This is what it does
Try it
Press the hotkey, drag a box over part of the screen, and the text inside it is read on your own machine and handed to a model that explains what it is looking at — instead of the five-tab ritual of snip, open a chat, paste, type the question.
A flow walkthrough — the capture, the OCR and the model cannot run in a browser page
- What it does: full-screen or region capture, local Tesseract OCR with control characters stripped, the model provider you configured, the answer copied straight to the clipboard, and a local history file so the last exchange is still there.
- What to try: drag a box over the stand-in page and read the manifest the pipeline carries forward — the region, the destination folder, and the coordinates’ coordinate space.
- What it runs on: the selection step is real, and so is the manifest. The other three stages are not: capture needs a Windows screen and the compiled .NET helper, OCR needs Tesseract on that machine, and the explanation needs your key.
- What it does not do yet: there is no installer and nothing hosted — the folder is the distribution. This page will not invent the text an OCR engine would have read.
Status
Where this actually stands today
- It is a working local tool. A GUI and a text-only command line run over the same capture, OCR and model pipeline, launched from seven one-click batch files for the two modes, the build step, the OCR install and the shortcut.
- It never used to check its API key, and that was a real defect. The config was created on first run by copying the example file, so every request carried the literal placeholder key; the call failed and the only feedback was three retry warnings. It never said the one thing that was actually wrong.
- That is fixed, and the fix is tested negatively. A config guard now rejects the shipped placeholder, an empty key, a whitespace-only key, a config with no key field and a null config before any request is sent. Its test reports 8 passed, 0 failed, and the last two cases confirm a realistic key is still accepted.
- The guard checks shape, not validity. It can prove a key is present and is not the placeholder; it cannot tell you the key works, and it does not claim to.
- Requirements are not optional. Windows 10 or 11, PowerShell 5.1 or newer, the Tesseract OCR engine installed separately, and a model API key.
The numbers
Constants from the build
8config-guard tests passed, 0 failed, including two that keep a good key accepted
7one-click batch launchers for the two modes, the build, the OCR install and the shortcut
2modes — a Windows Forms GUI and a text-only command line — over one pipeline
0installers and accounts; the folder is the distribution
Under the hood
What it is built on
- Stack
- PowerShell scripts behind a Windows Forms interface, a small compiled .NET C# capture helper, and Tesseract OCR running locally.
- Surfaces
- A GUI and a text-only command line over the same capture, OCR and model path, launched from one-click batch files.
- Runs where
- From the folder on Windows 10 or 11 with PowerShell 5.1 or newer. There is no installer; an install script handles the OCR engine through the system package manager.
- Data
- A local config file holding your provider key, and a local history file the answered exchanges are appended to.
- Payments
- None for the tool, which is MIT licensed. You supply your own model key and the provider bills you for what you use; the tool takes no share of it.
- Auth
- No account and no sign-up. The only credential is your provider key, and the guard can prove it is present and not the shipped placeholder — it cannot prove the key works.
- Tested by
- A PowerShell test script over the config guard: 8 cases, 0 failures, covering the placeholder key, an empty key, a whitespace-only key, a config with no key field, a null config, and two realistic keys that must still be accepted.
Plainly
What this is not
- It is not a cloud screenshot service. It is a folder on a Windows machine, and it runs only while you run it.
- It is not private in the way “local OCR” can imply. The reading is local; the recognised text is then sent to the model provider you configured. That distinction is worth understanding before you point it at a screen.
- It is not cross-platform. It is Windows-only, and parts of it are Windows-specific by design rather than by accident.
- It does not understand a screenshot on its own. The OCR engine does the reading and the model does the explaining; the tool is the pipeline that joins them to one hotkey.
- It is MIT-licensed, and it is not sold: your provider’s usage charges are the only bill attached to it.
Questions about Screenshot Assistant
What the OCR handles well, what happens to a screenshot after it is taken, or what a launcher does on first run — ask before you plan around any of it.