LLM Playground

Test your MCP server’s behavior against various LLMs

Use the LLM playground to test your server against different LLMs. This is helpful to simulate how LLMs interpret your server and test for hallucinations.

The playground is also great for simulating agent behaviors. You can customize the system prompt, swap models, and configure system prompts and temperature as if you would building an agent.

Set up LLM Playground

You need to set up at least one LLM to use the playground. Go to the settings tab in the inspector and follow instructions from there.

OpenAI

Get an API key from OpenAI Platform.

gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-4, gpt-5.1, gpt-5.1-codex, gpt-5.1-codex-mini, gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-chat-latest, gpt-5-pro, gpt-5-codex, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-3.5-turbo,o3-mini, o3, o4-mini, o1

GPT-5 models require organization verification. If you encounter access errors, visit OpenAI Settings and verify your organization. Access may take up to 15 minutes after verification.

GPT-5 models do not support temperature configuration. The temperature setting will be automatically disabled when using GPT-5 models.

Claude (Anthropic)

Get an API key from Anthropic Console.

claude-opus-4-1, claude-opus-4-0, claude-sonnet-4-5, claude-sonnet-4-0, claude-3-7-sonnet-latest, claude-haiku-4-5, claude-3-5-haiku-latest

Gemini

Get an API key from Google AI Studio

gemini-3-pro-preview, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.0-flash-exp, gemini-1.5-pro, gemini-1.5-pro-002, gemini-1.5-flash, gemini-1.5-flash-002, gemini-1.5-flash-8b, gemini-1.5-flash-8b-001, gemma-3-2b, gemma-3-9b, gemma-3-27b, gemma-2-2b, gemma-2-9b, gemma-2-27b, codegemma-2b, codegemma-7b

Deepseek

Get an API key from Deepseek Platform

deepseek-chat, deepseek-reasoner

Mistral AI

Get an API key from Mistral AI Console

mistral-large-latest, mistral-small-latest, codestral-latest, ministral-8b-latest, ministral-3b-latest

OpenRouter

Get an API key from OpenRouter Console

Select from any tool-capable model in the dropdown.

Ollama

Make sure you have Ollama installed, and the MCPJam Ollama URL configuration is pointing to your Ollama instance.

Start an Ollama instance with ollama serve <model>. MCPJam will automatically detect any Ollama models running.

LiteLLM Proxy

Use LiteLLM Proxy to connect to 100+ LLMs through a unified OpenAI-compatible interface.

1. Start LiteLLM Proxy: Follow the LiteLLM Proxy Quick Start Guide to set up your proxy server
2. Configure in MCPJam: Go to Settings → LiteLLM card → Click “Configure”
3. Enter Connection Details:
   • Base URL: Your LiteLLM proxy URL (default: http://localhost:4000)
   • API Key: Your proxy API key (use the same key you use in your API requests)
   • Model Aliases: Comma-separated list of model names configured in your proxy (e.g., gpt-3.5-turbo, claude-3-opus, gemini-pro)

Use the exact model names that work with your LiteLLM proxy’s /v1/chat/completions endpoint. These are typically the model names without provider prefixes (e.g., gpt-3.5-turbo instead of openai/gpt-3.5-turbo).

Example configuration:

# In your LiteLLM config.yaml
model_list:
  - model_name: gpt-3.5-turbo
    litellm_params:
      model: openai/gpt-3.5-turbo
      api_key: os.environ/OPENAI_API_KEY

Then in MCPJam, use gpt-3.5-turbo as the model alias.

Choose an LLM model

Once you’ve configured your LLM API keys, go to the Playground tab. On the bottom near the text input, you should see a LLM model selector. Select the model from the ones you’ve configured

System prompt and temperature

You can configure the system prompt and temperature, just like you would building an agent. The temperature is defaulted to the default value of the LLM providers (Claude = 0, OpenAI = 1.0).

Higher temperature settings tend to hallucinate more with MCP interactions

Using MCP prompts in chat

You can use MCP prompts directly in the playground chat by typing / to trigger the prompts menu. When you select a prompt, it appears as an expandable card above the chat input showing:

• Server name - Which MCP server provides the prompt
• Description - What the prompt does
• Arguments - Required and optional parameters
• Preview - A preview of the prompt content

Click the card to expand/collapse details, or click the X button to remove it before sending. This helps you understand what will be sent to the LLM before submitting your message.

Playground layout

The playground features a split-panel layout with:

• Chat panel (left) - Your conversation with the LLM, including tool calls and results
• JSON-RPC logger (right) - Real-time view of MCP protocol messages between Inspector and your servers

You can resize the panels by dragging the divider between them. This layout helps you understand exactly how the LLM interacts with your MCP servers.

Error handling

When errors occur during playground interactions, you’ll see an error message with a “Reset chat” button. For detailed debugging, click “More details” to expand additional error information, including JSON-formatted error responses when available. This helps you quickly identify and resolve issues with your MCP server or LLM configuration.

Elicitation support

MCPJam has elicitation support in the LLM playground. Any elicitation requests will be shown as a popup modal.

MCP-UI support

The playground supports rendering custom UI components from MCP servers using the MCP-UI specification. When an MCP tool returns a UI resource, it will be rendered inline in the chat with interactive capabilities.

MCP-UI components can:

• Display rich, interactive visualizations
• Trigger tool calls through button actions
• Send follow-up messages to the chat
• Open external links
• Show notifications

This enables MCP server developers to create custom user experiences beyond plain text responses.

ChatGPT Apps and MCP Apps support

The playground supports rendering custom UI components from MCP tools using both the OpenAI Apps SDK (ChatGPT apps) and MCP Apps.

ChatGPT Apps: When a tool includes an openai/outputTemplate metadata field pointing to a resource URI, the playground will render the custom HTML interface in an isolated iframe with access to the window.openai API.

MCP Apps: When a tool includes a ui/resourceUri metadata field, the playground will render the custom UI component inline in the chat.

This enables MCP servers to provide rich, interactive visualizations for tool results, including charts, forms, and custom widgets that can call other tools or send followup messages to the chat.

Display modes

OpenAI Apps can request different display modes to optimize their presentation:

• Inline (default) - Widget renders within the chat message flow
• Picture-in-Picture - Widget floats at the top of the screen, staying visible while you scroll through the chat
• Fullscreen - Widget expands to fill the entire viewport for immersive experiences

Widgets can request display mode changes using window.openai.requestDisplayMode({ mode: 'pip' }) or window.openai.requestDisplayMode({ mode: 'fullscreen' }). Users can exit PiP or fullscreen modes by clicking the close button in the top-left corner of the widget.

Device and locale testing

The playground includes controls for testing widgets across different device types and locales:

• Device selector - Switch between mobile, tablet, and desktop viewports to test responsive layouts
• Locale selector - Choose from common BCP 47 locales (e.g., en-US, es-ES, ja-JP) to test internationalization
• Theme toggle - Switch between light and dark modes

These settings are automatically passed to widgets via the window.openai API, allowing them to adapt their UI accordingly.

Widget debugging

When viewing tool results with custom UI components, you can access debugging information using the icon buttons in the tool header:

• Data (database icon) - View tool input, output, and error details
• Widget State (box icon) - Inspect the current widget state and see when it was last updated (ChatGPT Apps only)
• CSP (shield icon) - View CSP violations and suggested fixes for both ChatGPT Apps and MCP Apps

Click any icon to toggle the corresponding debug panel. Click again to close it.

The Widget State tab only appears for ChatGPT Apps (OpenAI SDK). MCP Apps (SEP-1865) do not support persistent widget state.

Content Security Policy (CSP)

The UI Playground includes CSP enforcement controls to help you test widget security configurations. You can switch between two CSP modes using the shield icon in the toolbar:

• Permissive (default) - Allows all HTTPS resources, suitable for development and testing
• Widget-Declared - Only allows domains declared in the widget’s CSP metadata field

The CSP mode applies to both ChatGPT Apps (openai/widgetCSP) and MCP Apps (ui/csp per SEP-1865).

When a widget violates CSP rules in widget-declared mode, you’ll see a badge on the CSP debug tab showing the number of blocked requests. Click the CSP tab to view:

• Suggested fix - Copyable JSON snippet to add to your widget’s CSP metadata field
• Blocked requests - List of all CSP violations with directive and URI details
• Declared domains - The connect_domains and resource_domains your widget currently declares

This helps you identify which external resources your widget needs and configure proper CSP declarations before deploying to production environments.

CSP mode only applies in the UI Playground. The Chat tab always uses permissive mode to avoid disrupting normal testing workflows.

UI Playground device settings

The UI Playground provides controls to simulate different device environments for testing ChatGPT Apps and MCP Apps. These settings affect how widgets receive device context.

Protocol-aware controls

The playground automatically detects which app protocol is in use and shows appropriate controls:

• ChatGPT Apps - Tools with openai/outputTemplate metadata
• MCP Apps - Tools with ui/resourceUri metadata

Some controls (like timezone) are specific to MCP Apps and only appear when an MCP Apps widget is active.

Device type

Select between mobile, tablet, or desktop device types. This affects the device context that widgets receive:

• ChatGPT Apps: window.openai.userAgent.device.type
• MCP Apps: hostContext.platform (derived as “mobile”, “web”, or “desktop”)

The device type selector is located in the playground toolbar.

Locale

Choose from common BCP 47 locales (e.g., en-US, ja-JP, es-ES) to test internationalization. Available for both ChatGPT Apps and MCP Apps:

• ChatGPT Apps: window.openai.locale
• MCP Apps: hostContext.locale

Timezone (MCP Apps only)

Select an IANA timezone identifier (e.g., America/New_York, Asia/Tokyo) to test timezone-aware widgets. This control only appears when testing MCP Apps:

• MCP Apps: hostContext.timeZone

The timezone selector includes common zones with UTC offset information for easy reference.

Device capabilities

Toggle hover and touch capabilities to simulate different input methods:

• Hover - Indicates whether the device supports hover interactions (typically enabled for desktop, disabled for mobile)
• Touch - Indicates whether the device supports touch input (typically enabled for mobile/tablet, disabled for desktop)

These settings are reflected in:

• ChatGPT Apps: window.openai.userAgent.capabilities.hover and window.openai.userAgent.capabilities.touch
• MCP Apps: hostContext.deviceCapabilities.hover and hostContext.deviceCapabilities.touch

Safe area insets

Configure safe area insets to simulate device notches, rounded corners, and gesture areas. Click the safe area button in the toolbar to open the editor, which provides:

• Visual preview - Shows how insets affect the content area
• Preset configurations - Quick access to common device profiles:
  • None - No insets (0px on all sides)
  • Notch - iPhone with notch (44px top, 34px bottom)
  • Island - iPhone with Dynamic Island (59px top, 34px bottom)
  • Android - Android gesture navigation (24px top, 16px bottom)
• Custom values - Manually adjust top, bottom, left, and right insets in pixels

Widgets receive these values through:

• ChatGPT Apps: window.openai.safeArea.insets
• MCP Apps: hostContext.safeAreaInsets

Fullscreen navigation

When a widget is in fullscreen mode, a navigation header appears at the top with:

• Back/Forward buttons - Navigate through the widget’s browsing history (enabled when navigation is available)
• Widget title - Displays the tool name
• Close button - Exit fullscreen mode

The navigation buttons mirror the widget’s internal history state, allowing you to navigate multi-page widgets without leaving fullscreen mode.