the background?
No. Each AI tool downloads and runs its own small, specific model directly in your browser — there is no server-side chatbot or API call involved in processing your text, image, or PDF.
Why does the first use of an AI tool take longer than later uses?
The model itself (tens to a few hundred megabytes) downloads the first time you use that specific tool, then is cached by your browser. Every use after that skips the download and runs immediately.
Is my data actually private with these tools?
Yes. Since the model runs entirely on your device after downloading, there is no server request carrying your text, image, or PDF content during processing — nothing is uploaded.
Why isn't the output as good as a large AI chatbot?
The models are intentionally small so they can run in a browser at all, which trades away some of the broad reasoning ability a large cloud model has. They're built to do one specific task well, not to hold an open-ended conversation.
Do I need a powerful device to use these AI tools?
A reasonably modern laptop or phone handles them fine, since the models are small by design. Very old or low-memory devices may run the tools more slowly.