Download

Just want to use it? Download the installer — no Node.js, no Python, no commands. Want to hack on it? Skip to Development and run from source.

⬇ Get the latest release

Requirements: Windows 10 / 11 (x64), no administrator rights, plus your own API key(s) — MeetingCopilot is bring-your-own-key and collects no fee.

⚠️ This beta is not code-signed yet, so Windows SmartScreen will warn you. Confirm the file came from the official release page, then More info → Run anyway. Each .exe ships a matching .exe.sha256 with the hash CI computed — verify it with Get-FileHash .\MeetingCopilot-<version>-win-x64.exe -Algorithm SHA256.

🍎 macOS: no packaged build in this release. macOS is supported when run from source — see INSTALL_MACOS.en.md.

User guides: Quick start · API keys · Troubleshooting · Windows install & data locations

---

Real capture, no mockup: the interviewer's voice is transcribed while they are still speaking (left, live gray subtitle), and a read-aloud answer grounded in your resume streams in automatically (right).

🎬 3-minute real-world walkthrough

Click to watch (with sound): using an English podcast and a Chinese vlog as the "other side" — live transcription of both languages, inline translation, auto answers in Chinese and English, 0.94 s end-to-end latency on screen.

5-minute setup

The first launch opens a setup wizard — a normal, screen-shareable window, so someone can walk you through it remotely. Five steps:

1. Welcome — what bring-your-own-key means, what an API key is, and exactly what leaves your machine. Pick 中文 / English here; that becomes the app's language.
2. Plan — Low latency (Alibaba Cloud realtime ASR + DeepSeek answers, 2 keys), Minimal (MiMo for both, 1 key, Beta), Transcription only (1 key), or Advanced / local (change nothing, configure it yourself).
3. Services — the provider tutorial is inline: open the key page, paste, "Save and test connection". Keys are encrypted with the OS credential store before they touch disk.
4. Connection test — a live level meter proves the app can hear your computer, plus the optional microphone channel and a checklist of what is still missing.
5. Done — the whole plan is written as one settings update and the main window opens.

Step 2 "Plan": each card states plainly how many platforms you have to sign up for and how many keys it needs; "Low latency" is preselected. Local backends stay out of the way behind the "Advanced setup and local modes" link.

Step 3 "Services": one card per service, with the official key page, the official docs and a step-by-step guide right inside it — no hunting through a browser. Paste, press "Save and test connection", and the key is encrypted by the OS credential store before it touches disk.

Then: play some speech, hit ▶ Start, click ⚡Ans on any line. Full walkthrough in QUICK_START.en.md.

You can reopen the wizard any time from ⚙ Settings → Run the setup wizard again, and the in-app Help & guides (tray menu or Settings) has the same content offline.

Features

• 🎧 Hears the other side directly — no meeting bot. Windows captures system loopback audio. macOS uses a selectable audio input (choose a virtual device such as BlackHole for meeting/system audio). An independent microphone channel transcribes your own voice separately.
• ⚡ Four switchable ASR backend families — local sidecars (FunASR by default; experimental MOSS-Transcribe 0.9B), local Whisper turbo (offline fallback, DirectML GPU), Alibaba Cloud fun-asr-realtime, and MiMo per-segment. FunASR provides live partials; MOSS emits a finalized utterance after a pause.
• 🌍 Bilingual (zh / en) out of the box — the ASR detects Chinese↔English switches automatically mid-meeting, with no settings to touch; one click on the answer-language toggle (A:EN) and the teleprompter output flips to English too. Built for English interviews and code-switching conversations.
• 🌐 Fully English or Chinese interface — every label, tooltip, dialog and status message is available in both languages. Switch under Settings → Appearance → UI Language; first launch follows your OS language automatically. UI language and answer language are independent, so you can run an English UI while reading Chinese answers, or vice versa.
• 🧠 First-person teleprompter answers — bring your own key, any OpenAI-compatible LLM (DeepSeek recommended). Answers are written to be read aloud verbatim: conclusion first, then 2-3 short points; STAR for behavioral questions; idea → key points → complexity for technical ones. Never invents experience beyond your resume.
• 📄 Per-session resume + JD slots — import .md/.txt/.docx/.pdf; parsing is local and deterministic, nothing gets uploaded. Question-type detection (behavioral / technical / smalltalk) appends a zero-latency answering hint.
• 🔁 Rolling interview memo — a structured summary (questions asked / facts you claimed / interviewer focus) updates asynchronously after each answer, so a 60-minute interview stays self-consistent while per-request tokens stay flat.
• 🚀 Prefix-cache prewarm — pressing ▶ fires a 1-token request that pre-builds the LLM provider's KV prefix cache, so the first real answer prefills from cache (verified via DeepSeek prompt_cache_hit_tokens); kept warm automatically during capture.
• 🖼️ Region-screenshot Q&A — drag-select any screen region (the selection overlay itself is invisible to recording) and ask a vision model (MiMo / Gemini) about it.
• 🥷 Capture protection — content protection plus a global hide/show hotkey. Windows excludes the window from supported captures; macOS cannot guarantee invisibility against modern ScreenCaptureKit clients.
• 🩺 Connection tests, service status and local diagnostics — one click tells you whether a key, the network or the account is at fault (14 normalized error codes), and the diagnostics report is built locally with no keys, transcripts or resume text in it.
• 🔔 System tray — show/hide, start/stop transcription, new session, settings, service status, help and quit, all without a taskbar button. Optional start-at-login, off by default.
• 🌗 Dark / light / follow-system themes, 3-step answer font size, latency HUD, inline translation, multi-session with fully isolated transcript + chat + material per meeting.

One click between the English and Chinese UI

Settings → Appearance → UI Language: the whole interface — title bar, panels, tooltips, dialogs — flips instantly. The screenshots above show the English UI; the Chinese one is in README.zh-CN.md.

Bilingual in one session

A Chinese question, then an English one — same session, nothing reconfigured. The local ASR picks up the language switch automatically (both at ~1.6 s), and after one click on 答:EN the answer streams out in English, still grounded in the same resume.

ASR Backends

Which one should you pick? If you installed the packaged build, use a cloud backend: it needs one API key and nothing else. The local backends are an advanced option — free and fully private, but you have to bring your own Python environment and let the model weights download, and the installer ships neither.

Local ASR & advanced setup

Requirements per backend

Audio capture, Python setup, stealth behavior and hotkeys differ per platform — each OS has its own guide:

Local streaming FunASR

One-time Python environment, then the app auto-spawns and reaps the sidecar (tools/funasr_stream_server.py, ws://127.0.0.1:10097) — selecting the preset in Settings is all you do. The selected model downloads automatically from ModelScope on first run (~880 MB for paraformer, ~1.7 GB for Nano). --device auto picks CUDA / Apple MPS / CPU with automatic CPU fallback.

• Windows (conda env, NVIDIA GPU): see docs/windows/SETUP.md
• macOS (project .venv, Apple MPS): see docs/macos/SETUP.md

If your Python lives elsewhere, set MC_FUNASR_PYTHON to its full path.

MOSS-Transcribe-Diarize 0.9B (experimental)

MOSS is a one-shot generative model, not native streaming ASR. MeetingCopilot invokes the isolated tools/moss_asr_server.py sidecar after an utterance ends instead of repeatedly decoding a growing buffer. CUDA BF16 is attempted first and initialization failures fall back to CPU; the existing FunASR environment and default remain unchanged.

• Windows: see docs/windows/SETUP.md
• Custom interpreter: set MC_MOSS_PYTHON; force a device with MC_MOSS_DEVICE=cuda:0 or cpu.

Local Whisper turbo

Place onnx-community/whisper-large-v3-turbo-ONNX under <userData>/models/onnx-community/whisper-large-v3-turbo-ONNX/ — %APPDATA%/MeetingCopilot/ on Windows, ~/Library/Application Support/MeetingCopilot/ on macOS (encoder_model_fp16.onnx, decoder_model_merged_quantized.onnx, plus config/tokenizer files). The encoder uses DirectML on Windows and CPU elsewhere.

Cloud endpoints

• Aliyun DashScope: endpoint wss://dashscope.aliyuncs.com/api-ws/v1/inference, model fun-asr-realtime or paraformer-realtime-v2.
• MiMo: https://api.xiaomimimo.com/v1, model mimo-v2.5-asr.

Privacy

• API keys are encrypted at rest with Electron safeStorage (Windows DPAPI / macOS Keychain) and never reach the renderer process.
• All data (settings / sessions / materials) lives under Electron's per-user userData directory (%APPDATA%/MeetingCopilot/ on Windows and ~/Library/Application Support/MeetingCopilot/ on macOS). No telemetry, no accounts, no server.
• With the local ASR backends, audio never leaves your machine; with BYOK LLMs, transcripts go only to the provider you configured.
• The diagnostics report is built locally and contains no keys, transcripts or resume text — it is safe to paste into a public issue.

Development

git clone https://github.com/JWM0203/MeetingCopilot.git
cd MeetingCopilot
npm install        # postinstall applies patches/ (transformers.js patch — do not remove)
npm run build      # builds main + preload + renderer into out/
npm start          # cross-platform; Windows can also use start.bat

Running from source starts the same first-run wizard as the packaged build. Set MC_DEV_DEFAULT_LOCAL_ASR=1 to skip it and go straight to the overlay with the local FunASR defaults.

npm test            # unit tests (prompt building / VAD / stores / doc parsing / ASR protocol / tray / links)
npm run typecheck   # dual tsconfig (main + renderer)
npm run dev         # vite HMR dev mode
npm run verify      # typecheck + tests + build, the pre-commit gate
npm run dist:dir    # unpacked build into release/win-unpacked
npm run dist:win    # nsis + portable installers
npm run smoke:packaged        # boots the packaged exe and asserts both startup paths
node tools/rt-asr-smoke.mjs   # streaming-ASR protocol smoke (set MC_RT_URL / MC_RT_KEY)

Platform setup for developers (python envs, audio routing, stealth): docs/windows/SETUP.md · docs/macos/SETUP.md.

🇨🇳 If npm / Electron downloads are slow in China, create a .npmrc containing registry=https://registry.npmmirror.com and electron_mirror=https://npmmirror.com/mirrors/electron/.

Architecture in one line: Electron main process (window / stealth / tray / IPC / LLM routing / ASR host) → ASR engines inside a utilityProcess (never the main process — DirectML inference wedges there) → React renderer (transcript pane + answer session pane); all state lives in plain JSON files, never DOM storage.

Disclaimer

This tool is intended for personal learning and assistive use. Whether and how real-time assistance may be used in meetings or interviews depends on your local laws and the policies of the other party — you are solely responsible for how you use this software.

License

Apache License 2.0 — free to use, modify and redistribute, including commercially, under the terms of the license.