Hear the interviewer's question and draft your answer — live, and entirely on your machine. Built for interview practice, mock interviews, and self-review.
Real-time interview assistant — live dual-channel transcription (Deepgram) + LLM answers (DeepSeek / Gemini / OpenAI / Ollama), grounded in your own documents.
A single-window Electron desktop app. It listens to a conversation, separates you (microphone) from the interviewer (system audio), and on a hotkey detects the interviewer's current question and drafts a concise, first-person answer — using the last ~15 turns of dialogue plus any documents you upload as context.
Static screenshot
中文说明见下方 中文。
This tool is built for interview preparation, practice, self-review, mock interviews, and accessibility assistance.
Using real-time answer generation during a live interview without the other party's knowledge may violate the policies of the company you're interviewing with, employment agreements, academic-integrity rules, or local law, and many people consider it dishonest. You are solely responsible for how you use this software and for complying with all applicable rules and laws. The authors provide it "as is" with no warranty (see LICENSE) and do not endorse deceptive use.
This app sends data to third-party APIs only for the providers you configure:
- Audio (microphone + captured system audio) is streamed to Deepgram for transcription.
- Transcripts, your question, and uploaded document text are sent to your chosen answer provider (DeepSeek / Gemini / OpenAI) to generate answers.
- API keys are stored locally only, in
app.getPath('userData')/settings.json(e.g.~/Library/Application Support/interview-copilot/settings.jsonon macOS). They are never committed to the repo or sent anywhere except the provider's own API.
Want zero cloud? Use Ollama as the answer provider (runs locally). A fully local STT option (Whisper) is on the roadmap.
- 🎙️ Dual-channel transcription — mic = candidate, system loopback = interviewer, transcribed separately and color-coded.
- ⌨️ One-key answering — press
Ctrl+A(configurable): detects the interviewer's current question from the latest turns and streams an answer. - ⚡ Auto-answer (optional) — flip the toggle and it answers on its own as soon as the interviewer finishes a question, no keypress needed.
- 🧠 Context-aware — answers are grounded in the last ~15 turns of dialogue + your uploaded résumé / JD / notes (a persistent Knowledge Base you can add to, update, and clear).
- 🎯 JD customization — paste or upload a job description in Settings; it's persisted and tailors every answer to the target role.
- 🔌 Switchable providers — DeepSeek, Gemini, OpenAI, or local Ollama, with automatic retry + model fallback.
- ✍️ Concise, outline-style answers (default ≤ 500 chars), shown alongside the detected question so you can edit and regenerate.
- 🖥️ Single, clean UI — live transcript on the left; question / answer / knowledge base on the right.
- macOS 13+, Windows, or Linux — runs on all three; system-audio capture is most seamless on macOS (see Platform support).
- Node.js ≥ 18 (Node 20 LTS recommended; see
.nvmrc). - API keys: Deepgram (STT, required) and at least one answer provider — DeepSeek / Gemini / OpenAI, or a local Ollama install.
git clone https://github.com/ericwang915/interview-copilot
cd interview-copilot
make run # installs deps if needed, then launches the appNo make? Use npm install && npm start. Run make help to see all targets (dev, test, lint, package, …).
On first launch, open Settings (gear icon) and enter your Deepgram key + one answer provider's key. Then:
- Upload a résumé / job description / notes (optional but recommended) into the Knowledge Base.
- Click Start Listening and grant microphone (and, for interviewer audio, Screen Recording) permission.
- When the interviewer asks something, press
Ctrl+A(or click Generate Answer).
# Build for THIS Mac's architecture (x64 or arm64), sign, and open — recommended
make app
# One .app that runs natively on both Intel and Apple Silicon
make universal
# Installers (electron-builder): dmg / nsis / AppImage
npm run dist # current platform
npm run dist:mac # mac — builds both x64 and arm64 dmgsPrebuilt Releases ship both an Apple-Silicon (-arm64.dmg) and an Intel (.dmg) build — download the one matching your Mac.
The built app lives in dist/. For everyday use, launch the packaged app (double-click), not npm start from a terminal — on macOS the Screen-Recording permission attaches to the launching process, so a terminal-launched dev build can't reliably capture system audio.
Prebuilt installers from the Releases page are not notarized (this is a free, open-source project without a paid Apple Developer certificate). So the first time you open it, Gatekeeper will warn that the developer can't be verified. To open it anyway:
- Right-click the app → Open, then click Open in the dialog (only needed once), or
- run
xattr -dr com.apple.quarantine "/Applications/Real Time Interview Copilot.app"in Terminal.
After that it launches normally. (Building from source — npm run package — produces a locally-signed app with no such prompt.)
| Platform | Run it | Mic (candidate) | System audio (interviewer) |
|---|---|---|---|
| macOS 13+ | make run (builds & opens the app) or a .dmg |
✅ | ✅ Screen-Capture loopback — grant Screen Recording |
| Windows | …Setup.exe, or npm install && npm start / make run |
✅ | ✅ getDisplayMedia loopback (tick "share system audio") |
| Linux | .AppImage, or npm install && npm start / make run |
✅ | getDisplayMedia loopback — pick a PulseAudio/PipeWire Monitor source under Interviewer audio |
make run adapts to the OS: on macOS it builds & opens the signed app (so the Screen-Recording permission attaches to the app); on Windows/Linux it launches via npm start. npm scripts are cross-platform (via cross-env).
No-permission / fallback (any OS): install a virtual audio device (BlackHole on macOS, VB-Cable on Windows) or use a system Monitor source, route the meeting app's output to it, and pick it under Interviewer audio.
src/
main/ Electron main process (Node)
main.js window, global hotkey, IPC, generation orchestration
preload.js contextBridge — safe renderer API
settings.js persisted settings (userData/settings.json)
config.js provider registry (models, base URLs, fallbacks)
prompt.js prompt builders (answer + question extraction) — pure, tested
llm.js provider-agnostic retry + fallback
gemini.js Gemini REST (SSE)
openaiCompat.js OpenAI-compatible client (DeepSeek / OpenAI / Ollama)
store.js knowledge base (context stuffing)
documents.js txt/md/pdf/docx parsing + chunking
renderer/ UI (no Node access)
index.html / styles.css / app.js
deepgram.js Deepgram live WebSocket client
pcm-worklet.js Float32 → 16-bit PCM AudioWorklet
test/ node:test unit tests
Key design notes:
- Deepgram runs in the renderer via a browser WebSocket using subprotocol token auth (
['token', key]) — no SDK/bundler. See SECURITY.md for the trade-off. - Answer providers run in the main process via REST/SSE — no third-party AI SDKs.
- Question detection uses only the latest turns; the answer is grounded in the broader recent history.
- Local STT (Whisper) for a fully offline pipeline
- Windows/Linux system-audio capture
- Optional vector RAG for large knowledge bases
- Auto-trigger on detected question end
- Answer history & export, i18n UI
PRs welcome — see CONTRIBUTING.md. Security issues: see SECURITY.md. Sharing it? There's ready-to-post copy in LAUNCH.md.
If Real Time Interview Copilot is useful to you, please ⭐ star the repo — it genuinely helps others find it.
基于 Deepgram（实时转写）+ DeepSeek / Gemini / OpenAI / Ollama（答案生成）的桌面应用：双声道识别（麦克风=面试者、系统声音=面试官），按 Ctrl+A 自动识别面试官当前的问题，并结合最近约 15 轮对话 + 你上传的资料，生成简洁的第一人称大纲式答案（默认 ≤500 字）。
🔐 隐私：音频会发送到 Deepgram 转写；转写文本、问题与资料会发送到你选择的答案 Provider；API Key 仅保存在本机 userData/settings.json，不入库、不上传。想完全本地：把 Provider 选成 Ollama。
运行：npm install && npm start，首次在「设置」里填 Deepgram Key + 任一答案 Provider Key。打包：npm run dist。日常使用请用打包后的 App（双击），不要从终端 npm start（否则 macOS 屏幕录制权限会算到终端头上）。