Understand what your AI is saying.

Highlight anything on your screen. Press a hotkey. Gloss explains it in plain language, in context, in your language.

Overview

AI coding agents brought millions of people into software who never learned to code. Their agents speak fluent developer. They do not. When Codex asks "should we use Docker?", every unfamiliar term is secretly a decision, and the choice is to click yes blindly or break your flow to ask.

Gloss is a Windows desktop overlay that removes that fork. Select any text in any app (Codex, ChatGPT, VS Code, a terminal, your browser), press the hotkey, and a small card appears with a plain language answer. Not a dictionary entry. A decision card:

Field	Answers
What it is	The term, in one plain sentence
Why it's suggested here	Why your agent proposed it in your project, read from your screen
If you say yes	What actually changes
If you say no	The alternative, and what stays the same
Reversibility	Easy to undo, some effort, or hard to change later

Codex Pets tell you what your agent is doing. Gloss tells you what it is saying.

A gloss (noun): a short note in the margin that explains a difficult word. That is the whole product. It has been the whole product for about five hundred years. This one is for the AI age.

Features

• Decision cards, not definitions. Every card is built for the moment an agent asks you something.
• Screen context. With one toggle, Gloss sends a single screenshot with each lookup so the model can read your screen and answer about your project instead of in general. Cards are labeled context on or context off.
• Any language. One tap translates the selection. Bangla by default, ten presets, or type any language. Technical terms stay in English with a short gloss in your language, so you learn the real words.
• Go deeper. A second press re-asks the flagship model for a fuller answer with concrete examples.
• A glossary that keeps score. Every card is saved. "You learned 9 concepts this week" is the product's win condition, written as UI.
• A desktop pet. A small scholar lives on your desktop, thinks while Gloss thinks, and speaks your cards as bubbles. Drag it anywhere. Click it for your glossary. Or hide it and cards appear at your cursor instead.
• Works everywhere. OS level overlay, not a browser extension. Terminal safe by design.
• Nothing to sign up for. No accounts, no server, no telemetry. Your own OpenAI key, stored on your device.

Make it yours

• Hotkey: rebind it in Settings with a live recorder. Default is Ctrl+Alt+E.
• Language: pick from ten presets or type any language. The card's translate button relabels instantly.
• Screen context: a visible toggle, on by default, mirrored in the tray.
• Pet: show or hide the scholar. Visible, your cards speak from it. Hidden, cards appear at your cursor.

Install

Judges: you only need Option A.

Option A: Installer (two minutes)

1. Download Gloss-Setup-0.1.0.exe from Releases.
2. Windows SmartScreen will warn about an unsigned app (hackathon build, no code signing certificate). Click More info, then Run anyway.
3. On first run, paste any OpenAI API key (get one here). It is stored on this device and sent only to OpenAI. A lookup costs about half a cent on GPT-5.6 Luna.
4. Test it: highlight any sentence in your browser and press Ctrl+Alt+E.

Option B: Run from source

git clone https://github.com/BlinkZ404/Gloss.git
cd Gloss
npm install
npm start        # dev run
npm run dist     # build the NSIS installer into dist/

Both options end in the same app. Platform: Windows 10/11. Node 20+ for source builds. Requires an OpenAI API key with GPT-5.6 access.

How a lookup works

1. You press the hotkey. Gloss first captures a screenshot of your display (if context is on), before any window appears.
2. It saves your clipboard, writes a sentinel value, waits for your modifier keys to physically release, then injects Ctrl+Insert into the focused app. Never Ctrl+C: in a terminal, Ctrl+C is an interrupt and would kill your running process.
3. It polls the clipboard until the copy lands, reads your selection, then restores your original clipboard, images included.
4. If a surface does not support copying, Gloss falls back gracefully: copy manually, press the hotkey, and the card is labeled "From your clipboard" instead of "You selected". Cards never pretend.
5. GPT-5.6 returns a structured JSON card. Luna answers instantly; Sol answers "Go deeper".

Built with Codex and GPT-5.6

Built solo in three days. Codex wrote most of the code, running on GPT-5.6 Sol. I planned, tested every task by hand, and made the calls.

• How we worked: the first message was the full plan: architecture, the capture pipeline step by step, the card JSON, the rules. Codex turned that into AGENTS.md in task one and followed it for the rest of the build. After that, one task per message. Sometimes a single line, sometimes a multi-part spec. Bug reports were pasted logs with timestamps, not descriptions.
• GPT-5.6 in the product: Luna answers the instant cards because it is fast and cheap. Sol answers "Go deeper". Screen context sends one screenshot per lookup so the model sees your project before explaining why your agent suggested something.
• Calls that were mine: decision cards, not a dictionary. Inject Ctrl+Insert, never Ctrl+C, because Ctrl+C kills a running terminal process. Clipboard sentinel, then restore what was there. The card never steals keyboard focus. Honest labels for where the text came from. No accounts, no server, your own key. The prompts, the margin-note design, the Education framing.

Favorite bug: capture failed for a day until one Codex diagnostic showed I hold my hotkey for almost 0.8 seconds, so the copy chord was arriving while Ctrl and Alt were still pressed. The fix was waiting for my fingers to leave the keys. Gloss takes a tenth of a second. I was the latency.

Mascot and brand art: ChatGPT image generation, iterated by hand until the character stopped looking like a baby and started looking like a scholar.

How GPT-5.6 powers Gloss

Job	Model	Why
Instant explain and translate cards	gpt-5.6-luna	Fast and inexpensive; a card in one to three seconds
Go deeper	gpt-5.6-sol	Flagship reasoning for fuller answers with examples
Screen context	gpt-5.6-luna (image input)	Reads a screenshot of your display so "why it's suggested here" is about your project

Tiering the models this way, fast for instant answers and flagship for depth, was a product design decision.

Privacy

• Your API key is stored locally with electron-store and sent only to api.openai.com.
• Screen context is a visible toggle. When on, one screenshot per lookup goes to OpenAI with the request. Gloss never stores screenshots, and nothing else leaves your machine.
• No accounts, no server, no analytics, no telemetry.

Known limitations

• Windows only, for now.
• Some application chrome text is painted, not selectable. For those, copy manually and press the hotkey.
• File lists on the clipboard cannot be restored after a capture (text and images are restored).
• Apps running as Administrator need Gloss run as Administrator for injection to reach them.
• The installer is unsigned, so SmartScreen will warn. This is expected for a hackathon build.

Project structure

main.js               # App entry: windows, capture flow, IPC
preload.js            # Renderer bridge
lib/                  # Main-process modules: capture, AI client, app plumbing
renderer/
├── card/             # The decision card overlay
├── glossary/         # Master detail learning record
├── pet/              # Desktop pet window
└── setup/            # First run and settings
assets/               # Icon, mascot, and brand art
AGENTS.md             # The build's rules, written by Codex in task one

Roadmap

• Sign in with ChatGPT. The disabled button on the setup screen is the promise. The day OpenAI's third party sign-in reaches apps like Gloss, the key field retires.
• Beyond code. The pipeline already reads anything on screen: contracts, forms, medical notes, news. Code is the beachhead, not the border.
• macOS support, a browser extension for in page popups, and an auto popup on selection with no hotkey to remember.
• On device OS models for an offline mode, and a managed backend for a key free consumer build.
• More pet. There is always more pet.

License & Credits

MIT. © 2026 Arifur Rahman Akash.

Built during OpenAI Build Week 2026 for the Education track. Mascot and brand art generated with ChatGPT image generation.