ai-meeting-notes-agent

Voice transcription meets autonomous AI agent. Like OpenClaw for your voice memos — transcribe, chat, and let an AI agent work with your files, all through Telegram.

Free and open source.

Self-host with your own API keys. Free forever. MIT licensed.

notesly.demo.0219.mp4

Core Value

Transcription + AI agent in one bot. Two things that belong together but nobody combines:

1. Transcription — send a voice memo, get a transcript with speaker labels and timestamps. Any language, any length. That's the entry point.
2. Autonomous AI agent — the same bot holds a conversation, summarizes long recordings, and spawns a Claude Code Agent that reads your stored files to answer questions. Like OpenClaw, but specialized for voice-first workflows.

Your Telegram chat becomes a personal AI workspace where voice memos, transcripts, files, and conversations all live together.

The Problem

• iPhone Voice Memos is the best recorder. One tap from the lock screen. No app to open, no meeting to join. It never crashes. Nothing else comes close.
• But voice memos are a dead end. You have hundreds on your phone. You'll never listen to them again.
• Recording apps solve the wrong problem. Granola, Otter, Fireflies — they replace Voice Memos instead of building on it.
• AI assistants don't know your context. ChatGPT, OpenClaw — powerful, but they can't search last Tuesday's meeting or find what your manager said about the deadline.
• What's missing is the bridge. Transcription that feeds into an AI agent with memory of everything you've recorded and stored.

What It Does

Send anything to the Telegram bot. It figures out what to do.

You send	Bot does
Voice memo or audio/video file	Transcribes with speaker labels + timestamps. Long recordings get an AI summary.
Text message	AI chat — ask questions, get help, have a conversation.
Text about your files — "what did we discuss yesterday?"	Searches your stored transcripts and files, answers with context.
Any other file (PDF, image, doc...)	Stores it for you. Ask about it later.

All files — audio, transcripts, uploads — are stored locally and optionally synced to S3. Survives container restarts.

How It Works

1. Record with Apple Voice Memos (or any voice recorder on your phone)
2. Share the recording directly to Telegram — no opening another app, no exporting, no emailing yourself
3. Read the full transcript — with speaker labels and timestamps — right in the chat

Works with any language. Handles multiple speakers. Transcripts come back in under a minute.

Prerequisites

You need two API keys to get started. Both are free:

Key	What it's for	Where to get it	Cost
TELEGRAM_BOT_TOKEN	Receive and reply to messages	Message @BotFather on Telegram, send /newbot	Free
ASSEMBLY_API_KEY	Transcription with speaker labels	assemblyai.com/app/account	Free tier included

Optional keys (unlock more features)

Key	What it unlocks
OPENAI_API_KEY	AI chat + summarization. Works with any OpenAI-compatible API (OpenAI, OpenRouter, DigitalOcean, etc.)
OPENAI_BASE_URL	Custom endpoint (default: https://api.openai.com/v1)
OPENAI_MODEL	Model for chat + summarization (default: gpt-4o-mini)
GLM_API_KEY	File analysis via Claude Agent SDK. Ask questions about your stored files.
GLM_MODEL	Model for file analysis (default: glm-4-plus)
ANTHROPIC_BASE_URL	Anthropic-compatible endpoint (default: https://api.z.ai/api/anthropic). Works with Z.AI, Anthropic, or any compatible provider.
S3_BUCKET	S3 storage sync — all files mirrored to S3, restored on container restart
BOT_NAME	Storage prefix (default: transcribe-bot)

Getting Started

1. Create a Telegram Bot (2 minutes)

1. Open Telegram and search for @BotFather
2. Send /newbot
3. Pick a name for your bot (e.g., "My Transcriber")
4. Pick a username (e.g., my_transcriber_bot)
5. BotFather gives you an API token — copy it

2. Run the Bot

git clone <repo-url> && cd ai-meeting-notes-agent
cp .env.example .env

Edit .env and fill in your API keys:

ASSEMBLY_API_KEY=your_assemblyai_key
TELEGRAM_BOT_TOKEN=your_bot_token

Start the bot:

uv run telegram_bot.py

That's it. Send a voice memo to your bot on Telegram and get a transcript back.

Deploy

Docker (any server)

docker compose up -d

The bot uses polling (no inbound ports needed), so it runs anywhere Docker runs — a $5 VPS, a Raspberry Pi, or your laptop.

AWS ECS (via GitHub Actions)

Fork this repo, add secrets in GitHub repo settings, and push. It deploys automatically.

Required secrets:

GitHub Secret	Value
AWS_ACCESS_KEY_ID	Your AWS access key
AWS_SECRET_ACCESS_KEY	Your AWS secret key
ASSEMBLY_API_KEY	Your AssemblyAI key
TELEGRAM_BOT_TOKEN	Your bot token from BotFather

Optional secrets (for AI + storage features):

GitHub Secret	Value
OPENAI_API_KEY	OpenAI-compatible API key for chat + summarization
OPENAI_BASE_URL	Custom endpoint URL
OPENAI_MODEL	Model name
GLM_API_KEY	API key for file analysis
GLM_MODEL	Model name for file analysis
ANTHROPIC_BASE_URL	Anthropic-compatible endpoint URL
S3_BUCKET	S3 bucket name for file sync
BOT_NAME	Storage prefix

Every push to main builds and deploys to ECS Fargate. You can also trigger it manually from the Actions tab.

Railway (one click)

Set your API keys as environment variables in the Railway dashboard.

Already have a backlog?

Got a folder full of voice memos you never transcribed? Do them all at once:

uv run transcribe.py -f /path/to/recordings/

Supported Formats

Voice notes from Telegram, iPhone Voice Memos, and any standard audio/video format: .m4a, .mp3, .ogg, .wav, .mp4, .mov, and more.

Architecture

Like OpenClaw, the bot uses a messaging platform (Telegram) as the primary interface, with modular AI capabilities behind it:

• Telegram bot (telegram_bot.py) — message router: voice → transcription, text → conversation, files → storage
• Transcription — AssemblyAI with speaker diarization, auto language detection, multi-format support
• Conversation — OpenAI-compatible LLM for chat, summarization, and Q&A
• Claude Code Agent — autonomous agent (via Claude Agent SDK) that reads your stored files and answers questions with full context
• Storage — unified data/{bot_name}/YYYY/MM/DD/ structure, identical paths locally and on S3
• S3 sync — bidirectional: pulls from S3 on startup, pushes after every write
• Web dashboard — React pipeline visualization showing module status, deployment info, and live configuration

What's Next

• Personalized notes — each participant gets notes relevant to them
• Calendar integration — auto-match recordings to meetings
• Team workspaces — shared transcripts across a group

Technical Details

See TECHNICAL.md for detailed architecture, configuration, Google Drive watcher, and deployment instructions.