convez.ai
SkillProdutividadeIntermediário
daniellee-ux

Seminarly — Anotações de Reunião com IA Local

daniellee-ux/Seminarly-AI

Gravação de reuniões local-first com transcrição e IA por provedor próprio

Ver no GitHub
InicialPluginScore · de 0 a 100

App macOS que grava áudio de reuniões, transcreve localmente com WhisperKit e gera notas estruturadas usando seu próprio provedor de LLM (Claude, ChatGPT, Gemini, etc). Roda 100% no seu Mac sem enviar dados para nuvem, com custo de ~R$0,15 por reunião.

Instalação

$npx convez-ai daniellee-ux/Seminarly-AI

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 1 mês
Versão
v0.1.16
Issues abertas
14
Contribuidores
2

Casos de uso

  • Documentar reuniões de trabalho sem enviar áudio para serviços cloud
  • Transcrever calls com identificação de palestrantes (diarização neural)
  • Gerar notas estruturadas expandindo anotações manuais com contexto de transcrição

Veredito

Solução sólida para quem prioriza privacidade e custo em anotações de reunião. Implementação técnica robusta (WhisperKit, FluidAudio, integração multi-provider), mas restrita a macOS. Fluxo 'jot and enhance' com LLM é bem pensado.

README

A tradução para o português deste README ainda está na fila.

Seminarly

A local-first macOS app that records meeting audio, transcribes it on-device with WhisperKit, identifies speakers with FluidAudio neural diarization with Chinese-aware k-means re-clustering on raw 256D embeddings, and uses your choice of LLM (Claude, ChatGPT, Gemini, and 8 others) to generate structured notes — typically ~$0.30/month on Claude Sonnet.

Why Seminarly

Most AI note-taking apps send your audio to the cloud, charge $18+/month, and join your call as a bot. Seminarly runs entirely on your Mac — no bot, no upload, no subscription. You keep your data; you pay only the per-call cost of the LLM provider you choose (~$0.03 per meeting on Claude Sonnet, similar or lower on other providers).

The notepad is the key difference. Instead of passively summarising everything, Seminarly lets you jot shorthand during the meeting. Claude uses your notes as signals of what matters, expands them with transcript context, and fills in what you missed — the same "jot and enhance" workflow that makes Granola popular, at a fraction of the cost.

Getting Started

1. Install

git clone https://github.com/daniellee-ux/Seminarly-AI.git
cd Seminarly-AI
brew install xcodegen   # if not already installed
xcodegen generate
open Seminarly.xcodeproj

Build and run in Xcode (⌘R).

2. First launch

On first launch, Seminarly downloads the Whisper large-v3 model (~3GB from Hugging Face). This is a one-time download — transcription runs locally after that.

Grant permissions when prompted:

  • Microphone — to capture your voice
  • Audio capture — to tap other apps' audio (macOS 14.4+)

3. Pick your AI provider and add an API key

Open Settings (gear icon) → choose a provider in the AI Provider picker, then paste your API key for that provider. The key is stored in macOS Keychain (one entry per provider) and never leaves your device.

Default is Anthropic Claude (console.anthropic.com). Other supported providers are listed below.

4. Record your first session

Option A — Auto-detect (easiest) Start a Zoom/Meet/Teams call. Within a few seconds, Seminarly shows a banner: "Zoom is producing audio — Record?" Click Record and you're in.

Option B — Manual Click the record button (⏺) in the toolbar or use the menu bar icon. Pick the audio source from the dropdown and hit Start Recording.

5. Take notes while recording (optional but recommended)

Once recording starts, the app transforms into a notepad. Type anything — shorthand, headings, fragments:

# Action items
- infra cost review → daniel
budget approved for Q2
sarah presenting next week

You don't need to write full sentences. The AI fills in the gaps.

6. Stop and enhance

Hit Stop. Seminarly will:

  1. Finalize the transcript
  2. Identify speakers (FluidAudio neural diarization)
  3. If you wrote notes → enhance them with transcript context
  4. If you wrote nothing → structure the transcript automatically

The result appears in the meeting detail view within ~15 seconds.


Note Templates

Pick the template that fits your session before recording:

TemplateBest for
Freeform (default)Any conversation — model picks natural topic groupings, with nested bullets
MeetingStandups, 1:1s, team meetings — captures decisions and action items
LectureClasses, talks — Cornell-method notes with key concepts and review questions
Study GuideExam prep — Q&A pairs, memory aids, practice problems
PodcastInterviews, panels — key insights, notable quotes, references
CustomAnything else — write your own instructions

The Notepad Workflow

The notepad is Seminarly's core UX. A few tips:

  • Use # headings to define sections — # Action Items tells the AI to create that section
  • Write fragments, not sentences — budget Q2 approved is enough
  • Note names and owners — migration → alex by friday produces a properly formatted action item
  • You don't have to write anything — the auto-structure path works well on its own

After the session, your notes appear in the detail view. You can edit them and click Enhance to regenerate at any time.


Features

  • System audio capture — tap any app's audio (Zoom, Meet, Teams) via Core Audio Taps API
  • On-device transcription — WhisperKit (Whisper large-v3) runs locally, no cloud fees
  • Speaker diarization — FluidAudio neural embeddings with Chinese-aware k-means re-clustering
  • Live notepad — jot notes during recording; AI expands them with transcript context
  • AI-structured notes — summaries, action items, decisions with per-item source attribution
  • Multi-provider LLM support — Claude, ChatGPT, Gemini, Grok, Kimi, GLM, MiniMax, DeepSeek, Doubao (China + International) — bring your own key, switch any time
  • Summary language picker — generate notes in any of 24 preset languages (or any custom language) independent of the spoken transcript; covers Simplified vs Traditional Chinese, Cantonese, Bokmål vs Nynorsk, Brazilian vs European Portuguese
  • Auto-detect audio sources — detects meeting apps and prompts to record
  • 6 note templates — freeform, meeting, lecture, study guide, podcast, custom
  • Local storage — all data stays on your Mac (SwiftData)
  • Menu bar app — start/stop recording from anywhere
  • Export — Markdown file or clipboard copy

Supported AI Providers

Pick whichever provider you have an API key for. Each is configured in Settings → AI Provider with its own key (stored separately in Keychain) and editable model field.

ProviderDefault modelWhere to get a key
Anthropic Claude (default)claude-sonnet-4-6console.anthropic.com
OpenAI ChatGPTgpt-5.5platform.openai.com
Google Geminigemini-3.1-flash-lite-previewaistudio.google.com
xAI Grokgrok-4.20console.x.ai
Moonshot Kimi (international + China)kimi-k2.6platform.moonshot.ai / platform.moonshot.cn
Zhipu GLMglm-5.1bigmodel.cn
MiniMaxMiniMax-M2.7platform.minimaxi.com
DeepSeekdeepseek-v4-flashplatform.deepseek.com
ByteDance Doubao (China + International)Endpoint ID (China) / seed-1-8-251228 (Int'l)volcengine.com / bytepluses.com

Model field is editable so you can switch to any other model the provider supports.


Cost Comparison

ServiceMonthly (10 meetings)Annual
Seminarly (Claude Sonnet)~$0.30~$3.60
Granola$18$216
Notion AI$20$240
Otter.ai Pro$8.33$100

Each note-generation call typically costs ~$0.02–0.04 on Claude Sonnet; comparable or lower on other providers (some Chinese providers are noticeably cheaper). No subscription, no base fee.


Requirements

  • macOS 14.4+ (required for Core Audio Taps)
  • Apple Silicon recommended — Intel Macs can build and run it, but on-device transcription is slow
  • 8 GB RAM minimum; 16 GB recommended if you want the default Whisper large-v3 model. The smaller models work fine on lower-memory machines:
    • large-v3 — best accuracy, ~3 GB RAM during inference
    • medium — good accuracy, ~1.5 GB RAM
    • small — usable for clear English, ~500 MB RAM
  • Xcode 16.3+ (matches project.yml)
  • An API key from one of the supported providers (see table above)

Architecture

System Audio → Core Audio Taps → 16kHz PCM → WhisperKit → Transcript
                                                              ↓
Microphone  → AVAudioEngine  → 16kHz PCM ──────────────→ Diarization
                                                              ↓
                                               User Notes (optional)
                                                              ↓
                                              LLMProvider (selected)
                                  Anthropic / OpenAI-compatible / Gemini
                                                              ↓
                                                     Structured Notes
                                                              ↓
                                                     SwiftData (local)

See docs/ARCHITECTURE.md for a full breakdown of every module.


Coding Agent Integration

A small read-only CLI (seminarly-cli) plus a checked-in Agent Skill lets coding agents — Claude Code, Codex, Cursor, Gemini CLI, anything that can shell out — read your sessions in Markdown form.

In-repo (one-off):

./scripts/build-cli.sh                       # build once
./Tools/seminarly-cli list                   # JSON of sessions, newest first
./Tools/seminarly-cli get latest             # Markdown for your most recent session
./Tools/seminarly-cli search "OKR"           # find sessions mentioning a topic

Global install (use from any directory):

./scripts/install-global.sh                  # builds CLI + symlinks the skill into
                                             # ~/.agents/skills/seminarly-cli/
                                             # (also Claude Code's ~/.claude/skills/)
                                             # and the binary into ~/.local/bin/

After install, open any project in your coding agent and the seminarly-cli skill is discoverable; the bare seminarly-cli command resolves via $PATH.

Output distinguishes the three data types (user notes, AI-enhanced notes, transcript) with separate sections and inline [user] / [transcript] source tags.

See docs/agent-access.md for the full design, JSON shapes, and instructions for extending it.


Contributing

Bug reports, feature requests, and pull requests are welcome. See CONTRIBUTING.md for setup, tests, and conventions.

License

GNU Affero General Public License v3.0 (AGPL-3.0). See LICENSE. Third-party dependencies and their licenses are listed in THIRD_PARTY_LICENSES.md.

Ver o original no GitHub

Bom para

Relacionados