Meetphant Meeting Assistant
An elephant never forgets. Turn meeting recordings into notes you can take away.
Record or import audio → transcribe → key points, action items & decisions.
Download for Windows · Website · 中文说明
Why Meetphant
Meetphant is a free, open-source, local-first desktop meeting assistant for Windows. It sits next to whatever meeting app you already use — no bots joining your call, no plugins, no host permissions.
| |
|---|
| 🎙️ Record any meeting | Microphone + system audio at once — Tencent Meeting, Feishu, Zoom, Teams or an in-person room |
| 🧠 Hotwords that learn your vocabulary | People, customers and product names go into every transcription request |
| 🔒 Your notes, your call | Summarize with any OpenAI-compatible model — cloud, or a local LLM such as Ollama |
| 📋 Notes you can act on | Key points · action items · decisions, in Chinese, English or both |
| 💾 Local-first | Meetings and settings in SQLite; API keys only in the OS keychain |
Features
01 · Record any meeting, no platform lock-in
- Capture your microphone and system audio (WASAPI loopback) in one click
- Live dual-track levels, so you know both sides are being recorded
- Already have a recording? Import
wav, mp3, m4a, flac, ogg, aac
- Saves M4A when FFmpeg is available, falls back to WAV without blocking you
02 · Custom hotwords for your team’s vocabulary
- Colleagues, customers, project code names — add once, use every time
- Hotwords only affect transcription; summaries get their own context setting
- Speakers are separated and can be renamed so notes show real names
03 · Connect a local LLM — your notes, your call
- Presets for Alibaba Cloud DashScope (Qwen, default), DeepSeek, OpenAI, Moonshot (Kimi), Zhipu and Volcengine Ark (Doubao)
- Or any custom endpoint, e.g.
http://localhost:11434/v1 for Ollama or a model on your company network
- With a local model, summarizing never sends meeting content to a third-party AI service
- API keys live in the OS keychain — never in the database, never shown in plain text
[!NOTE]
Speech-to-text currently uses the cloud-based Doubao Seed-ASR 2.0 file transcription service. Audio is uploaded through your own Volcengine TOS bucket.
And more
- Structured summaries — what was discussed, who does what, what was decided
- Workspace — meeting sidebar with a split transcript / summary view (tabs on narrow windows)
- Light & dark themes — follows your system or set it yourself
- Auto-update — checks for new versions at launch; download, install and restart in one click
How it works
| Step | | |
|---|
| 1 | Record or import | Capture mic + system audio during the meeting, or import an existing recording |
| 2 | Transcribe | Stopping a recording creates the meeting and starts transcription, with speakers separated and hotwords applied |
| 3 | Summarize | Get key points, action items and decisions — copy and share |
Download
Grab the latest installer from Releases (Windows 10 / 11, x64). The app interface is in Simplified Chinese.
| Installer | Contents | When to use |
|---|
Meetphant_<ver>_x64-setup.exe — lean (recommended) | App only | Normal installs. FFmpeg (~80–100 MiB) downloads on first need. Receives in-app updates |
Meetphant_<ver>_x64-offline-setup.exe | App + bundled FFmpeg | Slow or offline networks (manual updates) |
Both share one app id — installing one replaces the other.
Setup
Open Settings and fill in three credentials:
| Role | Provider | What you enter |
|---|
| Transcription | Doubao Speech | API Key (new speech console) |
| Audio upload for ASR (required) | Volcengine TOS | AK / SK + region / bucket |
| Summary | DashScope (default) or any OpenAI-compatible service | Provider, Base URL, API Key, model |
Secrets are stored only in the OS credential store — never in SQLite, never returned by settings_get.
📖 Step-by-step guide with screenshots (Chinese): Credentials guide
| File size | Path |
|---|
| ≤ 512 MiB (≤ 5 h) | Upload to TOS → pre-signed GET → Doubao Seed-ASR 2.0 submit / query (volc.seedasr.auc, up to 45 min polling) |
| > 512 MiB | Rejected with ASR_PAYLOAD_TOO_LARGE |
Hotwords are sent to ASR. context_text is used for summaries only and is not sent to Doubao.
Development
Prerequisites
npm install
npm run tauri dev # Vite on http://localhost:1420 + the Meetphant window
npm test # frontend tests (Vitest)
npm run typecheck
cd src-tauri && cargo test
src/ React UI (app / pages / features / ipc / shared)
src-tauri/ Tauri commands, services, SQLite, providers
website/ Product site (static HTML/CSS, zh at /, en at /en/)
scripts/ Installer packaging, FFmpeg prep, screenshot capture
npm run ffmpeg:prepare # cache the pinned FFmpeg essentials build
npm run pack:lean # lean installer
npm run pack:offline # offline installer
npm run pack:all # both → dist-installers/ (+ latest.json when signed)
Push a version tag (or run Actions → Release) to build both .exe files and latest.json and attach them to a GitHub release.
Auto-update
- The app checks
https://github.com/leoshcn/meetphant/releases/latest/download/latest.json on launch and from Settings → About.
- Only the lean installer is published on the updater channel.
- Release builds must be signed with a Tauri updater key:
- Local:
TAURI_SIGNING_PRIVATE_KEY or TAURI_SIGNING_PRIVATE_KEY_PATH (optional TAURI_SIGNING_PRIVATE_KEY_PASSWORD)
- CI: repository secrets
TAURI_SIGNING_PRIVATE_KEY and optional TAURI_SIGNING_PRIVATE_KEY_PASSWORD
- Generate keys once with
npm run tauri signer generate -- -w ~/.tauri/meetphant.key; keep the private key offline. The public key is embedded in src-tauri/tauri.conf.json.
- Windows Authenticode / SmartScreen signing is not configured yet (separate from Tauri update signatures).
Bundled FFmpeg is the Gyan essentials build (GPLv3) — see GyanD/codexffmpeg.
The product site lives in website/ — static HTML/CSS with no build step (Chinese at /, English at /en/).
npx serve website # or: py -3 -m http.server -d website 8000
npm run screenshots # re-render website/assets/screenshots/*.png and og.png
npm run screenshots renders the real React UI in headless Chromium with mocked Tauri IPC and sample data (scripts/screenshots/). Re-run it after UI changes and commit the images — this README uses them too.
Pushes to main that touch website/** deploy via .github/workflows/pages.yml. One-time setup: Settings → Pages → Source = GitHub Actions.
Tech stack
Tauri 2 · React 19 · TypeScript · Vite · Rust · SQLite · Doubao Seed-ASR 2.0 · OpenAI-compatible LLMs · OS keyring
License
MIT © 2026 Leo Li — free to use, modify and distribute.
