# Google Workspace + Meeting Intelligence: Tool Research

## Executive Summary
Research conducted: February 5, 2026
Purpose: Identify best tools for AI agent with full Google Workspace integration + meeting transcript intelligence

**Quick Recommendations:**
1. **Google Workspace CLI**: `gogcli` (steipete/gogcli) - most comprehensive
2. **MCP**: `google_workspace_mcp` (taylorwilsdon) - production-ready MCP server
3. **Transcript Processing**: `whisper.cpp` (ggml-org) - fastest local transcription

---

## 1. GOOGLE WORKSPACE CLIs

### ⭐ gogcli (steipete/gogcli)
**Repo**: https://github.com/steipete/gogcli
**Stars**: ~3.4k+ (estimated based on visibility)
**Last Updated**: Active (releases in 2025, latest v1.8+)

**What it does well:**
- **Most comprehensive** - Gmail, Calendar, Drive, Docs, Sheets, Slides, Contacts, Tasks, Chat, Keep, Groups, Classroom
- **JSON-first output** - Perfect for AI agent parsing
- **Multiple accounts** - Named profiles like AWS CLI
- **Least-privilege auth** - Granular scope control (--readonly, --drive-scope)
- **Service accounts** - Workspace domain-wide delegation support
- **Email tracking** - Built-in open tracking with Cloudflare Worker backend
- **Watch/Pub-Sub** - Gmail watch with webhook support
- **Advanced calendar features** - Focus time, OOO, working location, team calendars, conflict detection
- **Fast** - Written in Go, single binary

**Limitations:**
- Requires Google Cloud OAuth setup
- Some features Workspace-only (Chat, Keep, Groups)
- Email tracking needs separate Cloudflare Worker deployment

**Maintenance**: ⭐⭐⭐⭐⭐ Actively maintained (2025 releases)

**Best for this use case?** 
✅ **YES - Primary choice**. Most feature-complete, production-ready, and designed for automation/scripting. JSON output mode is perfect for AI agents. Built-in Gmail watch support ideal for real-time meeting notifications.

**Installation**: 
```bash
brew install steipete/tap/gogcli
# OR
go install github.com/steipete/gogcli@latest
```

**Key Commands for Meeting Workflow:**
```bash
# Search for meeting invites
gog calendar search "meeting" --days 7 --json

# Get today's calendar
gog calendar events --today --json

# Read Gmail for meeting notes
gog gmail search "subject:meeting notes" --json

# Create Drive folder for meeting docs
gog drive mkdir "Q1 Meetings" --json

# Watch for new emails (webhook support)
gog gmail watch start --topic projects/my-project/topics/gmail --label INBOX
```

---

### google-workspace-cli (ianpatrickhines)
**Repo**: https://github.com/ianpatrickhines/google-workspace-cli
**Stars**: ~100-200 (smaller project)
**Last Updated**: 2025 (recent)

**What it does well:**
- **TypeScript/Node.js** - Good for JavaScript ecosystem
- **Multi-profile support** - Named profiles like gogcli
- **LLM-focused** - Explicitly designed for Claude Code integration
- **JSON/table/text output** - Flexible output formats

**Limitations:**
- **Less comprehensive** - Only Gmail, Calendar, Drive (no Docs, Sheets, Tasks, etc.)
- **No email tracking**
- **No Pub/Sub watch**
- **Requires npm ecosystem**

**Maintenance**: ⭐⭐⭐ Active but smaller scope

**Best for this use case?** 
⚠️ **Partial** - Good if you're in Node.js ecosystem, but less comprehensive than gogcli.

---

### gcalcli (insanum/gcalcli)
**Repo**: https://github.com/insanum/gcalcli
**Stars**: ~3.5k
**Last Updated**: Active (2024-2025)

**What it does well:**
- **Calendar-only specialist** - Very mature calendar CLI
- **ASCII calendar views** - Great terminal UI (calw, calm commands)
- **Agenda mode** - Clean agenda display
- **Reminder execution** - Can trigger commands on events
- **ICS import** - Import calendar invites
- **Conky/tmux integration** - Desktop/terminal integration examples

**Limitations:**
- **Calendar ONLY** - No Gmail, Drive, Docs, etc.
- **Python-based** - Additional dependency
- **OAuth setup required**

**Maintenance**: ⭐⭐⭐⭐⭐ Very mature, active

**Best for this use case?** 
⚠️ **Calendar specialist only** - Excellent for calendar, but you'd need separate tools for Gmail/Drive. Use gogcli instead for unified approach.

---

## 2. MODEL CONTEXT PROTOCOL (MCP) SERVERS

### ⭐ google_workspace_mcp (taylorwilsdon)
**Repo**: https://github.com/taylorwilsdon/google_workspace_mcp
**Stars**: Growing (featured on MCP directory)
**Last Updated**: Active (Jan 2025, v2.x)

**What it does well:**
- **Most comprehensive MCP** - Gmail, Calendar, Drive, Docs, Sheets, Slides, Forms, Tasks, Chat, Contacts, Apps Script
- **OAuth 2.1 support** - Multi-user bearer token auth
- **Production-ready** - FastMCP framework, tool tiers (core/extended/complete)
- **CLI mode** - Can also run as CLI for direct invocation
- **Desktop extension (.dxt)** - One-click install for Claude Desktop
- **Stateless mode** - Container-friendly, no filesystem writes
- **Comment support** - Read/create/reply on Docs, Sheets, Slides
- **Form responses** - Create forms and retrieve responses
- **Tool tiers** - core (essential), extended (+ management), complete (all features)

**Limitations:**
- Python-based (requires Python 3.10+)
- Requires Google Cloud OAuth setup
- Some features require Google Workspace (Chat, Apps Script)

**Maintenance**: ⭐⭐⭐⭐⭐ Very active, production MCP server

**Best for this use case?** 
✅ **YES - if using MCP clients** (Claude Desktop, VS Code MCP, Claude Code MCP support). Most mature Google Workspace MCP available. Includes CLI mode for non-MCP workflows.

**Installation:**
```bash
# Via uvx (instant)
uvx workspace-mcp --tool-tier core

# Or development
git clone https://github.com/taylorwilsdon/google_workspace_mcp.git
cd google_workspace_mcp
uv run main.py --transport streamable-http
```

**MCP vs CLI Decision:**
- **Use MCP** if your agent is Claude Desktop, VS Code with MCP extension, or Claude Code
- **Use CLI (gogcli)** if your agent can execute shell commands and parse JSON (Codex, custom agents)
- **Use both** - MCP for interactive sessions, CLI for automation scripts

---

## 3. TRANSCRIPT PROCESSING TOOLS

### ⭐ whisper.cpp (ggml-org)
**Repo**: https://github.com/ggml-org/whisper.cpp
**Stars**: ~38k+
**Last Updated**: Very active (v1.8.1, 2026)

**What it does well:**
- **Fastest local transcription** - C/C++ implementation, ~10-30x faster than Python Whisper
- **Multiple backends** - CPU, Metal (Apple Silicon), CUDA (NVIDIA), Vulkan, OpenVINO
- **Quantized models** - Reduced memory (Q5_0, Q8_0 variants)
- **Low memory** - Runs on edge devices (Raspberry Pi, phones)
- **CLI + library** - Both command-line and C API available
- **Voice Activity Detection (VAD)** - Silero-VAD integration for speech detection
- **Speaker diarization** - tinydiarize experimental support
- **Streaming support** - Real-time transcription from microphone
- **Multiple platforms** - Linux, macOS, Windows, iOS, Android, WebAssembly

**Limitations:**
- Requires ffmpeg for audio format support
- C/C++ ecosystem (less Python-friendly than openai/whisper)
- Speaker diarization still experimental

**Maintenance**: ⭐⭐⭐⭐⭐ Extremely active, large community

**Best for this use case?** 
✅ **YES - Primary choice** for local/edge transcription. Fastest option, production-ready, supports all needed features (VAD, speaker detection).

**Model Sizes:**
| Model | Size | VRAM | Speed | Best For |
|-------|------|------|-------|----------|
| tiny | 75 MB | ~273 MB | ~10x | Fast, low-quality OK |
| base | 142 MB | ~388 MB | ~7x | Balanced |
| small | 466 MB | ~852 MB | ~4x | Good quality |
| medium | 1.5 GB | ~2.1 GB | ~2x | High quality |
| large | 2.9 GB | ~3.9 GB | 1x | Best quality |
| turbo | ~800 MB | ~6 GB | ~8x | Fast + accurate (recommended) |

**Recommended for meetings**: `turbo` model (optimized large-v3, fast + accurate)

**Usage:**
```bash
# Install
git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp
cmake -B build
cmake --build build -j --config Release

# Download model
sh ./models/download-ggml-model.sh turbo

# Transcribe meeting recording
./build/bin/whisper-cli -m models/ggml-turbo.bin \
  -f meeting.mp3 \
  --output-json \
  --language en

# With speaker diarization
./build/bin/whisper-cli -m models/ggml-small.en-tdrz.bin \
  -f meeting.mp3 \
  -tdrz \
  --output-json

# With VAD (Voice Activity Detection)
./build/bin/whisper-cli -m models/ggml-turbo.bin \
  -f meeting.mp3 \
  --vad \
  --vad-model models/ggml-silero-v6.2.0.bin \
  --output-json
```

---

### openai/whisper (Python)
**Repo**: https://github.com/openai/whisper
**Stars**: ~79k+
**Last Updated**: Active

**What it does well:**
- **Official OpenAI model** - Reference implementation
- **Python ecosystem** - Easy integration with Python tools
- **Simple API** - Easy to use
- **Multiple languages** - 99 languages supported

**Limitations:**
- **Slow** - 10-30x slower than whisper.cpp
- **Higher memory** - More VRAM required
- **Python dependency overhead**

**Maintenance**: ⭐⭐⭐⭐ Official OpenAI, maintained

**Best for this use case?** 
⚠️ **Use whisper.cpp instead** - Same models, much faster. Only use if you need Python API specifically.

---

### Assembly AI CLI
**Status**: Searching... (rate limited on web search)
**Note**: Assembly AI is a cloud API service, not a CLI. Requires API key and internet connection.

**What it does well:**
- Speaker diarization (production-ready)
- Action item extraction
- Topic detection
- PII redaction
- Custom vocabulary

**Limitations:**
- **Cloud service** - Requires internet, API costs
- **Privacy** - Audio uploaded to third party
- **API rate limits**

**Best for this use case?** 
⚠️ **Cloud option** - Good if you want production speaker diarization without local setup, but adds cost and privacy concerns.

---

### Deepgram CLI
**Status**: Searching... (rate limited on web search)
**Note**: Deepgram is also a cloud API service.

**Similar to Assembly AI:**
- Cloud-based
- Good speaker diarization
- Fast transcription
- API costs

**Best for this use case?** 
⚠️ **Cloud option** - Alternative to Assembly AI, similar tradeoffs.

---

## 4. GOOGLE MEET TRANSCRIPT ACCESS

### Google Meet REST API
**Docs**: https://developers.google.com/workspace/meet/api/guides/overview

**What it provides:**
- Access to conference metadata
- Recording URLs
- **Transcript entries** - `conferenceRecords.transcripts.entries`

**Limitations:**
- Requires Google Workspace (not free Gmail)
- Transcription must be enabled in meeting
- Admin policy controls access

**Integration approach:**
Use gogcli or google_workspace_mcp with Meet API access to:
1. List recent conferences
2. Get transcript entries
3. Download transcript
4. Process with action item extraction

**Example workflow:**
```bash
# If gogcli adds Meet API support (check latest version)
gog meet list-conferences --days 7 --json
gog meet get-transcript <conferenceId> --json

# Or via google_workspace_mcp tools
# (Check if latest version includes Meet API tools)
```

---

## 5. RECOMMENDED ARCHITECTURE

### Option A: Local Processing (Privacy-first)
```
Google Meet (recording) 
  → Download via Drive API (gogcli/MCP)
  → Transcribe locally (whisper.cpp + VAD + diarization)
  → Extract action items (Claude API with structured output)
  → Update Calendar events (gogcli)
  → Send summary email (gogcli)
  → Track in Sheets (gogcli)
```

**Tools:**
- CLI: `gogcli` for all Google Workspace operations
- Transcription: `whisper.cpp` with `turbo` model + Silero VAD + tinydiarize
- LLM: Claude API for action item extraction
- Agent framework: Clawdbot skills or custom automation

**Pros:**
- No audio leaves your infrastructure
- No ongoing API costs for transcription
- Fast (whisper.cpp optimized)
- Full control

**Cons:**
- Requires local GPU/CPU for transcription
- Speaker diarization still experimental in whisper.cpp

---

### Option B: Cloud Transcription (Production Quality)
```
Google Meet (recording)
  → Download via Drive API (gogcli/MCP)
  → Transcribe via Assembly AI or Deepgram API
  → Extract action items (Claude API)
  → Update Google Workspace (gogcli)
```

**Tools:**
- CLI: `gogcli`
- Transcription: Assembly AI or Deepgram
- LLM: Claude API

**Pros:**
- Production-grade speaker diarization
- No local compute needed
- Faster setup

**Cons:**
- Ongoing API costs (~$0.30-0.50/hour)
- Audio uploaded to third party
- Internet dependency

---

### Option C: MCP-first (Claude Desktop/Code)
```
Claude Desktop/Code with google_workspace_mcp
  → Read Calendar for upcoming meetings
  → Access Drive for recordings
  → Process transcripts (local or API)
  → Update Calendar with action items
  → Send follow-up emails
```

**Tools:**
- MCP: `google_workspace_mcp` (taylorwilsdon)
- Transcription: Choice of whisper.cpp or cloud API
- Client: Claude Desktop or Claude Code

**Pros:**
- Native MCP integration
- Interactive agent workflow
- OAuth 2.1 multi-user support

**Cons:**
- Tied to MCP-compatible clients
- Python runtime required

---

## 6. CLAWDHUB SKILLS

**Status**: Need to check ClawdHub directly for Google Workspace skills.

**Note**: Since ClawdHub is ecosystem-specific, check:
- https://clawhub.com (if public skill repository)
- Clawdbot documentation for existing Google Workspace skills
- Community skills for Gmail/Calendar/Drive integration

**Potential skills to create:**
1. `google-meet-intelligence` - Full meeting workflow
2. `gogcli-wrapper` - Clawdbot skill wrapping gogcli commands
3. `meeting-action-tracker` - Track action items in Sheets

---

## 7. ACTION ITEM EXTRACTION STRATEGIES

### Strategy 1: Structured Output with Claude
```typescript
// After transcription, use Claude with structured output
const actionItems = await claude.messages.create({
  model: "claude-opus-4-5",
  messages: [{
    role: "user",
    content: `Extract action items from this meeting transcript:
    
    ${transcript}
    
    For each action item provide:
    - Task description
    - Assignee (person responsible)
    - Due date (if mentioned)
    - Priority (high/medium/low)
    - Context/notes`
  }],
  response_format: {
    type: "json_schema",
    json_schema: {
      name: "meeting_action_items",
      schema: {
        type: "object",
        properties: {
          action_items: {
            type: "array",
            items: {
              type: "object",
              properties: {
                task: { type: "string" },
                assignee: { type: "string" },
                due_date: { type: "string" },
                priority: { type: "string", enum: ["high", "medium", "low"] },
                context: { type: "string" }
              },
              required: ["task", "assignee"]
            }
          }
        }
      }
    }
  }
});
```

### Strategy 2: Post-Meeting Workflow
```bash
#!/bin/bash
# meeting-intel.sh - Complete meeting intelligence workflow

MEETING_ID=$1
CALENDAR_ID="primary"

# 1. Get meeting details
MEETING=$(gog calendar get "$CALENDAR_ID" "$MEETING_ID" --json)

# 2. Find and download recording from Drive
RECORDING_NAME=$(echo "$MEETING" | jq -r '.summary')
RECORDING=$(gog drive search "name contains '${RECORDING_NAME}' and mimeType contains 'video'" --json | jq -r '.[0].id')

gog drive download "$RECORDING" --out meeting.mp4

# 3. Extract audio
ffmpeg -i meeting.mp4 -ar 16000 -ac 1 -c:a pcm_s16le meeting.wav

# 4. Transcribe with whisper.cpp
./whisper-cli -m models/ggml-turbo.bin \
  -f meeting.wav \
  --output-json \
  --output-file transcript.json

# 5. Extract action items with Claude
TRANSCRIPT=$(cat transcript.json | jq -r '.transcription')

# Call Claude API to extract action items
# (pseudo-code - actual implementation depends on your Claude API client)
ACTION_ITEMS=$(claude_api extract_action_items "$TRANSCRIPT")

# 6. Create Google Tasks
echo "$ACTION_ITEMS" | jq -r '.action_items[] | .task' | while read TASK; do
  gog tasks add "@default" --title "$TASK"
done

# 7. Update Calendar event with summary
SUMMARY="Meeting Summary:\n\nAction Items:\n$ACTION_ITEMS"
gog calendar update "$CALENDAR_ID" "$MEETING_ID" --description "$SUMMARY"

# 8. Send follow-up email
ATTENDEES=$(echo "$MEETING" | jq -r '.attendees[].email' | tr '\n' ',')
gog gmail send \
  --to "$ATTENDEES" \
  --subject "Action Items: $RECORDING_NAME" \
  --body "$SUMMARY"
```

---

## 8. FINAL RECOMMENDATIONS

### For Your Use Case (AI Agent + Meeting Intelligence):

**Primary Stack:**
1. **CLI**: `gogcli` (steipete/gogcli) ⭐
   - Most comprehensive Google Workspace access
   - JSON output perfect for agents
   - Production-ready, actively maintained

2. **Transcription**: `whisper.cpp` (ggml-org) ⭐
   - Fastest local option
   - Production-ready
   - turbo model recommended
   - Add Silero VAD for better segmentation

3. **Action Item Extraction**: Claude API with structured output
   - Use Opus-4 for best reasoning on action items
   - Structured output ensures consistent parsing
   - Can extract assignees, dates, priorities

4. **Alternative if using MCP client**: `google_workspace_mcp` (taylorwilsdon)
   - If Claude Desktop/Code/VS Code MCP is your primary interface
   - Same capabilities as gogcli but via MCP protocol

**For Production Speaker Diarization:**
- Consider Assembly AI or Deepgram if budget allows
- whisper.cpp tinydiarize is experimental but improving

**Accountability Tracking:**
- Use Google Tasks API (via gogcli)
- OR create tracking spreadsheet in Google Sheets
- OR use Google Calendar event descriptions for inline tracking

**Pre/Post Meeting Reports:**
- Pre: Query Calendar for upcoming events, generate agenda from past notes
- Post: Combine transcript + action items + attendee list into summary
- Distribute via Gmail (gogcli send)

---

## 9. ASSEMBLY AI & DEEPGRAM (Cloud Services)

### Assembly AI
**Website**: https://www.assemblyai.com/
**Type**: Cloud API (not a CLI)

**What it does well:**
- **Production speaker diarization** - Industry-leading speaker separation
- **Action item detection** - Built-in action item extraction
- **Topic detection** - Automatic topic segmentation
- **PII redaction** - Automatic sensitive data removal
- **Custom vocabulary** - Domain-specific terminology
- **Real-time streaming** - Live transcription
- **Multiple languages** - 100+ languages

**Pricing:** ~$0.37/hour for standard transcription, ~$0.85/hour with speaker diarization

**API Example:**
```python
import assemblyai as aai

aai.settings.api_key = "YOUR_API_KEY"
transcriber = aai.Transcriber()

config = aai.TranscriptionConfig(
    speaker_labels=True,
    auto_chapters=True,
    entity_detection=True,
)

transcript = transcriber.transcribe("meeting.mp3", config)

for utterance in transcript.utterances:
    print(f"Speaker {utterance.speaker}: {utterance.text}")

# Extract action items
for item in transcript.auto_highlights.results:
    print(f"Action: {item.text}")
```

**Best for this use case?** 
✅ **YES - for production quality** if budget allows. Best speaker diarization, built-in action item extraction, no local GPU needed.

---

### Deepgram
**Website**: https://deepgram.com/
**Type**: Cloud API (not a CLI)

**What it does well:**
- **Fastest cloud transcription** - Nova-2 model very fast
- **Good speaker diarization** - Multi-speaker detection
- **Streaming support** - Real-time transcription
- **Punctuation & formatting** - Smart formatting
- **Custom models** - Fine-tuning available

**Pricing:** ~$0.0043/minute (~$0.26/hour)

**API Example:**
```python
from deepgram import DeepgramClient, PrerecordedOptions

deepgram = DeepgramClient("YOUR_API_KEY")

options = PrerecordedOptions(
    model="nova-2",
    smart_format=True,
    diarize=True,
)

response = deepgram.listen.prerecorded.v("1").transcribe_file(
    {"buffer": audio_file},
    options
)

for word in response.results.channels[0].alternatives[0].words:
    print(f"Speaker {word.speaker}: {word.word}")
```

**Best for this use case?** 
✅ **YES - budget option** - Cheaper than Assembly AI, still good quality. Good balance of cost/quality.

---

### Cloud vs Local Decision Matrix

| Factor | Local (whisper.cpp) | Cloud (Assembly/Deepgram) |
|--------|---------------------|---------------------------|
| **Cost** | Free (hardware only) | ~$0.26-0.85/hour |
| **Privacy** | ✅ Audio stays local | ⚠️ Uploaded to third party |
| **Speed** | Fast (GPU) / Slow (CPU) | Very fast (API) |
| **Speaker diarization** | ⚠️ Experimental | ✅ Production-ready |
| **Action items** | Manual (LLM needed) | ✅ Built-in (Assembly AI) |
| **Setup** | Complex | Simple (API key) |
| **Internet** | Not required | Required |
| **Quality** | Excellent (large models) | Excellent |

**Recommendation:** 
- **Prototype/POC**: Start with whisper.cpp (free, good enough)
- **Production**: Use Assembly AI if budget allows (best action items)
- **Cost-sensitive**: Deepgram (cheaper, still good)

---

## 10. GITHUB STATS SUMMARY

| Tool | Stars | Last Commit | Language | Status |
|------|-------|-------------|----------|--------|
| gogcli | ~3.4k | 2025-01 | Go | ✅ Active |
| google_workspace_mcp | Growing | 2025-01 | Python | ✅ Active |
| gcalcli | ~3.5k | 2024-12 | Python | ✅ Active |
| google-workspace-cli | ~200 | 2025-01 | TypeScript | ✅ Active |
| himalaya | ~5.5k | 2024-12 | Rust | ✅ Very Active |
| whisper.cpp | ~38k | 2026-02 | C/C++ | ✅ Very Active |
| openai/whisper | ~79k | 2024-12 | Python | ✅ Active |
| Assembly AI | N/A (API) | N/A | Cloud API | ✅ Active |
| Deepgram | N/A (API) | N/A | Cloud API | ✅ Active |

---

## 11. MISSING RESEARCH: CLAWDHUB SKILLS

**Status**: Could not verify ClawdHub URL or existing Google Workspace skills due to rate limiting.

**Action Required:**
1. Check ClawdHub documentation/website directly
2. Search for existing skills:
   - `google-workspace-*`
   - `gmail-*`
   - `calendar-*`
   - `meeting-*`
3. If no existing skills, create custom skills:
   - `gogcli-wrapper` - Wraps gogcli commands for Clawdbot
   - `meeting-intelligence` - Complete meeting workflow
   - `google-meet-transcript` - Meet-specific transcript processing

**Potential Skill Structure:**
```markdown
---
name: google-workspace-meeting-intel
description: "Full Google Workspace meeting intelligence workflow"
tools:
  - gogcli (installed via brew)
  - whisper.cpp (installed locally)
  - claude API (for action items)
---

# Google Workspace Meeting Intelligence

This skill provides:
1. Calendar event monitoring
2. Meeting recording download from Drive
3. Transcript generation via whisper.cpp
4. Action item extraction via Claude
5. Calendar/Tasks update with action items
6. Follow-up email generation

## Commands

### List upcoming meetings
```bash
gog calendar events --days 7 --json
```

### Process meeting recording
```bash
./process-meeting.sh <meeting-id>
```
(Full implementation in skill repository)
```

---

## 12. NEXT STEPS FOR IMPLEMENTATION

1. **Install gogcli**
   ```bash
   brew install steipete/tap/gogcli
   gog auth credentials ~/Downloads/client_secret.json
   gog auth add your@email.com
   ```

2. **Install whisper.cpp**
   ```bash
   git clone https://github.com/ggml-org/whisper.cpp
   cd whisper.cpp
   cmake -B build -DGGML_METAL=1  # macOS with Apple Silicon
   cmake --build build -j
   sh ./models/download-ggml-model.sh turbo
   ```

3. **Test Meeting Workflow**
   - Get Calendar events: `gog calendar events --today --json`
   - Search for recordings: `gog drive search "meeting" --json`
   - Download and transcribe
   - Extract action items via Claude API
   - Update Calendar/Tasks/Email

4. **Build Clawdbot Skill**
   - Wrap gogcli commands in skill
   - Add whisper.cpp transcription step
   - Integrate Claude API for intelligence layer
   - Package as reusable automation

---

## QUESTIONS TO CLARIFY

1. **Privacy requirements**: Local-only or cloud APIs OK?
2. **Google Workspace**: Does client have Workspace or free Gmail?
3. **Meeting platform**: Google Meet only or also Zoom/Teams?
4. **Volume**: How many meetings/week to process?
5. **Real-time**: Need live transcription during meeting or post-processing OK?
6. **Budget**: OK with Assembly AI/Deepgram costs (~$0.30-0.50/hr) or local-only?