An AI-powered agent that automatically summarizes sales calls using Vercel's Sandbox architecture. The agent analyzes call transcripts and generates structured summaries with objections, action items, and insights.

Template Note: This template uses Gong as a starting example for call transcript integration. You can adapt it to work with other call recording platforms (Zoom, Google Meet, etc.) by modifying the webhook handler and transcript fetching logic.

Extensible: This template can be extended to integrate with Salesforce (or another CRM of your choice), Slack (to post call summaries), and other services. The demo files in /demo-files/context/ demonstrate how CRM data and other context can be provided to the agent.

• Structured Summaries - AI-generated summaries with tasks, objections, and key insights
• Sandbox Agent - Uses Vercel Sandbox for secure code execution and file exploration
• bash-tool - Generic bash tool for AI agents, compatible with AI SDK
• Demo Mode - Works out of the box with mock data (no Gong credentials needed)
• Objection Tracking - Identifies and scores how well objections were handled
• Durable Workflows - Built with Vercel Workflow DevKit for reliability

You can deploy to Vercel and try it immediately with demo data with one click:

Demo mode is enabled by default - no Gong credentials required to test!

git clone https://github.com/vercel-labs/call-summary-agent
cd call-summary-agent
pnpm install

vercel link
vercel env pull

pnpm dev

# Trigger the agent with demo data
curl -X POST http://localhost:3000/api/gong-webhook \
  -H "Content-Type: application/json" \
  -d '{}'

To use real Gong API data instead of demo data:

Add these to your Vercel project settings or .env.local:

GONG_ACCESS_KEY=your_gong_access_key
GONG_SECRET_KEY=your_gong_secret_key

Note: Demo mode is automatically enabled when Gong credentials are missing.

1. Go to your Gong settings > Integrations > Webhooks
2. Create a new webhook with:
   • URL: https://your-app.vercel.app/api/gong-webhook
   • Events: Select "Call completed"
3. Save and test the webhook

flowchart TD
    subgraph entry [Entry Point]
        Webhook[POST /api/gong-webhook]
    end

    subgraph workflow [Workflow Layer]
        WF["workflowGongSummary<br/>(use workflow)"]
        Steps["Workflow Steps<br/>(use step)"]
    end

    subgraph agent [Agent Layer]
        Agent[ToolLoopAgent]
        Sandbox[Vercel Sandbox]
    end

    subgraph tools [Sandbox Tools]
        Bash[bash-tool]
    end

    subgraph external [External Services]
        Gong[Gong API]
    end

    Webhook --> WF
    WF --> Steps
    Steps --> Agent
    Agent --> Sandbox
    Agent --> tools
    Steps --> Gong

Demo mode uses realistic mock data including a sample 20-minute product demo call. The demo files are organized in the /demo-files folder:

demo-files/
├── webhook-data.json              # Mock Gong webhook payload
├── transcript.json                # 20-minute call transcript
└── context/
    ├── gong-calls/previous/       # Historical calls
    │   ├── demo-call-000-discovery-call.md
    │   └── demo-call-intro-initial-call.md
    ├── research/                  # Background info
    │   ├── company-research.md
    │   └── competitive-intel.md
    └── playbooks/
        └── sales-playbook.md

These files are loaded into the sandbox for the agent to explore.

Override the default system prompt:

AGENT_SYSTEM_PROMPT="You are a sales call analyst..."

1. Webhook Received: Gong sends call data when a call completes (or mock data in demo mode)
2. Workflow Started: The durable workflow begins processing
3. Transcript Fetched: Call transcript is retrieved from Gong API (or mock data)
4. Sandbox Created: A secure sandbox is created with call files
5. Agent Runs: The AI agent explores transcripts using bash-tool
6. Summary Generated: Structured output with tasks, objections, insights

Each step uses the "use step" directive for:

• Automatic retries on failure
• State persistence
• Observability in Vercel dashboard

The agent uses bash-tool for exploring call transcripts via shell commands:

# List call files
ls gong-calls/

# Search for pricing discussions
grep -r "pricing" gong-calls/

# View call metadata
cat gong-calls/metadata.json

# Find objections
grep -i "concern\|issue\|problem" gong-calls/*.md

All bash commands are logged for observability:

[bash-tool] INFO: Bash command starting { command: 'grep -r "pricing" gong-calls/' }
[bash-tool] INFO: Bash command completed { command: '...', exitCode: 0, stdoutLength: 605 }

The agent generates structured output:

{
  summary: string,           // Comprehensive call summary
  tasks: [{
    taskDescription: string,
    taskOwner: string,
    ownerCompany: 'internal' | 'customer' | 'partner'
  }],
  objections: [{
    description: string,
    quote: string,
    speaker: string,
    speakerCompany: string,
    handled: boolean,
    handledAnswer: string,
    handledScore: number,    // 0-100
    handledBy: string
  }]
}

Add playbook detection by configuring config.playbooks in lib/config.ts.

# Install dependencies
pnpm install

# Run development server
pnpm dev

# Build for production
pnpm build

sales-call-summary-agent/
├── app/
│   ├── api/gong-webhook/    # Webhook endpoint
│   ├── layout.tsx
│   └── page.tsx             # Status page
├── demo-files/              # Demo mode files
│   ├── webhook-data.json
│   ├── transcript.json
│   └── context/             # Additional context files
├── lib/
│   ├── agent.ts             # ToolLoopAgent configuration
│   ├── config.ts            # Centralized configuration
│   ├── gong-client.ts       # Gong API helpers
│   ├── mock-data.ts         # Demo mode loader
│   ├── sandbox-context.ts   # File generation for sandbox
│   ├── tools.ts             # Agent tools (bash-tool)
│   ├── types.ts             # TypeScript types
│   └── logger.ts            # Logging utility
└── workflows/
    └── gong-summary/
        ├── index.ts         # Main workflow
        └── steps.ts         # Workflow steps

Receives Gong webhook payloads and triggers the summary workflow.

Request Body: Gong webhook payload or empty {} in demo mode.

Response:

{
  "message": "Workflow triggered",
  "callId": "1234567890"
}

The function returns a string output which can be configured into an outputSchema of your choice.

Contributions are welcome! Please open an issue or submit a pull request.