Meeting Transcription

This use case walks through examples/meeting_transcriber.py — a complete, ~280-line script that joins a Discord server voice channel, transcribes every speaker live with their display name, writes to a local markdown file, and on exit asks Claude for a Summary / Key decisions / Action items / Open questions breakdown.

It exists because the most common voice ask isn’t “make the bot speak” — it’s “give me a searchable record of what was said”. This is that.

What it does

• Connects to a server voice channel by ID using a regular bot token
• Streams audio to Deepgram (or OpenAI Whisper) and prints each finalised utterance to console with the speaker’s display name
• Appends every line to ~/.discli/transcripts/meeting-<YYYYMMDD-HHMMSS>.md
• On Ctrl+C, sends the full transcript to Claude via the Agent SDK and appends a structured summary to the same file

What it does not do

• DM voice or group-DM voice — bot tokens cannot join those per Discord’s API
• Live summary during the call — only on exit
• Speaker recognition across calls — display names come from the current guild membership

Setup

pip install 'discord-cli-agent[voice,deepgram]' claude-agent-sdk
discli config set token YOUR_BOT_TOKEN
export DEEPGRAM_API_KEY=...

Verify with:

discli doctor

You want green ticks across CORE and VOICE, plus DEEPGRAM_API_KEY set under STT.

Running it

Get the voice channel ID (right-click in Discord with Developer Mode on, or discli channel list --type voice):

python examples/meeting_transcriber.py 1016638171854938152

To use OpenAI Whisper instead of Deepgram:

export OPENAI_API_KEY=...
python examples/meeting_transcriber.py 1016638171854938152 --stt openai

During the meeting, the console fills with speaker-labelled lines:

Connected as MyBot#1234 (12345)
Listening to #standup. Transcript: /home/me/.discli/transcripts/meeting-20260514-103000.md
Press Ctrl+C to stop and generate a summary.

- **[10:30:14] Roy:** ok let's go around — what did everyone do yesterday
- **[10:30:22] Sara:** finished the auth migration, started on the rate limiter
- **[10:30:35] Roy:** nice — anything blocking

Press Ctrl+C to end. You’ll see:

Stopping listener…
Generating summary from 47 line(s)…
Summary cost: $0.0084

=== Meeting Summary ===

## Summary
Standup covering yesterday's work and today's plan. Sara finished the auth
migration; rate limiter is next.

## Key decisions
- Move forward with read-path-first for the migration this sprint.

## Action items
- Sara: finish rate limiter today.
- Roy: write the migration runbook by Friday.

## Open questions
- Do we backfill old sessions or expire them?

Full transcript + summary saved to: /home/me/.discli/transcripts/meeting-20260514-103000.md

The output file

The transcript file is plain markdown — readable, grep-able, easy to commit to a private repo or paste into a doc:

# Meeting transcript — 2026-05-14 10:30:00

- **Channel:** #standup (Engineering)
- **Bot:** MyBot#1234

## Transcript

- **[10:30:14] Roy:** ok let's go around — what did everyone do yesterday
- **[10:30:22] Sara:** finished the auth migration, started on the rate limiter
- **[10:30:35] Roy:** nice — anything blocking

---

## Summary
...

How it works (~280 LOC)

The script is intentionally small — it reuses VoiceEngine from discli rather than rebuilding the voice stack:

from discli.voice_engine import VoiceEngine

engine = VoiceEngine(config={"stt_provider": stt_provider})
engine.set_event_handler(on_event)
await engine.connect(channel)
await engine.listen_start(channel.guild.id)

The on_event callback receives every voice_speech_detected event, resolves the user ID to a display name via the guild membership, appends a markdown line to the transcript file, and prints it to console:

def on_event(event: dict) -> None:
    if event.get("event") != "voice_speech_detected" or not event.get("is_final"):
        return
    uid = int(event.get("user_id") or 0)
    if uid == client.user.id:
        return
    name = resolve_name(uid)
    ts = datetime.now().strftime("%H:%M:%S")
    line = f"- **[{ts}] {name}:** {event['text']}"
    transcript_lines.append(line)
    print(line, flush=True)
    with path.open("a", encoding="utf-8") as f:
        f.write(line + "\n")

On Ctrl+C, the finally block stops the listener, disconnects, and pipes the whole transcript through Claude:

async with sdk.ClaudeSDKClient(options) as claude:
    await claude.query(prompt)
    async for msg in claude.receive_response():
        ...

SUMMARY_SYSTEM_PROMPT instructs Claude to emit exactly four sections (Summary, Key decisions, Action items, Open questions) using the display names already in the transcript.

Customising

Common changes:

Privacy

The transcript captures everyone in the channel by display name. Treat it like a recording:

• Tell participants the bot is transcribing — most jurisdictions require consent for recording
• Store the transcript file somewhere private (the default ~/.discli/transcripts/ is your local home)
• If you post summaries to a public channel, scrub any sensitive content first

Where to go next

• See Voice for the full voice stack overview
• See Building Agents if you want a persistent agent that transcribes alongside other tasks
• The complete file is in the repo at examples/meeting_transcriber.py