Call.md — Turn Meetings into Live Agent Loops with Real-Time AI Intelligence

By Prahlad Menon 4 min read

What if your meetings could think alongside you?

Not just transcription after the fact — but live intelligence during the call. Questions to ask. Things to say. Automatic tool lookups when someone mentions something you need to research. Real-time coaching nudges when you’re monologuing too long.

Call.md makes this real. It’s an open-source Electron app that turns meetings into live agent loops.

GitHub: video-db/call.md

The Problem with Meeting AI

Most meeting tools follow the same pattern: record → transcribe → summarize → forget. You get a summary email hours later that you’ll never read. The actual intelligence arrives too late to matter.

Call.md flips this. The AI works during the meeting — when you can actually use it.

How It Works

Before the call:

  • AI generates probing questions based on your meeting description
  • Dynamic checklist created from context
  • Google Calendar sync pulls your upcoming meetings

During the call:

  • Dual-channel transcription (you vs. them) — separate tracks for mic and system audio
  • Live Assist generates contextual suggestions in real-time
  • Conversation metrics: talk ratio, speaking pace, questions asked, monologue detection
  • Coaching nudges when the conversation needs steering
  • MCP auto-triggering — detects when you need information and calls tools automatically

After the call:

  • Three parallel summaries: narrative overview, key points by topic, action items
  • Webhook export to n8n, Zapier, or CRMs
  • Full markdown export with transcript and metrics

The MCP Integration is the Killer Feature

This is where Call.md diverges from every other meeting tool.

During your call, an MCP (Model Context Protocol) agent listens to the conversation. When it detects an information need — “What’s our pricing for enterprise?” or “When did we last ship to that client?” — it automatically triggers your connected tools.

Results appear inline in a panel during the meeting. No alt-tabbing. No “let me look that up.” The information just appears.

You connect MCP servers in Settings (stdio for local, HTTP for remote). The agent handles the rest.

Architecture

The tech stack is modern and well-structured:

  • Electron 34 for desktop
  • React 19 + TypeScript 5.8 for the UI
  • tRPC 11 for type-safe IPC between main and renderer
  • Drizzle ORM + SQLite for local-first storage
  • VideoDB SDK for transcription (requires API key)
  • MCP SDK 1.0 for tool integrations

Everything runs locally. Your data stays on your machine. The SQLite database and logs live in ~/Library/Application Support/call-md/.

The Copilot Services

The intelligence layer is organized into focused services:

ServicePurpose
context-managerTracks conversation state
conversation-metricsTalk ratio, WPM, questions
nudge-engineRate-limited coaching alerts
sales-copilotCore orchestrator
summary-generatorPost-meeting extraction
transcript-bufferManages live transcripts

The MCP layer has its own stack: connection orchestrator, intent detector, tool aggregator, result handler. It’s built to scale to many connected tools.

Quick Start

# macOS (Apple Silicon & Intel)
curl -fsSL https://artifacts.videodb.io/call.md/install | bash

Launch the app, grant microphone + screen recording permissions, enter your free VideoDB API key from console.videodb.io, and start recording.

Who This Is For

  • Sales teams who want real-time battlecard lookups and objection handling
  • Customer success tracking conversation health metrics
  • Founders who take a lot of external calls and want automated CRM updates
  • Anyone who wants their tools to work with them during conversations, not after

The Bigger Picture

Call.md represents where meeting AI is heading: from passive recording to active participation. The meeting itself becomes an agent runtime — your tools, your knowledge, your workflows, all accessible through natural conversation.

And it’s open source. Fork it. Add your own intelligence services. Connect your own MCP servers. The architecture is clean enough to extend.

Links: