Skip to main content
Glama

Qulaq («құлақ» — ear)

An agent for parents in Kazakhstan whose school and kindergarten live in WhatsApp groups. It listens to the voice notes you missed, pulls out dates, deadlines, money and things to bring, and gives you a short "don't miss this" report in the evening or once a week.

Status: Submitted to the Meta Global AI Developer Hackathon (Indie Developer track).

Two ways in (no unofficial WhatsApp access)

  1. Export chat — in WhatsApp: group → ⋮ → More → Export chat → Include media. Upload the .zip. WhatsApp keeps only the latest ~10,000 messages with media; older media is marked <Media omitted>, and Qulaq reports what it could not hear instead of guessing.

  2. Forward — forward voice notes and messages to the Qulaq number (WhatsApp Business Cloud API). Planned.

Qulaq does not read groups by scraping or by unofficial libraries.

Related MCP server: whatsapp-profiles

Pipeline

export / forward → ffmpeg (mono 16-bit WAV) → Muse Voice Transcribe → Muse Spark extracts events → code rules decide what is important (date, payment, bring item, RSVP) → digest.

The model extracts fields and writes text; importance is decided by plain code rules.

Never a confident wrong number. Speech recognition can mishear "бес жүз" (500) as "үш жүз" (300). Every voice note is read three ways (language hints Kazakh+Russian, none, and one single-language hint). If amount, date or time differ, the item is marked ⚠️ with all variants and the original words. A payment with no readable amount is never dropped: plain code adds a flagged item with the quote.

Agent actions (MCP)

python -m qulaq.mcp_server exposes tools to any MCP client (Claude Desktop, etc.): read_chat_export, list_events, explain_doubts, export_calendar (.ics with a reminder at 18:00 the day before), check_conflicts. Doubtful events are not exported until the parent confirms them.

Claude Desktop config (claude_desktop_config.json):

{"mcpServers": {"qulaq": {"command": "python", "args": ["-m", "qulaq.mcp_server"],
  "cwd": "C:\\path\\to\\qulaq"}}}

Run

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env        # add MODEL_API_KEY
pytest
python scripts/check_asr.py voice.opus --dry-run     # test ffmpeg only
python scripts/check_asr.py ru.opus kk.opus mix.opus # real speech check

Needs ffmpeg on PATH. See scripts/sample_texts.md for the speech check.

Privacy

Real chats involve children and teachers. Nothing real is committed (.gitignore); the demo uses invented chats only.

What we found testing it

Tested on synthetic school chats recorded by one speaker (about six voice notes in Russian, Kazakh and mixed speech). This is a small sample, not a benchmark.

  • Russian voice notes were transcribed very well.

  • Kazakh needs a language hint. With no hint the transcript came back as garbled Latin text; with Kazakh+Russian hints it was mostly correct Cyrillic.

  • Numbers are the weak spot. "бес жүз" (500) came back as 300 or as 5 depending on the hint and the run. This is why every voice note is read three ways and disagreements are shown with the original words.

  • Results change between runs. The same recording gave a correct "жұмаға дейін" (until Friday) in one run and a garbled word, so no date, in another. Qulaq shows "no date" or ⚠️ instead of guessing.

  • Bugs that testing found in our own code: a payment mentioned without an amount was silently lost (now a flagged item); a doubtful date was "confirmed" by an unrelated message (removed); one disputed date flagged unrelated items in the same voice note (comparison is now per item); two events on the same day were merged and one disappeared (events with different times now stay separate).

  • Known gaps: the forward-to-WhatsApp path is not built; spoken amounts are interpreted by the model, not by code; we have not tested real chats, only invented ones.

Other WhatsApp chat summarizers exist (text-focused GitHub summarizers, ChatBrief, ThreadRecap, Kaption, WhatsApp's own message summaries, Amar Khobor). Qulaq differs by focusing on voice notes in Kazakh/Russian school groups and on deadlines and actions rather than a generic summary.

Provenance

All code in this repository was written during the hackathon period (Oct 27 – Nov 4, 2026). The idea grew out of an earlier project (School Bridge); no code from it is reused.

Demo

https://youtu.be/R85CqYKlhLo

License

MIT

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables local querying of WhatsApp message history via SQLite, with profile separation, media transcription, and operational summaries.
      1
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      A read-only, local MCP server that enables Claude Desktop and Claude Code to brief users on selected WhatsApp chats, surfacing mentions, urgent items, follow-ups, bills, and payment screenshots through local OCR while keeping data encrypted on the user's PC.
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables turning pasted syllabus text into dated calendar events, expanding recurring classes while skipping holidays and reading weeks, building valid .ics files with reminders, and merging calendars to surface conflicts. All processing runs locally without login or network access.
      4
      103 npm
      MIT