Minutes会议记录助手

Any platform — from source (requires Rust + cmake; Windows also needs LLVM)

cargo install minutes-cli # macOS/Linux cargo install minutes-cli --no-default-features # Windows (see install notes below)

Or use Mistral API (requires MISTRAL_API_KEY)

[summarization] engine = "mistral" mistral_model = "mistral-large-latest"

Or from source (requires Rust + cmake)

export CXXFLAGS="-I$(xcrun --show-sdk-path)/usr/include/c++/v1" cargo install --path crates/cli ```

Requires (1) parakeet.cpp installed (https://github.com/Frikallo/parakeet.cpp)

Set your name (required for Levels 0-2)

Setup (one-time)

Step 1: Create a sync folder — pick one that syncs between your phone and desktop:

```bash

First-time install

claude plugin marketplace add silverstein/minutes claude plugin install minutes

Install

Download pre-built binary from GitHub releases, or build from source:

Requires: Rust, cmake, MSVC build tools, LLVM (for libclang)

Install LLVM (needed by whisper-rs bindgen):

winget install LLVM.LLVM [Environment]::SetEnvironmentVariable("LIBCLANG_PATH", "C:\Program Files\LLVM\bin", "User")

Full build (includes speaker diarization):

cargo install --path crates/cli

Apple Metal (macOS) — already enabled in the release DMG; use this for source builds

cargo install --path crates/cli --features metal

Setup (all platforms)

```bash

Install ffmpeg for best transcription quality (strongly recommended for non-English audio)

brew install ffmpeg # macOS

apt install ffmpeg # Linux

AND (2) a Minutes CLI compiled with `--features parakeet`. The downloadable

CLI (`brew install silverstein/tap/minutes`) and bare `cargo install minutes-cli`

do not. See docs/PARAKEET.md for the source-build walkthrough.

minutes setup --parakeet # Multilingual v3 (tdt-600m, ~1.2GB) minutes setup --parakeet --parakeet-model tdt-ctc-110m # English-only compact model (~220MB)

Also installs native Silero VAD weights for the parakeet.cpp --vad path

macOS — build from source

export CXXFLAGS="-I$(xcrun --show-sdk-path)/usr/include/c++/v1" export MACOSX_DEPLOYMENT_TARGET=11.0 cargo tauri build --bundles app --features parakeet,metal

Windows — build desktop installer from source

cargo install tauri-cli --version 2.10.1 --locked cd tauri/src-tauri cargo tauri build --ci --bundles nsis --no-sign ```

Tagged GitHub releases can include both a Windows NSIS installer as minutes-desktop-windows-x64-setup.exe and a raw desktop binary as minutes-desktop-windows-x64.exe. The installer is currently unsigned, so treat it as an advanced-user / preview distribution surface until Windows signing is added.

The desktop app adds a system tray icon, recording controls, audio visualizer, Recall, and a meeting list window. The current Windows desktop build covers recording, transcription, search, settings, and Recall. Calendar suggestions, call detection, tray copy/paste automation, and the native dictation hotkey remain macOS-only for now.

Release workflow details live in:

• docs/RELEASE-MACOS.md
• docs/RELEASE-WINDOWS.md

For macOS development, use a dedicated signed dev app identity:

• Production app: /Applications/Minutes.app (com.useminutes.desktop)
• Development app: ~/Applications/Minutes Dev.app (com.useminutes.desktop.dev)

If you are testing hotkeys, Screen Recording, Input Monitoring, or repeated macOS permission prompts, launch only Minutes Dev.app via ./scripts/install-dev-app.sh. Avoid the repo symlink ./Minutes.app, raw target/ binaries, or ad-hoc local bundles for TCC-sensitive testing.

This repository is open source, so local development does not require the maintainer's Apple signing credentials:

- ./scripts/install-dev-app.sh works with ad-hoc signing by default - for more stable macOS permission behavior across rebuilds, set MINUTES_DEV_SIGNING_IDENTITY to a consistent local codesigning identity - release signing and notarization remain maintainer/release workflows

For dictation, the recommended path is the standard shortcut in the desktop app (Cmd/Ctrl + Shift + D by default). The raw-key path for keys like Caps Lock is available as an advanced option but remains more fragile and permission-heavy.

Privacy: All Minutes windows are hidden from screen sharing by default — other participants on Zoom/Meet/Teams won't see the app. Toggle via the tray menu: "Hide from Screen Share ✓".

Then replace /Applications/Minutes.app with the new build from

parakeet_vocab = "tdt-600m.tokenizer.vocab" # Safer when multiple Parakeet models are installed

vad_model = "silero-v6.2.0" # Silero VAD model (auto-downloaded by setup). Empty = disable.

# Prevents whisper hallucination loops on non-English/noisy audio.

[summarization] engine = "none" # Default: Claude summarizes conversationally via MCP # "auto" = auto-detect an installed agent CLI for pipeline summaries # "agent" = uses your Claude Code, Codex, OpenCode, or Pi subscription (no API key) # "ollama" = local, free # "openai-compatible" = OpenRouter, Vercel/Cloudflare gateways, llama.cpp, LM Studio, etc. # "claude" / "openai" = direct API key (legacy) agent_command = "claude" # Which CLI to use when engine = "agent" (claude, codex, opencode, pi, etc.) ollama_url = "http://localhost:11434" ollama_model = "llama3.2" openai_compatible_base_url = "http://localhost:11434/v1" openai_compatible_model = "llama3.2" openai_compatible_api_key_env = "" # Blank means no Authorization header for local endpoints. Desktop cloud endpoints can still use a saved Keychain key without rewriting config.

[diarization] engine = "auto" # "auto" (default — uses pyannote-rs if models downloaded, otherwise skips), # "pyannote-rs" (always on — native Rust, no Python), # "pyannote" (legacy — requires pip install pyannote.audio), # "none" (explicitly disabled)

Building your own agent on Minutes

Minutes is designed as infrastructure for AI agents. Files are the durable substrate; MCP is the active interface; live transcript JSONL and the local event log are the real-time paths. The MCP server is the primary integration surface today:

• Read meetings: list_meetings, search_meetings, get_meeting return structured JSON
• Track people: get_person_profile builds cross-meeting profiles with topics, open commitments
• Monitor consistency: consistency_report flags conflicting decisions and stale commitments
• Record + process: start_recording, stop_recording, process_audio for pipeline control
• Live coaching: start_live_transcript, read_live_transcript for real-time mid-meeting access
• Local event stream: minutes events --follow --since-seq N tails newline-delimited events, including finalized live utterances, for agents that want a durable cursor
• Voice profiles: list_voices, confirm_speaker for speaker identification workflows
• Resources: Stable URIs (minutes://meetings/recent, minutes://actions/open) for agent context injection

Any agent framework that speaks MCP can use Minutes as its conversation memory layer — the agent handles the intelligence, Minutes handles the recall.

TypeScript SDK — for direct programmatic access without MCP:

npm install minutes-sdk

import { listMeetings, searchMeetings, parseFrontmatter } from "minutes-sdk";

const meetings = await listMeetings("~/meetings", 20);
const results = await searchMeetings("~/meetings", "pricing");

Built with: Rust, whisper.cpp (transcription), pyannote-rs (speaker diarization), Silero VAD (voice activity detection), symphonia (audio decoding), cpal (audio capture), Tauri v2 (desktop app), ureq (HTTP). Optional: ffmpeg (recommended for non-English audio decoding).

Quick start

```bash

Optional: sidecar metadata

If your phone workflow also saves a .json file alongside the audio (same name, .json extension), Minutes reads it for enriched metadata:

{"device": "iPhone", "source": "voice-memos", "captured_at": "2026-03-24T08:41:00-07:00"}

This adds device and captured_at to the meeting's frontmatter. Works with any automation tool (Apple Shortcuts, Tasker, etc.).

Supports .m4a, .mp3, .wav, .ogg, .webm. Format conversion is automatic — uses ffmpeg when available (recommended for non-English audio), falls back to symphonia.

If a desktop call capture leaves a raw file under ~/.minutes/native-captures/, process that audio file directly with minutes process <path> --type meeting. For compatibility, minutes import <audio-file> also routes to the same meeting-processing path; minutes import granola remains the Granola history importer.

Optional: automated summarization

```toml

Desktop users can paste cloud gateway keys in Settings; Minutes stores them

shared config. CLI users can set any env var and name it below. Local servers

Optional: knowledge base integration

Maintain a living knowledge base from your conversations — person profiles, decision history, and a chronological log that compounds over time. Inspired by Karpathy's LLM Wiki pattern.

[knowledge]
enabled = true
path = "~/wiki"        # or your Obsidian vault, PARA system, etc.
adapter = "wiki"       # "wiki" (flat markdown), "para" (atomic facts), "obsidian" (wiki + [[links]])
engine = "none"        # "none" = structured YAML only (safest), "agent" = LLM extraction
min_confidence = "strong"

After each meeting, structured facts (decisions, action items, commitments) flow into person profiles automatically. Every fact carries provenance back to its source meeting.

minutes ingest --dry-run --all   # Preview what would be extracted
minutes ingest --all              # Backfill existing meetings
minutes ingest ~/meetings/call.md # Process a single meeting

Three output formats: - Wiki — people/{slug}.md with facts grouped by category - PARA — areas/people/{slug}/items.json with atomic facts (id, status, supersededBy) - Obsidian — Wiki format with [[wikilinks]] for cross-references

Safety: default engine = "none" extracts only from parsed YAML frontmatter. No LLM call, zero hallucination risk. Confidence thresholds filter speculative facts. Corrupt data is backed up, never silently destroyed.

Restart your terminal after setting LIBCLANG_PATH

produce loops on non-English audio. ffmpeg is optional but recommended.

Enable speaker diarization (optional, ~34MB ONNX models)

minutes setup --diarization

In your config file (`$XDG_CONFIG_HOME/minutes/config.toml` when set,

otherwise `~/.config/minutes/config.toml`):

[identity] name = "Your Name"

Configuration

Optional — minutes works out of the box.

```toml

By default: ~/.config/minutes/config.toml

Or: $XDG_CONFIG_HOME/minutes/config.toml when XDG_CONFIG_HOME is set

[transcription] engine = "whisper" # "whisper" (default), "parakeet" (opt-in, lower WER), or "apple-speech" (experimental) model = "small" # whisper: tiny (75MB), base, small (466MB), medium, large-v3 (3.1GB)

engine = "apple-speech" # Experimental: standalone `minutes live` only. Configure via config file or CLI, not desktop settings.

macOS — CLI only

brew tap silverstein/tap && brew install minutes

CLI only (terminal recording, search, vault sync)

brew tap silverstein/tap brew install minutes

DMG and tagged CLI release binaries include the feature; the Homebrew Formula

macOS CLI (Homebrew)

brew upgrade silverstein/tap/minutes

Phone → desktop voice memo pipeline

No phone app needed. Record a thought on your phone, and it becomes searchable memory on your desktop. Claude even surfaces recent memos proactively — "you had a voice memo about pricing yesterday."

The watcher is folder-agnostic — it processes any audio file that lands in a watched folder. Pick the sync method that matches your setup:

Claude integration

minutes is a native extension for the Claude ecosystem. No API keys needed — Claude summarizes your meetings when you ask, using your existing Claude subscription.

You: "Summarize my last meeting"
Claude: [calls get_meeting] → reads transcript → summarizes in conversation

You: "What did Alex say about pricing?"
Claude: [calls search_meetings] → finds matches → synthesizes answer

You: "Any open action items for me?"
Claude: [calls list_meetings] → scans frontmatter → reports open items

Claude Code (Plugin)

Install the plugin from the marketplace: ```bash

Browser-based integrations such as Google Meet are opt-in on purpose.

Alternative: use Parakeet engine (opt-in, local GPU via parakeet.cpp)

Troubleshooting

No speech detected / blank audio: The most common cause is microphone permissions. Check System Settings → Privacy & Security → Microphone and ensure your terminal app (or Minutes.app) has access.

tmux users: tmux server runs as a separate process that doesn't inherit your terminal's mic permission. Either run minutes record from a direct terminal window (not inside tmux), or use the Minutes.app desktop bundle which gets its own mic permission.

Build fails with C++ errors on macOS 26+: whisper.cpp needs the SDK include path. Set CXXFLAGS as shown above before building.

Dictation hotkey still fails after you enabled it in System Settings: The native hotkey uses macOS Input Monitoring, which is separate from Screen Recording. The fastest way to test the exact installed desktop identity is:

./scripts/diagnose-desktop-hotkey.sh "$HOME/Applications/Minutes Dev.app"

Use ./scripts/install-dev-app.sh first so you are testing the stable development app identity rather than a raw target/ build. The helper intentionally launches the app through LaunchServices; direct shell execution of Contents/MacOS/minutes-app --diagnose-hotkey can misreport TCC status.