testmcpy 0.1.3

A comprehensive testing framework for validating LLM tool calling capabilities with MCP services

testmcpy - MCP Testing Framework

A comprehensive testing framework for validating LLM tool calling capabilities with MCP (Model Context Protocol) services, specifically designed for testing Superset operations.

Quick Start

Installation

From source (development):

git clone https://github.com/preset-io/testmcpy.git
cd testmcpy
pip install -e .

From PyPI (once published):

pip install testmcpy

Via Homebrew (once published to PyPI):

brew tap preset-io/testmcpy
brew install testmcpy

See INSTALLATION.md for detailed installation instructions and distribution options.

Quick Usage

# First-time setup: Create user config file
testmcpy setup

# View current configuration
testmcpy config-cmd

# List MCP tools
testmcpy tools
testmcpy tools --detail --filter chart

# Research LLM capabilities
testmcpy research --model claude-sonnet-4.5-20250929 --provider anthropic

# Run test suites
testmcpy run tests/ --model claude-3-5-haiku-20241022 --provider anthropic

# Interactive chat
testmcpy chat --provider anthropic --model claude-sonnet-4.5-20250929

# Compare test results
testmcpy report reports/model1.yaml reports/model2.yaml

# Initialize new project
testmcpy init my_project

Framework Structure

mcp_testing/
├── research/               # Research scripts for testing LLM capabilities
│   └── test_ollama_tools.py
├── src/                    # Core framework modules
│   ├── mcp_client.py      # MCP protocol client
│   ├── llm_integration.py # LLM provider abstraction
│   └── test_runner.py     # Test execution engine
├── evals/                  # Evaluation functions
│   └── base_evaluators.py # Standard evaluators
├── tests/                  # Test cases (YAML/JSON)
│   ├── basic_test.yaml
│   └── example_mcp_tests.yaml
├── reports/                # Test reports and comparisons
└── cli.py                  # CLI interface

Writing Test Cases

Test cases are defined in YAML files:

version: "1.0"
name: "My Test Suite"

tests:
  - name: "test_chart_creation"
    prompt: "Create a bar chart showing sales by region"
    expected_tools:
      - "create_chart"
    evaluators:
      - name: "was_mcp_tool_called"
        args:
          tool_name: "create_chart"
      - name: "execution_successful"
      - name: "final_answer_contains"
        args:
          expected_content: ["chart", "created"]
      - name: "within_time_limit"
        args:
          max_seconds: 30

Available Evaluators

Generic Evaluators

• was_mcp_tool_called - Verify MCP tool was called
• execution_successful - Check for successful execution
• final_answer_contains - Validate response content
• answer_contains_link - Check for links in response
• within_time_limit - Verify performance
• token_usage_reasonable - Check token/cost efficiency

Superset-Specific Evaluators

• was_superset_chart_created - Verify chart creation
• sql_query_valid - Validate SQL syntax

Supported LLM Providers

Anthropic (Recommended) ✅

The Anthropic API (anthropic) provider is recommended for most users:

# Add to ~/.testmcpy
ANTHROPIC_API_KEY=sk-ant-your-key-here
DEFAULT_PROVIDER=anthropic
DEFAULT_MODEL=claude-3-5-haiku-20241022

Available Models:

• claude-sonnet-4.5-20250929 - Newest, most capable
• claude-3-5-haiku-20241022 - Fast, cost-effective (recommended)
• claude-3-5-sonnet-20241022 - Balanced performance
• All Claude models via API

Features:

• ✅ Full support for HTTP-based MCP services (like Superset MCP)
• ✅ Best tool calling accuracy
• ✅ Production-ready
• ✅ Simple API key setup

Get an API key: https://console.anthropic.com/

Ollama (Local, Free)

For local development without API costs:

# 1. Install Ollama
# macOS
brew install ollama

# Linux
curl -fsSL https://ollama.com/install.sh | sh

# 2. Start Ollama service
ollama serve

# 3. Pull a model with tool calling support
ollama pull llama3.1:8b

# 4. Configure testmcpy
# Add to ~/.testmcpy:
OLLAMA_BASE_URL=http://localhost:11434
DEFAULT_PROVIDER=ollama
DEFAULT_MODEL=llama3.1:8b

Recommended Models:

• llama3.1:8b - Best tool calling support
• mistral-nemo - Good alternative
• qwen2.5:7b - Fast, smaller model

Note: Requires Ollama running locally. Not recommended for production testing (less reliable tool calling than Claude).

OpenAI

# Add to ~/.testmcpy
OPENAI_API_KEY=sk-your-key-here
DEFAULT_PROVIDER=openai
DEFAULT_MODEL=gpt-4-turbo

Other Providers

• Claude Agent SDK (claude-sdk) - ⚠️ Only for stdio-based MCP servers (not HTTP)
• Local (local) - Transformers-based local models
• Claude CLI (claude-cli) - Uses Claude Code binary

Configuration

testmcpy uses a multi-layer configuration system with clear priority ordering:

Priority Order (highest to lowest):

1. Command-line options
2. .env file in current directory
3. ~/.testmcpy user configuration file
4. Environment variables
5. Built-in defaults

First-Time Setup

Create your user configuration file with helpful comments:

testmcpy setup

This creates ~/.testmcpy with examples for all configuration options. Edit the file to add your API keys and preferences.

View Current Configuration

testmcpy config-cmd

This displays all configuration values with their sources and checks which config files exist.

Configuration Files

User Config: ~/.testmcpy

Create with testmcpy setup, or manually create ~/.testmcpy to set your personal defaults:

# MCP Service Configuration
MCP_URL=http://localhost:5008/mcp/

# Option 1: Static Bearer Token
MCP_AUTH_TOKEN=your_token_here

# Option 2: Dynamic JWT Token (for Preset/Superset)
# MCP_AUTH_API_URL=https://api.app.preset.io/v1/auth/
# MCP_AUTH_API_TOKEN=your_preset_api_token
# MCP_AUTH_API_SECRET=your_preset_api_secret

# Default LLM Settings
DEFAULT_MODEL=claude-3-5-haiku-20241022
DEFAULT_PROVIDER=anthropic

# API Keys
ANTHROPIC_API_KEY=sk-ant-...
# OPENAI_API_KEY=sk-...

See .testmcpy.example for a complete example with detailed comments.

Project Config: .env

Create .env in your project directory to override user defaults:

# Project-specific settings
MCP_URL=https://my-project.mcp.example.com/mcp/
MCP_AUTH_TOKEN=project_specific_token
DEFAULT_MODEL=claude-sonnet-4.5-20250929

Authentication Options

testmcpy supports two methods for MCP authentication:

1. Static Bearer Token (simplest):

MCP_AUTH_TOKEN=your_bearer_token

2. Dynamic JWT Generation (for Preset/Superset):

Instead of manually managing JWT tokens, configure API credentials and testmcpy will automatically fetch and cache JWT tokens:

MCP_AUTH_API_URL=https://api.app.preset.io/v1/auth/
MCP_AUTH_API_TOKEN=your_api_token
MCP_AUTH_API_SECRET=your_api_secret

When configured, testmcpy will:

• Call the auth API with your credentials
• Extract the JWT access token from the response
• Cache the token for 50 minutes (tokens typically expire in 1 hour)
• Automatically refresh when needed

Note: Static MCP_AUTH_TOKEN takes priority. If both are configured, the static token is used.

Environment Variables

All configuration keys can also be set via environment variables:

# For Claude providers
export ANTHROPIC_API_KEY="sk-ant-..."

# For OpenAI provider
export OPENAI_API_KEY="sk-..."

# MCP service
export MCP_URL="http://localhost:5008/mcp/"
export MCP_AUTH_TOKEN="your_token"

# Or use dynamic token generation
export MCP_AUTH_API_URL="https://api.app.preset.io/v1/auth/"
export MCP_AUTH_API_TOKEN="your_api_token"
export MCP_AUTH_API_SECRET="your_api_secret"

# Default LLM settings
export DEFAULT_MODEL="claude-sonnet-4.5-20250929"
export DEFAULT_PROVIDER="anthropic"

Development Status

Phase 0: Research & Prototype ✅

• Research local LLM options with tool calling
• Build minimal Python script for LLM+MCP integration
• Validate tool calling with selected LLM
• Create basic framework structure

Phase 1: Foundation (In Progress)

• CLI framework with typer + rich
• Basic test execution engine
• MCP protocol client
• LLM provider abstraction
• Core evaluation functions
• Integration with existing Superset tests

Phase 2: Core Features (Planned)

• Multi-model comparison support
• Advanced reporting with charts
• Test suite versioning
• Parallel test execution

Phase 3: Advanced Capabilities (Future)

• CI/CD integration
• Interactive test development mode
• Performance profiling
• Cost optimization insights

Known Limitations

• Claude SDK Provider: Only supports stdio-based MCP servers (command-line tools)
  • Not compatible with HTTP-based MCP services (like Superset MCP)
  • Use anthropic provider for HTTP MCP services
• HTTP MCP Services: Use anthropic provider (fully supported)
• Ollama models: Require specific formatting for reliable tool calling
• CPU-only execution: May be slow for larger local models
• Tool calling accuracy: Varies by model (Claude models generally most reliable)
• Cost: Claude API providers (anthropic) incur API costs; consider using Ollama for development

Contributing

This framework follows the patterns established by promptimize and superset-sup. When contributing:

1. Use modern Python practices (type hints, async/await)
2. Follow the existing code style
3. Add tests for new evaluators
4. Document new features in this README

License

Same as the parent promptimize project.