Playground

The Agent Playground provides an interactive chat interface to test your MCP servers and tools before deploying them to production. After configuring your MCP server, use the Playground to validate tool behavior, debug issues, and see exactly how AI models interact with your integrations.

Overview

The Playground simulates how AI agents interact with your MCP tools. You can:

• Select an MCP Server: Choose which connected app's tools to test
• Pick Active Tools: Enable specific tools for the session
• Choose an AI Model: Test with different models (GPT-4o, Claude, etc.)
• Chat Naturally: Ask the AI to perform tasks using your tools
• Monitor Execution: Watch tool calls, parameters, and results in real-time

Getting Started

Step 1: Access the Playground

Navigate to Integration Studio → Playground to open the testing environment.

Step 2: Select Your MCP Server

From the MCP Server dropdown, select the server you want to test:

Step 3: Configure Active Tools

Once you select an MCP server, you'll see all available tools in the left panel:

• Tools Tab: Browse and enable/disable specific tools
• Search: Filter tools by name or description
• Tool Cards: Each tool shows its name, description, and status

Click on tools to add them to your active session. Only enabled tools will be available to the AI model.

Step 4: Select an AI Model

At the bottom of the chat interface, select which AI model to use for testing:

Step 5: Start Chatting

Type your request in the chat input and send. The AI will:

1. Understand your request
2. Decide which tools to use
3. Execute the tools with appropriate parameters
4. Return the results in a conversational format

The Playground Interface

The Playground interface with Zendesk MCP server, active tools, and a completed search query

Interface Components

Execution Monitoring

One of the Playground's most powerful features is real-time execution monitoring. When the AI uses your tools, you can see every step of the process.

Execution Progress

When a tool executes, you'll see an expandable Execution Complete section showing:

• Step Count: Total number of execution steps (e.g., "24 steps")
• Status: Success, failure, or in-progress indicator
• Expandable Details: Click to see full execution log

What You Can Monitor

Example: Search Results

When searching for data (like Zendesk tickets), you'll see structured results:

I've found several Zendesk tickets related to login issues:

1. Ticket ID 179: "Account Login Issue"
   - Created: 2025-12-23T04:43:17Z
   - Status: Open
   - Priority: High
   - Description: The customer reports an issue with the account login...

2. Ticket ID 159: "Login Issues on Platform"
   - Created: 2025-12-23T04:32:16Z
   - Status: Open
   - Priority: Normal
   - Description: Thank you for reaching out...

Testing Workflows

Basic Tool Testing

Test individual tools with simple prompts:

"Search Zendesk tickets about 'login'"
"Create a new contact with name John Doe and email john@example.com"
"Get the status of order #12345"
"List all open support tickets"

Multi-Tool Operations

Test complex workflows that require multiple tools:

"Find the customer John Smith, then create a support ticket
for them about a billing issue"

The AI will:

1. Use the search tool to find John Smith
2. Use the create ticket tool with John's customer ID
3. Return confirmation of both operations

Edge Case Testing

Test how tools handle unusual inputs:

Error Handling

Test how tools respond to errors:

"Delete ticket with ID 99999999"  // Non-existent ID
"Create a duplicate contact"       // Constraint violation
"Access restricted data"           // Permission error

Chat vs Form Mode

The Playground offers two input modes:

Chat Mode

Natural language interface for conversational testing:

• Type requests as you would to an AI assistant
• Great for testing how AI interprets user intent
• Best for complex, multi-step operations

Form Mode

Structured input for precise tool testing:

• Select a specific tool from a dropdown
• Fill in each parameter explicitly
• Execute with exact values you specify
• Best for testing specific parameter combinations

Toggle between modes using the Chat / Form button in the input area.

Working with Tool Results

Understanding Response Format

Tool responses in the Playground show:

1. Conversational Summary: AI's natural language interpretation
2. Structured Data: Formatted results (lists, tables, etc.)
3. Raw Response: Expandable view of the actual API response
4. Execution Metadata: Timing, steps, and status

Copying Results

You can copy:

• The full AI response
• Individual data fields
• Raw JSON responses

Use copied data to verify results in the actual connected application.

Debugging Tools

Common Issues and Solutions

Using Execution Logs

Click on Execution Complete to expand detailed logs:

1. Identify the failing step: Look for red status indicators
2. Check parameters: Verify the AI passed correct values
3. Review errors: Read error messages for specific issues
4. Test fixes: Adjust tool schema and retry

Iterating on Tool Schemas

If the AI consistently misuses a tool:

1. Review the description: Is it clear what the tool does?
2. Check parameter names: Are they intuitive?
3. Add examples: Include example values in descriptions
4. Reduce ambiguity: Make tool purposes distinct

Best Practices

From Playground to Production

Once your tools work correctly in the Playground:

1. Verify All Test Cases Pass

Ensure you've tested:

• Normal operations (happy path)
• Edge cases and error handling
• Multi-tool workflows
• Different AI models

2. Review Tool Schemas

Based on Playground testing, refine:

• Tool descriptions for clarity
• Parameter descriptions and examples
• Required vs optional fields
• Default values

3. Deploy to Agents

Add your validated MCP server tools to production agents:

1. Go to Agent Builder
2. Add the MCP server as a tool source
3. Enable specific tools for the agent
4. Test the agent with real users

4. Monitor in Production

After deployment, monitor tool usage in the MCP Hub dashboard:

• Success/error rates
• Response times
• Most-used tools
• Common errors

Next Steps

Summary

The Agent Playground is your testing ground for MCP tools:

• Interactive Chat Interface: Test tools with natural language prompts
• MCP Server Selection: Choose which connected app's tools to test
• Tool Management: Enable/disable specific tools for focused testing
• Multi-Model Support: Test with GPT-4o, Claude, and other models
• Full Execution Visibility: Monitor every step of tool execution with detailed logs
• Chat & Form Modes: Natural language or structured parameter testing
• Debugging Tools: Identify and fix issues before production deployment

Use the Playground to validate your MCP tools work correctly, then confidently deploy them to your production AI agents.