MCP Server Tester

The MCP Server Tester is a tool for automated testing of MCP servers.

Why this tool?

Building reliable MCP servers that work consistently and are spec-compliant is harder than it should be:

• the Inspector tool helps to build and test the intial server, but won't protect you from regressions
• tools works perfectly in direct testing, but LLMs use them incorrectly and inconsistently in the real world
• most MCP SDKs aren't fully spec-compliant, and it's hard to know what's missing
• the spec evolves rapidly, making it hard to stay current with best practices

The mcp-server-tester is a implementation-agnostic CLI testing tool that provides three focused commands that help solve these core issues:

• tools - Run tests that verify your tools work correctly with direct MCP calls
• evals - Run tests that verify LLMs can consistently understand and use your tools correctly
• compliance - Automated server spec validation as the protocol evolves (Work in Progress)

---

Commands

tools - Direct MCP Tool Testing

Test your tools work correctly with direct API calls:

npx mcp-server-tester tools tool-tests.yaml --server-config filesystem-server-config.json

Sample output:

✅ Tool Discovery: Found 3 tools (read_file, write_file, delete_file)
✅ Write file successfully
✅ Read file returns content
✅ Write and read workflow
❌ Handle invalid path - Expected error containing 'permission' but got 'access denied'
✅ Delete file cleanup

Results: 5/6 tests passed

evals - LLM Integration Testing

Test that LLMs can discover and use your tools effectively:

npx mcp-server-tester evals eval-tests.yaml --server-config filesystem-server-config.json

Sample output:

🤖 Testing with claude-3-5-haiku-latest
✅ LLM can create and read files
✅ File manipulation workflow
✅ Error handling test
❌ Documentation query task: FAILED
    Prompt: "How do I search for configuration files in the project?"
    • Required tool 'search_docs' was not called (actual calls: list_files)
❌ Lists all available tools: FAILED
    Prompt: "Please list all available file management tools you have access to."
    • LLM judge failed: Score: 0.3/1.0 (threshold: 0.8). Assistant did not provide complete tool list.

Results: 3/5 tests passed

Note: Requires ANTHROPIC_API_KEY environment variable.

compliance - MCP Specification Compliance Testing (WIP)

Test that your server is spec compliant:

npx mcp-server-tester compliance --server-config filesystem-server-config.json

See the Compliance Command (WIP) section below for details and example output.

Quick Start

1. Create server config (filesystem-server-config.json):

   {
     "mcpServers": {
       "filesystem": {
         "command": "node",
         "args": ["./filesystem-server.js"]
       }
     }
   }
   

2. Create tool and eval test files manually:

   tool-tests.yaml:

   tools:
     expected_tool_list: ['write_file']
     tests:
       - name: 'Write file successfully'
         tool: 'write_file'
         params: { path: '/tmp/test.txt', content: 'Hello world' }
         expect: { success: true }
   

   eval-tests.yaml:

   evals:
     models: ['claude-3-5-haiku-latest']
     tests:
       - name: 'LLM can write files'
         prompt: 'Create a file at /tmp/greeting.txt with the content "Hello from Claude"'
         expected_tool_calls:
           required: ['write_file']
         response_scorers:
           - type: 'llm-judge'
             criteria: 'Did the assistant successfully create the file?'
             threshold: 0.8
   

   See the Tools Testing and Evals Testing sections for comprehensive syntax examples.

3. Create tool and eval tests automatically using an agent:

   Try out this prompt, replacing the server config information with your own:

   Please create tool tests and eval tests for me to use with the mcp server tester tool.
   To see how to use it, run the tool's documentation and schema commands:
   
     `npx -y mcp-server-tester --help`
   
   My server config file is at ./filesystem-server-config.json. To know what tools you need to create tests for, run this command:
   
     `npx -y @modelcontextprotocol/inspector --cli --config filesystem-server-config.json --server filesystem-server --method tools/list`
   
   Please follow these steps:
   
   1. Create tool tests
     - Create a file called `tool-tests.yaml` that contains a single test for each tool. Follow these guidelines:
       - Do NOT force an individual test to pass; if the expected output is not returned, the test should fail
       - if there is a clear dependency between tool calls, you can chain them using the "calls" property
     - Run the tests and confirm that the syntax is correct and that each test runs (they do not have to pass)
   
   2. Create eval tests
     - Create a file called `eval-tests.yaml` with eval tests that will test the server's behavior. Follow these guidelines:
       - start with a few simple evals, and then build up to more complex ones
       - create between 5 and 10 eval tests
     - Run the tests and confirm that the syntax is correct and that each test runs (they do not have to pass)
   
   3. Provide a summary, which includes:
     - a list of the tools that are being tested and what you chose to test
     - a list of the eval tests and your reasoning for why you chose them
     - an explanation of how to run the tool tests and eval tests
   

4. Run tests:

   # Run tools tests (fast, no API key needed)
   mcp-server-tester tools tool-tests.yaml --server-config filesystem-server-config.json
   
   # Run LLM evaluation tests (requires API key)
   export ANTHROPIC_API_KEY="your-key"
   mcp-server-tester evals eval-tests.yaml --server-config filesystem-server-config.json
   

Tools Testing

The tools command tests your MCP server tools with direct API calls. This is fast, requires no API keys, and helps verify your tools work correctly before testing with LLMs.

Basic Syntax

tools:
  expected_tool_list: ['tool1', 'tool2'] # Optional: verify specific tools are available
  tests:
    - name: 'Test description'
      tool: 'tool_name'
      params: { param1: 'value1' }
      expect: { success: true }

Single Tool Tests

Test individual tool calls:

tools:
  tests:
    # Basic success test
    - name: 'Write file successfully'
      tool: 'write_file'
      params: { path: '/tmp/test.txt', content: 'hello world' }
      expect: { success: true }

    # Error handling test
    - name: 'Handle invalid path'
      tool: 'write_file'
      params: { path: '/invalid/path.txt', content: 'test' }
      expect:
        success: false
        error: { contains: 'permission denied' }

    # Result validation test
    - name: 'Read file returns content'
      tool: 'read_file'
      params: { path: '/tmp/existing.txt' }
      expect:
        success: true
        result: { contains: 'expected content' }

Multi-Step Workflow Tests

Test sequences of tool calls:

tools:
  tests:
    - name: 'Write and read workflow'
      calls:
        - tool: 'write_file'
          params: { path: '/tmp/workflow.txt', content: 'test data' }
          expect: { success: true }

        - tool: 'read_file'
          params: { path: '/tmp/workflow.txt' }
          expect:
            success: true
            result: { equals: 'test data' }

        - tool: 'delete_file'
          params: { path: '/tmp/workflow.txt' }
          expect: { success: true }

Expectation Types

Success/Failure:

expect: { success: true }   # Tool should succeed
expect: { success: false }  # Tool should fail

Error Validation:

expect:
  success: false
  error: { contains: 'permission denied' } # Error message must contain text

Result Validation:

expect:
  success: true
  result: { contains: 'partial text' }      # Result must contain text
  result: { equals: 'exact match' }         # Result must exactly match

Evals Testing

The evals command tests that LLMs can discover and use your tools effectively. This requires an ANTHROPIC_API_KEY and tests real-world usage patterns.

Basic Syntax

evals:
  models: ['claude-3-5-haiku-latest']
  max_steps: 5 # Optional: limit conversation turns
  tests:
    - name: 'Test description'
      prompt: 'Task for the LLM to complete'
      expected_tool_calls:
        required: ['tool1', 'tool2']
        allowed: ['tool3']
      response_scorers:
        - type: 'llm-judge'
          criteria: 'Did the assistant complete the task?'

Tool Call Validation

Control which tools the LLM should use:

evals:
  tests:
    - name: 'LLM must use specific tools'
      prompt: 'Write a paragraph of poetry to file "poetry.md"'
      expected_tool_calls:
        required: ['write_file'] # Must use these tools
        allowed: ['read_file'] # Can use this tool, but all other tool uses will throw error

Response Scoring

Evaluate the quality of LLM responses:

Regex Scorer:

response_scorers:
  - type: 'regex'
    pattern: 'success|completed|done' # Response must match pattern

LLM Judge Scorer:

response_scorers:
  - type: 'llm-judge'
    criteria: 'Did the assistant successfully complete the file operations?'
    threshold: 0.8 # Score must be >= 0.8

Advanced Options

IDE Schema Validation

Add this line to the top of your test files to enable automatic schema validation and autocomplete in IDEs like VS Code:

# yaml-language-server: $schema=https://raw.githubusercontent.com/steviec/mcp-server-tester/refs/heads/main/src/schemas/tests-schema.json

Environment Variable Replacement:

Use ${VAR_NAME} syntax to inject environment variables into your test configurations:

export API_BASE_URL="https://api.example.com"
export TEST_USER="alice"
mcp-server-tester tools test.yaml --server-config server.json

tools:
  tests:
    - name: 'Test ${TEST_USER} API access'
      tool: 'api_call'
      params:
        url: '${API_BASE_URL}/users/${TEST_USER}'
        method: 'GET'
      expect:
        success: true
        result: { contains: '${TEST_USER}' }

Environment variables work in test names, parameters, expectations, and all other config values. Missing variables will cause the config loading to fail with a helpful error message.

Debug Mode:

See full conversation output and scoring details:

mcp-server-tester evals eval-tests.yaml --server-config filesystem-server-config.json --debug

Multiple Models:

evals:
  models: ['claude-3-5-haiku-latest', 'claude-3-5-sonnet-latest']
  tests: [...]

Complex Evals:

evals:
  tests:
    - name: 'Multi-step file management'
      prompt: |
        Create a file called notes.txt with "Meeting notes from today".
        Then read it back and confirm the content is correct.
        Finally, create a backup copy called notes-backup.txt.
      expected_tool_calls:
        required: ['write_file', 'read_file']
        allowed: ['write_file', 'read_file', 'copy_file']
      response_scorers:
        - type: 'llm-judge'
          criteria: 'Did the assistant create the original file, read it back, and create a backup?'
          threshold: 0.8
        - type: 'contains'
          text: 'Meeting notes from today'

Unified Test Files:

You can combine both tools and evals sections in a single test file for convenience:

# Combined test file with both tools and evals
tools:
  expected_tool_list: ['write_file', 'read_file']
  tests:
    - name: 'Write file successfully'
      tool: 'write_file'
      params: { path: '/tmp/test.txt', content: 'Hello world' }
      expect: { success: true }

evals:
  models: ['claude-3-5-haiku-latest']
  tests:
    - name: 'LLM can write files'
      prompt: 'Create a file at /tmp/greeting.txt with content "Hello from Claude"'
      expected_tool_calls:
        required: ['write_file']
      response_scorers:
        - type: 'llm-judge'
          criteria: 'Did the assistant successfully create the file?'
          threshold: 0.8

When running commands, each will only use its respective section:

• mcp-server-tester tools unified-test.yaml uses only the tools: section
• mcp-server-tester evals unified-test.yaml uses only the evals: section

Compliance Command (WIP)

⚠️ Work in Progress: The compliance command does not fully cover the spec yet.

I'm not 100% convinced that a compliance checker should live in this tool. But it's been useful to me, so I'll include it here. It's far from complete.

npx mcp-server-tester compliance --server-config filesystem-server-config.json

Example output:

🏥 MCP SERVER COMPLIANCE
Diagnosing server: Filesystem MCP (MCP Protocol 2024-11-05)
Started: 7/19/2025, 7:53:57 AM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📋 MCP SPECIFICATION COMPLIANCE SUMMARY

🔍 BASE PROTOCOL
   └─ Transport Layer      ✅  3/3 passed
   └─ JSON-RPC 2.0         ❌  0/2 passed

🔍 LIFECYCLE
   └─ Initialization       ✅  2/2 passed
   └─ Capability Negotiation ✅  1/1 passed
   └─ Protocol Version     ❌  0/1 passed

🔍 SERVER FEATURES
   └─ Tools                ❌  3/4 passed
   └─ Resources            skipped (not advertised)
   └─ Prompts              skipped (not advertised)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📊 OVERALL MCP COMPLIANCE: 80/100 (2 features skipped)

Server Capabilities: tools ✅ | prompts ⏭️ | resources ⏭️

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
⚠️  SPECIFICATION WARNINGS (4)

⚠️  SDK: JSON-RPC Protocol Compliance [Base Protocol - JSON-RPC 2.0]
   Expected: Full JSON-RPC 2.0 compliance
   Actual:   1 compliance issues detected
   → Fix: Check capability advertisement matches implementation
   → Fix: Verify method name spelling and casing

⚠️  SDK: Error Response Validation [Base Protocol - Error Handling]
   Expected: Proper JSON-RPC error codes and format
   Actual:   2 error handling issues detected
   → Fix: Ensure server returns proper JSON-RPC error codes
   → Fix: Verify that the server properly validates requests
   → Fix: Ensure error handling is implemented according to JSON-RPC spec

⚠️  Lifecycle: Protocol Version Negotiation
   Protocol version issues detected (2 findings, 4 validations)

⚠️  Server Features: Tools - Execution (tools/call)
   Expected: Proper tool execution or error handling
   Actual:   1 execution issues detected
   → Fix: Verify error response format compliance

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📈 DETAILED COMPLIANCE BREAKDOWN
   base-protocol: 3/5 tests passed
   lifecycle: 7/8 tests passed
   server-features: 3/12 tests passed

🔗 SPECIFICATION REFERENCES
• MCP Specification: https://spec.modelcontextprotocol.io/
• JSON-RPC 2.0: https://www.jsonrpc.org/specification
• Error Codes: https://spec.modelcontextprotocol.io/specification/basic/error-handling/

License

MIT