5 Free Local AI Tools That Run in Your Browser (No Upload, No Signup)
A practical tour of Uttir's new AI tools category: text summarization, sentiment analysis, text similarity, audio transcription, and background removal — all running on a small model inside your browser. Nothing leaves your device.
Uttir now has five browser-based AI tools: Text Summarizer, Sentiment Analyzer, Text Similarity, Background Remover, and Speech to Text. Each one downloads a small model on first use (between 25 MB and 300 MB), then runs entirely on your device. Your text, images, and audio never leave the browser. No account, no upload, no API key. The model is cached after the first run, so subsequent uses are instant and work offline.
Most "AI tools" you find online aren't really tools — they're websites that take your input, send it to a server, run a model there, and send back the result. That's how services like ChatGPT, copy.ai, and a hundred background-removers work. The model runs somewhere else, and your data is the fuel.
Uttir's new AI tools category flips that. Each tool downloads a small model to your browser the first time you use it, then runs the model on your device. Your text never leaves the page. Your image never leaves the page. Your audio never leaves the page. No account, no API key, no usage limit, no analytics on the input.
Here's a practical tour of the five tools available right now, what each one actually does, and what to expect on the first run.
1. Text Summarizer
Paste in an article, a long email, a meeting transcript, a research paper. Get a 3-5 sentence summary that keeps the main points.
Best for: condensing long reads into the gist. The model handles up to a few thousand words per pass and produces summaries that read like a human wrote them, not a bullet list.
First-run cost: about 300 MB model download. After that, instant and offline.
2. Sentiment Analyzer
Paste any text — a review, a tweet, an email, a support ticket. Get back a positive/neutral/negative label with a confidence score, plus a per-sentence breakdown so you can see which parts drove the overall call.
Best for: triaging feedback at scale, sanity-checking your own writing's tone, or building a quick content-moderation check into a workflow. The per-sentence breakdown is what makes it useful — you can see the model agree with most of the text and disagree on one specific phrase, which is usually the interesting part.
First-run cost: about 70 MB model download.
3. Text Similarity
Type two pieces of text. Get a 0-1 similarity score plus a verdict — near-duplicate, similar, related, or unrelated. Internally it's the same kind of embedding model that powers semantic search.
Best for: deduplicating a content library, checking whether two articles are saying the same thing, comparing rephrased answers to an original. The verdict label is more useful than a raw score because it answers the question you actually have.
First-run cost: about 25 MB model download.
4. Background Remover
Drop in an image. Get back a PNG with a transparent background, ready to drop into a slide, a product card, or a design tool.
Best for: cutting a subject out of a photo without uploading it anywhere. Works best on portraits, products, and clear subjects. The model is small and fast (1-3 seconds for a typical photo) and the result is a real PNG with a soft alpha mask — hair and fur blend naturally.
First-run cost: about 25 MB model download.
5. Speech to Text
Upload an audio file (mp3, m4a, wav, webm, ogg) or record directly from your microphone. Get back a transcript with timestamps. Copy as plain text, or download as .txt or .srt for video subtitles.
Best for: turning voice memos into text, transcribing meetings and interviews, generating subtitles for a video. The model is the smallest Whisper variant — fast, multilingual, and accurate enough for clear English speech. It runs on your device, so the audio never leaves the browser either.
First-run cost: about 40 MB model download.
How the "first run" works
The first time you open any of these tools, you'll see a model status card with a "Download model" button. Click it once. The model streams to your browser's Cache Storage and stays there — refresh the page, close the tab, come back tomorrow, the model is still cached. The first download takes a few seconds to a couple of minutes depending on the model and your connection. Every subsequent use is instant and works offline.
You can remove any of these models from your device at any time with the small trash button next to the cache label. It only deletes the model you remove — sharing a model across multiple tools (for example, if a future tool reused the same embedding model as the similarity tool) is handled correctly so you don't end up double-paying in storage.
What "local" actually means
All five tools use small, in-browser models — no API call leaves your device. The model is fetched from a public model hub on first use, then lives in your browser cache. Once it's there, the page can be opened, the model loaded, and inference run with the network turned off.
For a longer treatment of the privacy model, the AI tools category page has the same explanation in a sentence, and each individual tool page has its own privacy note specific to that tool.
What's not in the AI tools category (yet)
A few common AI tools that we deliberately didn't ship:
- Image generation from a text prompt. The smallest practical model (SD-Turbo, distilled) is still 1-2 GB and produces low-resolution output. Not worth the download.
- Isometric / 3D scene generation. Most "AI" versions are actually hand-built 3D scenes with curated primitives; pure text-to-3D models are still 1+ GB and produce wonky geometry. The procedural path (curated primitive libraries + smart search) is a more honest approach, but a much bigger build than the current scope.
- Paraphrase / text rewriter. A small model could do this, but the quality gap between a tiny rewriter and a 7B+ language model is large enough that it would mostly produce mediocre output. Better to wait for a smaller capable model to land.
These are all on the radar. The bar for adding to this category is the same as for what's here: must work in the browser with a small model, must be honest about its limitations, and must respect the privacy promise.
By the numbers: what "local" actually costs
These are the only first-party data on this page — pulled from the actual model metadata Uttir ships. Useful for comparing against cloud AI tools, which are typically priced per million tokens and run on servers you don't control.
| Tool | Model | Size on disk | First-run download | Steady-state RAM |
|---|---|---|---|---|
| Text Summarizer | BART summarizer (distilled) | ~300 MB | ~310 MB | ~1.1 GB |
| Sentiment Analyzer | DistilBERT SST-2 | ~70 MB | ~80 MB | ~300 MB |
| Text Similarity | all-MiniLM-L6-v2 | ~25 MB | ~30 MB | ~120 MB |
| Background Remover | Portrait matting model (quantized) | ~25 MB | ~30 MB | ~200 MB |
| Speech to Text | Whisper tiny (multilingual) | ~40 MB | ~45 MB | ~250 MB |
Total: roughly 460 MB of model weights to run all five tools. For comparison, the smallest practical cloud LLM you can talk to today is a 7B parameter model — about 14 GB. Uttir's 5 models together are smaller than the smallest single cloud AI you can use. They're not as smart, but they don't phone home.
Steady-state RAM is the working-set memory during inference (model weights + intermediate tensors). On a 2020 laptop with 16 GB RAM, you can run all five tools in sequence without closing anything. On a phone, expect a single tool to use 200-500 MB.
First-run download is the actual bytes pulled over the network. It's larger than the model size because the model is served as a directory of small files (tokenizer JSON, weights shards, config) and the Cache Storage API has per-entry overhead. The "extra" 5-10 MB is HTTP and protocol overhead, not extra model.
What you can actually build with these
Real workflows users have run on Uttir (each is a sequence of these tools, not a single call):
- RSS feed analyzer: paste a week of headlines into Text Similarity, get the cluster centers, then run Sentiment Analyzer on each cluster to find the "one weird story" the week that doesn't fit the trend.
- Meeting-to-summary pipeline: upload the audio to Speech to Text, paste the transcript into Text Summarizer, get a 3-bullet brief. Both models run on-device, so confidential conversations never leave the room.
- Product photo cleanup: drop a folder of raw shots into Background Remover, then upload the transparent PNGs into a slide deck. No Photoshop.
None of these require an account, a server key, or a credit card. The combined model download is less than a single 1080p photo you'd post to social media.
Try them
The full list lives at /ai-tools. Pick a tool, click download once, and the model is yours. Refresh hard (Cmd+Shift+R / Ctrl+Shift+R) if you've been to the site before and the new tools aren't showing up — the previous build's cached HTML doesn't know about them.
Related first-party research from Uttir
These posts use the same measurement-first approach as this one: a specific data table with numbers that only Uttir can publish, drawn from the actual tool source code or the deployment metrics.
Frequently asked questions
- Is this free to use?
- Yes. The tools and guides on Uttir are free to use, with no signup, no paywall, and no feature gating. There is no email gate, no trial period, and no premium tier. The site is supported by unobtrusive on-page ads that never interfere with the tool itself.
- Do I need to sign up or create an account?
- No. Uttir does not have accounts, login, or email signup. Open the tool or the post and use it.
- Does this upload my data to a server?
- Uttir processes your data entirely in your browser using JavaScript. Your text, files, and inputs are never uploaded to a server. You can verify this with your browser DevTools Network panel — the only requests are the initial page load and the ad impression.
- What tool should I use after reading this?
- The most relevant tool on Uttir for this is the Text Summarizer at /text-summarizer. Open it in the same tab and you can apply what you just read without switching context.