Join meeting

CallingClaw — AI Meeting Assistant

Join meetings as an active AI participant — not just a note-taker.

CallingClaw is fundamentally different from other meeting bots. It uses OpenAI Realtime API for sub-second voice response (~300ms), Playwright for interactive screen sharing (click, scroll, navigate real websites), and a dual-system architecture (System 1: instant voice + System 2: deep reasoning) for natural conversation.

Prerequisites

• CallingClaw Desktop running on localhost:4000
  • Download from callingclaw.com
  • Or build from source: cd callingclaw-desktop && npm install && npm start
• OpenAI API key with Realtime API access
• Google account signed into Chrome (for Google Meet)

First-Time Setup

If this is your first time, the agent should run through this checklist:

1. Check CallingClaw is running: curl http://localhost:4000/api/status
2. Check API keys: curl http://localhost:4000/api/keys
3. Check Google OAuth: curl http://localhost:4000/api/google/auth-status
4. Check permissions: curl http://localhost:4000/api/onboarding/permissions
5. Full readiness: curl http://localhost:4000/api/onboarding/ready

If any check fails, use the corresponding setup endpoints to fix it. See the MCP server (mcp/) for the full tool-based setup flow.

Quick Start

Join a meeting

curl -X POST http://localhost:4000/api/meeting/join \
  -H "Content-Type: application/json" \
  -d '{"meetUrl": "https://meet.google.com/abc-defg-hij"}'

This takes 30-60 seconds (Playwright launches Chrome, navigates to Meet, joins).

Say something

curl -X POST http://localhost:4000/api/voice/text \
  -H "Content-Type: application/json" \
  -d '{"text": "Hey everyone, I'\''m CoCo, your AI meeting assistant."}'

The text goes through OpenAI Realtime API and is spoken naturally — not robotic TTS. Response latency is ~300ms.

Share screen

curl -X POST http://localhost:4000/api/screen/share \
  -H "Content-Type: application/json" \
  -d '{"url": "https://your-slides.com/deck"}'

Unlike other bots, CallingClaw uses Playwright — it can interact with the page (click buttons, scroll, fill forms, navigate). This makes it ideal for live product demos and sales presentations.

Inject context

curl -X POST http://localhost:4000/api/voice/inject \
  -H "Content-Type: application/json" \
  -d '{"text": "The customer'\''s company is Acme Corp. They use Salesforce and Jira. Budget is $50K."}'

Context injection feeds information to the AI without speaking it aloud. The AI absorbs it silently and uses it to answer questions naturally.

Get transcript

curl http://localhost:4000/api/meeting/transcript

Leave meeting

curl -X POST http://localhost:4000/api/meeting/leave

MCP Server (Recommended)

For the best experience, use the MCP server which exposes all CallingClaw capabilities as MCP tools — compatible with Claude Code, Cursor, Claude Desktop, VS Code Copilot, and any MCP client.

{
  "mcpServers": {
    "callingclaw": {
      "command": "npx",
      "args": ["-y", "callingclaw-mcp"]
    }
  }
}

Or run from this repo:

cd mcp && npm install && npm run build && node dist/index.js

20 tools available: join_meeting, leave_meeting, speak, present_url, get_transcript, inject_context, send_chat_message, get_status, get_action_items, set_voice_provider, check_health, check_api_keys, set_api_keys, check_google_auth, setup_google_oauth, google_chrome_login, get_config, set_config, check_capabilities, check_audio.

API Reference

Meeting Lifecycle

Endpoint	Method	Description
/api/meeting/join	POST	Join a meeting (30-60s)
/api/meeting/leave	POST	Leave meeting
/api/meeting/status	GET	Meeting state
/api/meeting/transcript	GET	Full transcript
/api/meeting/summary	POST	Generate summary
/api/meeting/validate	POST	Validate meeting URL
/api/meeting/prepare	POST	Prepare context/brief
/api/meeting/talk-locally	POST	Voice chat without meeting

Voice

Endpoint	Method	Description
/api/voice/text	POST	Speak text naturally
/api/voice/inject	POST	Inject silent context
/api/voice/start	POST	Start voice session
/api/voice/stop	POST	Stop voice session
/api/voice/instructions	GET/POST	Get/set AI instructions

Screen Sharing

Endpoint	Method	Description
/api/screen/share	POST	Share URL or screen
/api/screen/stop	POST	Stop sharing
/api/screen/scroll	POST	Scroll the shared page
/api/screen/snapshot	GET	DOM snapshot
/api/screen/present	POST	Start guided presentation

Setup & Config

Endpoint	Method	Description
/api/status	GET	Health check
/api/keys	GET/POST	API key management
/api/config	GET/POST	Configuration
/api/google/auth-status	GET	Google OAuth status
/api/google/scan	GET	Scan for OAuth credentials
/api/google/apply	POST	Apply found credentials
/api/google/chrome-login	POST	Open Chrome for sign-in
/api/onboarding/ready	GET	Full readiness check
/api/onboarding/permissions	GET	macOS permission status
/api/capabilities	GET	Available features

Calendar

Endpoint	Method	Description
/api/calendar/events	GET	List upcoming events
/api/calendar/create	POST	Create calendar event
/api/scheduler/schedule	POST	Schedule auto-join

How CallingClaw Works

User speaks → Mic → OpenAI Realtime API → AI responds (~300ms)
                         ↕
                    System 2 (Haiku/Opus)
                    Deep reasoning when needed
                    Context retrieval
                    Tool execution
                         ↕
                    Playwright Browser
                    Screen share + interact
                    Navigate + click + scroll

Dual-System Architecture

• System 1 (Realtime API): Handles instant conversation. Sub-second response. Always-on during meetings. Natural voice with emotion.
• System 2 (Haiku/Sonnet): Handles complex tasks in background. Knowledge retrieval, file search, code review. Results injected back into System 1's context silently.

Why Not Just TTS?

Other meeting bots (like AgentCall) use traditional TTS which requires: STT (1s) → LLM reasoning (3-8s) → TTS (1s) = 5-12 second delay

CallingClaw's Realtime API is end-to-end: Audio in → Realtime API → Audio out = ~300ms

This makes natural conversation possible. In sales demos and business meetings, a 10-second pause kills the flow. 300ms feels like a real person.

Comparison with AgentCall

	CallingClaw	AgentCall
AI Brain	Built-in (Realtime API)	None (depends on host agent)
Voice Latency	~300ms	7-12 seconds
Screen Share	Interactive (Playwright)	Read-only HTML polling
Setup	Local Desktop app	Cloud API + API key
Best For	Sales, demos, business meetings	Coding agent voice control
License	MIT	MIT

Tips

• First meeting: Use talk-locally first to test voice without joining a real meeting
• Sales demos: Pre-inject customer context before the meeting starts
• Presentations: Use /api/screen/present/prepare to plan slide flow, then /api/screen/present/start for synchronized voice + visuals
• Multi-language: Set language in config — works for Chinese, English, Japanese, and more
• Voice provider: Default openai works best. Use grok for budget ($0.05/min) or gemini for cheapest (~$0.02/min). Note: gemini WebSocket may not work in China.