Test and benchmark LLMs with MCP tools in minutes.

A testing framework for validating how LLMs call tools via Model Context Protocol (MCP) — compare Claude, GPT-4, Llama, and other models' accuracy, cost, and performance.

---

Documentation | Getting Started | CLI Reference | Examples | Contributing | Discussions

---

Why testmcpy?

• Validate tool calling: Ensure LLMs call the right tools with correct parameters
• Compare models: Find the best price/performance balance for your use case
• Prevent regressions: Catch breaking changes in your MCP service with CI/CD
• Optimize costs: Track token usage and identify the most cost-effective models

How it compares

	testmcpy	MCP Inspector	MCPJam	promptfoo
Automated LLM-driven evals of MCP servers	✅ YAML suites, 40+ evaluators	❌ manual testing	✅	⚠️ generic LLM eval with an MCP provider
Multi-provider (Claude / GPT / Gemini / Ollama / Bedrock…)	✅ 11 providers incl. agent SDKs	n/a	✅	✅
CI gate with exit codes + JUnit	✅ --gate, --junit-xml	❌	✅	✅
Cost & token tracking per test/model	✅	❌	⚠️	⚠️
Multi-turn, mutation & metamorphic testing	✅	❌	❌	⚠️
Auth testing (JWT/OAuth/mTLS) + debugger	✅ 7 auth types	⚠️ OAuth only	✅ OAuth debugger	❌
Python-native (pip/uvx, pytest-friendly)	✅	❌ npm	❌ npm	❌ npm

Use MCP Inspector for quick manual poking; reach for testmcpy when you want repeatable, scored, CI-gated evaluation of how real models use your server.

Quick Start

# Install testmcpy
pip install testmcpy

# Run interactive setup
testmcpy setup

# Start testing
testmcpy chat                     # Interactive chat with MCP tools
testmcpy research                 # Test LLM tool-calling capabilities
testmcpy run tests/              # Run your test suite

That's it! No complex configuration needed to get started.

Key Features

Multi-Provider LLM Support

Test with Claude, GPT, Gemini, Llama, and other models. Works with both paid APIs and free local models via Ollama. Includes agent-SDK providers (Claude, Codex, Gemini) with native MCP support.

Provider	Config name	Models	Features
Anthropic	anthropic	claude-opus-4, claude-sonnet-4-5, claude-haiku-4-5	Native MCP, extended thinking, vision, token caching
OpenAI	openai	gpt-4, gpt-4-turbo, gpt-4o	Function calling, vision, cost tracking
Ollama	ollama	Llama, Mistral, etc. (local)	Free, local execution, no API costs
Claude SDK	claude-sdk (aliases: claude-cli, claude-code)	claude-sonnet-4-5, claude-opus-4	Claude Agent SDK, native MCP, CLI OAuth login
Codex SDK	codex-sdk (aliases: codex-cli, codex)	gpt-5-codex, o3, o4-mini	openai-agents SDK, native MCP, Codex CLI OAuth or API key
Gemini SDK	gemini-sdk	gemini-sdk-flash, gemini-sdk-pro	google-adk, native MCP
Google Gemini	gemini (alias: google)	gemini-2.5-flash, gemini-2.5-pro	Direct Gemini API, function calling
Gemini CLI	gemini-cli	gemini-2.5-flash, gemini-2.5-pro	Subprocess-based Gemini CLI
AWS Bedrock	bedrock (alias: aws-bedrock)	Claude models via AWS	IAM auth, no Anthropic key needed
xAI	xai (alias: grok)	grok models	Function calling
OpenRouter	openrouter	100+ models with one API key	Function calling, cost tracking

Built-in Evaluators

Comprehensive validation out of the box. Each evaluator returns a score from 0.0 to 1.0 with pass/fail status and detailed reasoning.

Tool Calling:

• was_mcp_tool_called — Verify specific tool was invoked (supports prefix/gateway matching)
• tool_call_count — Validate number of tool calls
• tool_called_with_parameter — Check specific parameter was passed (fuzzy matching)
• tool_called_with_parameters — Validate multiple parameters at once
• parameter_value_in_range — Ensure numeric parameters are within bounds

Execution & Performance:

• execution_successful — Check for errors or failures in tool results
• within_time_limit — Performance validation against max_seconds
• final_answer_contains — Validate response content
• token_usage_reasonable — Cost efficiency validation
• response_time_acceptable — Latency threshold checking
• auth_successful — Authentication flow validation

Extensible: Extend BaseEvaluator and implement evaluate(context) -> EvalResult to create custom evaluators for your domain.

YAML Test Definitions

Define test suites as code for repeatable, version-controlled testing:

version: "1.0"
name: "Chart Operations Test Suite"

config:
  timeout: 30
  model: "claude-sonnet-4-5"
  provider: "anthropic"

tests:
  - name: "test_create_chart"
    prompt: "Create a bar chart showing sales by region"
    evaluators:
      - name: "was_mcp_tool_called"
        args:
          tool_name: "create_chart"
      - name: "execution_successful"

  # Multi-turn test
  - name: "test_multi_turn"
    steps:
      - prompt: "List all dashboards"
        evaluators:
          - name: "was_mcp_tool_called"
            args:
              tool_name: "list_dashboards"
      - prompt: "Show me the first one"
        evaluators:
          - name: "final_answer_contains"
            args:
              content: "dashboard"

  # Load testing
  - name: "test_load"
    prompt: "List dashboards"
    load_test:
      concurrent: 5
      duration: 60

CLI & Web UI

• Rich terminal UI: Progress bars, colored output, formatted tables
• Optional web interface: Visual tool explorer, interactive chat, analytics dashboards
• Real-time feedback: Watch tests execute with live updates via WebSocket

Architecture

testmcpy connects your LLM provider to your MCP service and validates the interactions:

graph TB
    subgraph UI["User Interface Layer"]
        CLI["CLI Commands<br>(Typer)"]
        WebUI["Web UI<br>(React + Vite + Tailwind)"]
        TUI["Terminal Dashboard<br>(Textual)"]
    end

    subgraph Core["Core Framework"]
        Runner["Test Runner"]
        LLM["LLM Integration"]
        Evals["Evaluators"]
    end

    subgraph MCP_Layer["MCP Integration Layer"]
        Client["MCP Client<br>(FastMCP)"]
        Auth["Auth Manager"]
        Discovery["Tool Discovery"]
    end

    subgraph External["External Services"]
        LLM_APIs["LLM APIs<br>(Anthropic, OpenAI, Ollama)"]
        MCP_Services["MCP Services<br>(HTTP/SSE)"]
        Storage["Storage<br>(SQLite + JSON)"]
    end

    UI --> Core
    Core --> MCP_Layer
    MCP_Layer --> External
    Core --> External

Loading

How it works:

1. Define test cases in YAML with prompts and expected behavior
2. testmcpy sends prompts to your chosen LLM (Claude, GPT-4, Llama, etc.)
3. LLM calls tools via MCP protocol to your service
4. Evaluators validate tool selection, parameters, execution, and performance
5. Get detailed pass/fail results with metrics and cost analysis

Installation

# Install base package
pip install testmcpy

# With web UI support
pip install 'testmcpy[server]'

# All optional features
pip install 'testmcpy[all]'

Requirements: Python 3.10-3.12

Getting Started

1. Configuration

Run the interactive setup wizard:

testmcpy setup

This creates two config files:

.llm_providers.yaml — LLM configuration:

default: prod

profiles:
  prod:
    name: "Production"
    providers:
      - name: "Claude Sonnet"
        provider: "anthropic"
        model: "claude-sonnet-4-5"
        api_key: "your-anthropic-api-key"
        timeout: 60
        default: true

.mcp_services.yaml — MCP server profiles:

default: prod

profiles:
  prod:
    name: "Production"
    mcps:
      - name: "My MCP Service"
        mcp_url: "https://your-service.example.com/mcp"
        auth:
          auth_type: "jwt"  # or "bearer", "oauth", "none"
          api_url: "https://auth.example.com/v1/auth/"
          api_token: "your-api-token"
          api_secret: "your-api-secret"
        timeout: 30
        rate_limit_rpm: 60
        default: true

Configuration priority: CLI options > Profile files > .env > User config (~/.testmcpy) > Environment variables > Built-in defaults

The setup command is idempotent — safe to run multiple times. Use --force to overwrite existing files.

TESTMCPY_CHAT_OAUTH_LOGIN (default true): when a chat message hits an OAuth (oauth_auto_discover) MCP profile with no cached token, the server opens the interactive browser OAuth flow and retries. This assumes a browser is available on the machine running the server — in headless deployments set TESTMCPY_CHAT_OAUTH_LOGIN=false so the request fails fast with a clear error instead of blocking on a login that can never complete.

2. Explore Your MCP Service

# List available MCP tools
testmcpy tools

# Interactive chat to explore your tools
testmcpy chat

# Run automated research on tool-calling capabilities
testmcpy research --model claude-haiku-4-5

3. Create and Run Test Suites

# tests/my_tests.yaml
version: "1.0"
name: "My MCP Service Tests"

tests:
  - name: "test_tool_selection"
    prompt: "Create a bar chart showing sales by region"
    evaluators:
      - name: "was_mcp_tool_called"
        args:
          tool_name: "create_chart"
      - name: "execution_successful"
      - name: "within_time_limit"
        args:
          max_seconds: 30

testmcpy run tests/ --model claude-haiku-4-5

Commands Reference

The highlights are below — the full reference for all 38 commands lives at preset-io.github.io/testmcpy/cli.

Command	Description
Setup
testmcpy setup	Interactive configuration wizard
testmcpy doctor	Diagnose installation issues
Discovery
testmcpy tools	List available MCP tools
testmcpy profiles	List MCP profiles (table)
testmcpy status	Show MCP connection status
testmcpy explore-cli	Browse tools (non-interactive)
Testing
testmcpy run <path>	Execute test suite
testmcpy research	Test LLM tool-calling capabilities
testmcpy chat	Interactive chat with MCP tools
testmcpy compare	Multi-model comparison
Quality & Benchmarking
testmcpy bench	Run a suite across models × profiles × repeats
testmcpy conformance	Run the official MCP spec conformance suite
testmcpy score	Grade tool surface for LLM usability (0-100, A-F)
testmcpy scan	Static security scan of tool metadata (SARIF output)
testmcpy matrix / leaderboard / flaky	Per-test × per-config analytics
Advanced
testmcpy baseline-save	Save current test results as a named baseline
testmcpy baseline-compare	Compare a run against a saved baseline
testmcpy baseline-list	List saved baselines
testmcpy mutate	Prompt mutation testing
testmcpy metamorphic	Metamorphic testing
testmcpy generate	AI-assisted test generation
testmcpy smoke-test	Quick smoke test against an MCP service
testmcpy coverage	Tool coverage report for a test suite
testmcpy multi-env	Run the same suite against multiple MCP profiles
testmcpy export-db	Export the SQLite results database
UI
testmcpy serve	Start web UI server (default port 8000)
testmcpy config-cmd	View current configuration
testmcpy config-mcp	Print MCP client snippets for Claude Desktop / Code

Common options: --profile, --llm-profile, --model, --provider, --timeout, --verbose, --output

Inline MCP Auth (No Config File Needed)

Pass MCP auth credentials directly on the command line, bypassing .mcp_services.yaml:

# JWT auth (e.g., Preset workspaces)
testmcpy run tests/ \
  --mcp-url https://workspace.example.com/mcp \
  --auth-type jwt \
  --jwt-url https://auth.example.com/v1/auth/ \
  --jwt-token $MCP_JWT_TOKEN \
  --jwt-secret $MCP_JWT_SECRET

# Bearer token auth
testmcpy run tests/ \
  --mcp-url https://workspace.example.com/mcp \
  --auth-type bearer \
  --auth-token $MCP_BEARER_TOKEN

# No auth (public MCP endpoint)
testmcpy run tests/ \
  --mcp-url https://workspace.example.com/mcp \
  --auth-type none

Environment variables are also supported: MCP_AUTH_TOKEN, MCP_JWT_URL, MCP_JWT_TOKEN, MCP_JWT_SECRET.

Web Interface

Optional React-based UI for visual testing and analytics — every page is documented at preset-io.github.io/testmcpy/web-ui:

# Install with UI support
pip install 'testmcpy[server]'

# Start server
testmcpy serve

The UI accepts loopback Host headers by default. For LAN, container, or reverse-proxy access, bind on all interfaces and explicitly list every hostname or IP clients will use (the option is repeatable):

testmcpy serve --host 0.0.0.0 \
  --allowed-host testmcpy.example.com \
  --allowed-host 192.0.2.10 \
  --no-browser

TESTMCPY_ALLOWED_HOSTS=testmcpy.example.com,192.0.2.10 provides the same host policy for deployments configured through environment variables. Values are hostnames or IP addresses only, without a URL scheme or port. A global * is rejected by testmcpy serve because it would disable DNS-rebinding protection.

Route	Page	Description
/	MCP Explorer	Tool discovery, smoke tests, schema viewing
/tests	Test Manager	YAML test browser, execution, results
/reports	Reports	All test results, evaluations, cost analysis
/chat	Chat Interface	Multi-turn conversation with MCP tools
/performance	Performance	Per-test matrix and config leaderboard (also serves /metrics, /compare)
/servers	Servers	Health monitoring + cross-server schema compatibility (also serves /mcp-health, /compatibility)
/security	Security Dashboard	Security evaluator results and risk summary
/generation-history	Generation History	AI test generation logs
/auth-debugger	Auth Debugger	Auth flow debugging
/config	Configuration	Settings and environment
/mcp-profiles	MCP Profiles	MCP server configuration
/llm-profiles	LLM Profiles	LLM provider configuration

Access at http://localhost:8000.

More screenshots

Generation History — AI-assisted test generation runs	Auth Debugger — step through OAuth / JWT / Bearer flows
Performance — per-test results across model and MCP configurations	Leaderboard — configs ranked by pass rate, cost-per-pass, latency
Security Dashboard — security evaluator results and risk summary	Schema Compat — cross-server tool schema compatibility matrix
Servers — MCP server health monitoring	MCP Profiles — manage MCP service connections
LLM Profiles — provider configurations with model pricing	Configuration — current settings and client snippets

LLM Providers

Anthropic (Recommended)

Best tool-calling accuracy, native MCP support:

# .llm_providers.yaml
prod:
  name: "Production"
  providers:
    - name: "Claude Sonnet"
      provider: "anthropic"
      model: "claude-sonnet-4-5"
      api_key_env: "ANTHROPIC_API_KEY"
      default: true

Ollama (Free, Local)

Perfect for development without API costs:

brew install ollama  # macOS
ollama serve
ollama pull llama3.1:8b

local:
  name: "Local Only"
  providers:
    - name: "Ollama Llama"
      provider: "ollama"
      model: "llama3.1:8b"
      base_url: "http://localhost:11434"
      default: true

OpenAI

openai:
  name: "OpenAI"
  providers:
    - name: "GPT-4"
      provider: "openai"
      model: "gpt-4-turbo"
      api_key_env: "OPENAI_API_KEY"
      default: true

CI in 60 Seconds

Gate your MCP service on eval results in any CI system — no wrapper required:

# .github/workflows/mcp-tests.yml
jobs:
  mcp-tests:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: astral-sh/setup-uv@v5
      - name: Run MCP eval suite
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          uvx testmcpy run tests/ \
            --mcp-url "$MCP_URL" \
            --gate --min-pass-rate 85 \
            --junit-xml junit.xml

• --gate exits non-zero when the run fails your thresholds, so the build fails. Tune thresholds in .testmcpy-gate.yaml:

  min_pass_rate: 85.0       # % of tests that must pass
  max_failures: 3           # absolute failure budget
  required_tests:           # these must always pass
    - critical_auth_flow
  block_on_regression: true # fail on baseline regressions

• --junit-xml emits JUnit XML for CI systems that ingest it natively (Jenkins, GitLab, CircleCI, Buildkite). On GitHub Actions, pair it with an action like dorny/test-reporter — or just rely on the next bullet.

• Inside GitHub Actions, the markdown eval report is automatically appended to the job summary — results render on the workflow run page with zero extra steps.

Or use the bundled reusable Action — adds a sticky PR comment, JUnit artifact upload, and structured outputs (pass-rate, gate_passed):

- uses: preset-io/testmcpy@v1
  with:
    test_path: tests/
    mcp_url: ${{ vars.MCP_URL }}
    pass_threshold: '85'
    pr_comment: 'true'
    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}

Custom Evaluators

Extend testmcpy with domain-specific validation:

from testmcpy.evals.base_evaluators import BaseEvaluator, EvalResult

class MyEvaluator(BaseEvaluator):
    def evaluate(self, context: dict) -> EvalResult:
        response = context.get("response", "")
        passed = "expected" in response
        return EvalResult(
            passed=passed,
            score=1.0 if passed else 0.0,
            reason=f"Check passed: {passed}",
        )

See the Evaluator Reference and the Custom Evaluators guide for complete documentation.

Examples

Check out the examples/ directory for:

• Basic test suites — Simple examples to get started
• CI/CD integration — GitHub Actions and GitLab CI workflows
• Custom evaluators — Building domain-specific validation
• Multi-model comparison — Benchmarking different LLMs

Contributing

We welcome contributions! Whether it's bug reports, feature requests, documentation improvements, or code contributions.

Read the Contributing Guide to get started.

Community & Support

• Issues: Report bugs or request features
• Discussions: Ask questions and share ideas
• Documentation: preset-io.github.io/testmcpy (agent-facing source docs live in context/)
• Examples: Explore examples/ for sample code

License

Apache License 2.0 — See LICENSE for details.

---

Built by @aminghadersohi at Preset.