plaud-pipeline

plaud-pipeline

Claude Code plugin with an MCP server for local pipeline operations on already-synced Plaud voice recordings.

Plaud is a voice recording device (Pin for personal, Note for meetings). This plugin focuses on the parts that live on your disk after sync: status overview, classification by category + sensitivity, and health checks. For raw recording access (list recordings, get transcripts, download audio), use the official Plaud MCP — registered separately in Claude Code.

Quick start

Prerequisites

• Claude Code installed
• A Plaud account with recordings at web.plaud.ai
• Python 3.10+
• Node.js 20+ (for the official Plaud CLI/MCP)

1. Install this plugin

claude plugin marketplace add https://github.com/dkushnikov/plaud-pipeline
claude plugin install plaud-pipeline

2. Install Plaud's official CLI + MCP

npm install -g @plaud-ai/cli @plaud-ai/mcp
plaud login          # OAuth via browser, writes tokens to ~/.plaud/tokens.json

Add the official MCP to your Claude Code config (~/.claude.json):

{
  "mcpServers": {
    "plaud-official": {
      "command": "npx",
      "args": ["-y", "@plaud-ai/mcp@latest"]
    }
  }
}

3. Point PLAUD_DATA_DIR at your recordings

You need a local directory with your Plaud recordings as source.md files. Expected structure:

your-recordings/
  └── YYYY-MM-DD_recording-id/
      ├── source.md      ← YAML frontmatter + transcript
      └── audio.wav      ← optional

Add to your ~/.zshrc (or ~/.bashrc):

export PLAUD_DATA_DIR="$HOME/plaud-recordings"

Sync is handled outside this plugin — by the plaud CLI directly or by an external sync script (e.g., ~/Atlas/bin/plaud-sync.py) that shells out to plaud CLI to populate the data directory.

4. Test

Start a new Claude Code session and run /plaud-status. You should see your recording counts.

What it does

Tool	Description
plaud_status	Pipeline overview — counts by stage (synced/extracted/processed)
plaud_classify	Classify recordings by category + sensitivity using configurable rules
plaud_health	Integrity checks — duplicates, ghost folders

For listing recordings, fetching transcripts, downloading audio, or AI summaries — use Plaud's official MCP tools (list_files, get_transcript, get_note, get_file).

Optional: classification config

Create plaud.yaml in your recordings directory to customize classification:

# Map your device serial numbers to types
# Find serials: web.plaud.ai → Settings → Device Info, or check source.md frontmatter (device_serial field)
devices:
  "YOUR_PIN_SERIAL": pin
  "YOUR_NOTE_SERIAL": note

# Keywords for classification rules
classification:
  work_keywords: [meeting, standup, sprint, review, planning]
  personal_keywords: [diary, reflection, journal]
  default_category_by_device:
    pin: personal-diary
    note: work-meeting
    zoom: work-meeting

Without this file, built-in defaults apply (English keywords, no device mapping — all devices treated as "unknown").

Finding your device serial: look at any source.md file — the device_serial field in frontmatter is your serial number. Or check the Plaud app settings.

Classification

Six categories, assigned by device type + keyword matching + speaker count:

Category	Typical source
work-meeting	Note device, work keywords in title
work-1on1	"1-1" / "1on1" pattern in title
personal-diary	Solo Pin recording
personal-conversation	Multi-speaker Pin with personal keywords
voice-memo	Short recordings (< 3 min solo)
interview	"interview" in title

Sensitivity routing: work → shared, personal → personal only, private/unknown → hold for review.

Notes

• plaud_classify is read-only — returns proposed classifications but doesn't write back to frontmatter yet.
• Auth is delegated to the official @plaud-ai/cli and @plaud-ai/mcp (OAuth). This plugin no longer reads PLAUD_TOKEN.

Development

pip install -e ".[dev]"
pytest tests/ -v
python -m plaud_pipeline --transport stdio

License

MIT