Questions, answered plainly.

What みて is, where it runs, and what it does and does not do.

What it does

Does みて translate whole sentences?: No, and that is on purpose. みて is a reading aid, not a machine translator. It gives you the reading, the dictionary form, English glosses, inflection notes, and furigana for each word, so you keep reading the Japanese yourself instead of replacing it with English.
How is this different from Yomitan or 10ten?: Those are excellent browser extensions, and for ordinary web pages they are the better fit: they read the page's own text directly.みて's niche is everything outside the browser: native games, visual novels, and apps a browser extension can never reach.
Will the overlay get in the way of the game?: The overlay is layered and click-through. It does not steal focus or interrupt game input: your clicks and keys pass straight through to the game. みて sits unnoticed until you want a reading, then gets out of the way.
What is it built on?: PP-OCRv5 detection and recognition models run through ONNX Runtime; Lindera segments the recognized text into words; and definitions come from JMdict, ordered by JPDB frequency so the most likely meaning leads.

Is みて really fully local?: Yes. Capture, OCR, segmentation, and dictionary lookup all run on your machine. There is no cloud OCR and no network round-trip: nothing about what you are reading is sent off the device. Because みて is open source under AGPL-3.0, you can read the code and confirm it yourself.
Is it free and open source?: Both. みて is licensed AGPL-3.0. The capture path, the OCR models, the segmentation, and the lookup are all open to read and audit.

Which platforms does みて support?: みて is Windows-first (Windows 10 and 11) and optimized for NVIDIA hardware. It is not built for macOS or Linux. The overlay is a native Win32 layer, so the experience is specific to Windows.
Do I need an NVIDIA GPU?: An NVIDIA GPU is recommended for the fastest path. みて uses TensorRT where it can, falls back to CUDA, and falls back again to CPU, so it runs without a dedicated GPU, just not as quickly as the reference numbers.
Which apps does it work with?: みて captures the pixels of a target window, so it does not need any cooperation from the app it reads. That covers native games, visual novels, manga and comic readers, and video (paused or live). If Windows can put it on screen, みて can read it.

How fast is it?: On the reference NVIDIA setup, a fresh full-screen read of a 4K game frame takes about 200ms (p95, on the native TensorRT path). Your hardware sets the exact number.
How accurate is the OCR?: Measured against several hundred hand-checked 4K screenshots, about 94 of every 100 text lines are read perfectly, and roughly one character in eighty is wrong or missing. Mistakes concentrate in tiny, faint, or heavily stylized text; ordinary dialogue and menus read very reliably.
What happens when it gets a character wrong?: When a misread happens, it usually shows up as an unknown word rather than a confident wrong definition, so a definitionみて shows is one it actually found.

How do I install みて?: Download the desktop app installer from the releases page and run it. The app installs the engine, downloads the recognition models, and walks you through NVIDIA's optional GPU setup, then you pick a window and read. It keeps itself and the engine up to date after that.
Do I need a terminal or to build it myself?: No. The desktop app is the normal way to install and runみて: there is no command line and no build step. Because it is open source under AGPL-3.0, you can build from source if you want to: clone the repo and run bootstrap-dev.ps1 in PowerShell. See the GitHub repository.