minutes-openclaw

Local whisper.cpp audio transcription for OpenClaw, powered by the Minutes CLI.

Registers Minutes as a MediaUnderstandingProvider so any OpenClaw channel (WhatsApp, Telegram, etc.) gets private offline transcription without sending audio to a cloud API.

Requirements

• minutes binary in PATH (or set MINUTES_BIN=/path/to/minutes) that includes the minutes transcribe --json command (added upstream in silverstein/minutes#380, merged to main after v0.19.0). If minutes --help doesn't list a transcribe subcommand, build from source or wait for the next tagged release.
  • Install: cargo install minutes-cli or download from github.com/silverstein/minutes/releases
• Whisper model downloaded: minutes setup --model small
• ffmpeg (recommended for ogg/opus/webm from WhatsApp/Telegram):
  • macOS: brew install ffmpeg
  • Linux: apt install ffmpeg
  • Without ffmpeg, WAV and MP3 are still supported natively

Installation

One-liner — installs and builds directly from GitHub, then configures OpenClaw:

openclaw plugins install github:maosuarez/minutes-openclaw
curl -fsSL https://raw.githubusercontent.com/maosuarez/minutes-openclaw/master/install.sh | bash

Custom binary or language:

MINUTES_BIN=/usr/local/bin/minutes MINUTES_LANGUAGE=es \
  curl -fsSL https://raw.githubusercontent.com/maosuarez/minutes-openclaw/master/install.sh | bash

From a local clone (for development or offline):

git clone https://github.com/maosuarez/minutes-openclaw
cd minutes-openclaw
openclaw plugins install .   # prepare script builds automatically
./install.sh

install.sh env vars:

| Var | Default | Description | |---|---|---| | MINUTES_BIN | minutes | Path to the minutes binary | | MINUTES_LANGUAGE | (auto) | Default language code (en, es, fr, …) |

Enable for audio understanding

Installing the plugin registers the provider, but OpenClaw only uses it once you wire it into tools.media.audio.models (same as any other audio provider — installation alone is not enough). One validated write:

openclaw config patch --stdin <<'EOF'
{ tools: { media: { audio: { enabled: true, models: [ { type: "provider", provider: "minutes", model: "whisper.cpp" } ] } } } }
EOF
openclaw plugins doctor          # should report "No plugin issues detected"
# restart the gateway so the change applies (e.g. systemctl --user restart openclaw-gateway)

Verify end-to-end without sending a real voice note:

openclaw infer audio transcribe --file /path/to/audio.wav --json
# routes through the minutes provider and prints the transcript

How it works

When OpenClaw receives a voice note attachment:

1. The audio buffer is written to a temp file.
2. minutes transcribe <temp> --json [-l language] is called locally — whisper.cpp transcribes it and prints a JSON envelope to stdout. No meeting files are written and no summarization runs.
3. data.text from the envelope is returned to OpenClaw for delivery.
4. The temp audio file is always deleted.

Nothing is persisted to disk beyond the temp audio file, which is removed as soon as transcription finishes (or fails).

Priority

autoPriority: { audio: 10 } — lower than all bundled cloud providers (openai=20, xai=25, …), so Minutes is tried first when installed.

Configuration

| Key | Type | Default | Description | |-----|------|---------|-------------| | minutesBin | string | "minutes" | Path to the minutes binary (or set MINUTES_BIN env var) | | language | string | (auto) | Default language code ("en", "es", "fr", …) |

Development

npm install
npm run build   # tsc
npm test        # vitest unit tests (no disk/network)

# Live test (requires minutes binary + whisper model):
MINUTES_OPENCLAW_LIVE=1 npm test   # transcribes test/fixtures/demo.wav

Design: self-contained community plugin

This is a community-owned package, independent of the Minutes core release train (per the maintainer's guidance on coupling — OpenClaw ships ~monthly with an evolving pluginApi, so the plugin tracks OpenClaw on its own cadence):

• Version 0.2.0, independent of the Minutes workspace version.
• No source imports from Minutes crates — the sole runtime dependency is the minutes binary, invoked via subprocess.
• Pinned to OpenClaw's pluginApi / pluginSdkVersion in package.json.
• Shells out to the stable minutes transcribe --json contract (silverstein/minutes#376, shipped in #380) instead of scraping a generated markdown memo. No meeting files are written by this plugin — earlier versions (<0.2.0) shelled out to minutes process -t memo and had a persistMemo option to keep that memo file; that option is gone since the new contract never creates one.