stormydayy/halo-overlay

Halo is for people who want an on-screen prompter during a live call: it captures the other party's audio, optionally your microphone, and screen frames, then asks an OpenAI-compatible model for a short line to say. The README is explicit that many interviews, exams, employers and platforms ban outside help or recording, and recording laws vary, so the rules where you use it matter more than the feature list.

The problem

In a live conversation you often need to recall a detail, phrase an answer, or catch up after losing the thread, while typing a prompt into a chat window would give you away. Halo sits as an overlay: audio and screen context are captured in the background, and you ask for a suggestion with a global shortcut. The new part is Rewind: after a lapse, one keystroke turns the last 30 seconds to five minutes of transcript plus selected keyframes into a three-part answer — what happened, what is being asked, and a ready-made reply.

What's different

The README positions Halo as a from-scratch rewrite of cheating-daddy with one OpenAI-compatible client instead of separate code paths per provider, so Gemini, Groq, OpenAI, OpenRouter, Ollama, LM Studio or your own server all plug into the same settings screen. Speech recognition is provider-independent too: audio is cut at pauses by a VAD and sent to any Whisper-compatible endpoint. Screen capture does not screenshot on a timer; a small grayscale copy is compared every interval and only frames that change and settle enter a rolling buffer, which should keep duplicate or mid-animation images out of model requests.

The trade-off is latency and heuristics. Because recognition runs in chunks, suggestions typically lag speech by one to three seconds. Auto-answer relies on a heuristic question detector whose false positives are why it is off by default. The app also expects Node.js on the desktop and ships no build step, but installers are produced separately with electron-builder; that model suits people comfortable running a repo locally.

Who should try it

It fits a developer or technical user who wants a private, configurable prompter for meetings, sales calls, negotiation or study, and who is willing to own model and speech keys. The fully local path — Ollama with a vision-capable model plus a local Whisper server — is the main reason to choose it over a hosted-only tool, provided your hardware can run both. It is a poor fit if you need streaming, low-latency transcription, or if the setting explicitly forbids assistive tools and recording; the README's disclaimer is not boilerplate here. macOS users should also check the audio limitation below before committing.

Requirements and limitations

The documented setup needs Node.js 20+ and Python 3.11+, then npm install and npm start; there is no bundled binary path in the README. You must configure a model and a speech-recognition provider, and local endpoints such as localhost or LAN addresses skip the API key requirement project instructions. System audio is the weak point across platforms: Windows uses loopback, macOS relies on a bundled SystemAudioDump helper built for Apple Silicon, and Linux only has system audio if the environment supports loopback; otherwise Halo continues with microphone and screen only. On macOS you must grant Screen Recording and Microphone permissions.

Ongoing upkeep is real: the README warns that default model names go stale, so the free-text Model field needs updating as providers change their catalogs. Session data and optional .webm recordings live locally and can be wiped from History or with a shortcut; API keys are encrypted via the OS keychain, but audio and screenshots still go to whichever providers you configure. If you want everything offline, expect to run and maintain two local services.

HotRepo's take

HotRepo only has a single dated observation: 325 total stars and #22 on the NEW-REPOSITORY board (新项目榜, not the overall daily board) on 2026-10-27, for a repository created two days earlier. That is not enough comparable daily history to judge whether interest persists. Before investing, verify the audio path for your exact OS and Mac architecture, since that is the most likely thing to break the intended workflow.