MCP Testbench

Docker-isolated security testing for Model Context Protocol (MCP) servers.

MCP Testbench is a comprehensive security testing harness that runs MCP servers in isolated Docker containers and performs automated vulnerability scanning, fuzzing, and penetration testing.

Why MCP Testbench?

With critical vulnerabilities like CVE-2025-6514 (CVSS 9.6) and CVE-2025-49596 (CVSS 9.4) affecting MCP servers, security testing is no longer optional. MCP Testbench provides:

• 🌐 Universal Testing: Works with both stdio (local) and HTTP (remote) MCP servers
• 🔒 Docker Isolation: Test potentially malicious servers safely with --network none
• 🔍 CVE Scanning: Automatic detection of known vulnerabilities
• ⚡ Fuzzing Engine: Protocol conformance testing with malformed inputs
• 🛡️ Injection Testing: Prompt injection, SQL injection, XSS, and path traversal tests
• 📊 Security Scoring: Simple A-F grades with shields.io badges for READMEs
• 🚀 CI/CD Ready: JSON reports and GitHub Actions integration

Proven Results: Found critical F-grade vulnerabilities in popular MCP servers like time-mcp and docker-mcp.

Installation

CLI Installation

pip install mcp-testbench

Or from source:

git clone https://github.com/ubermorgenland/mcp-testbench
cd mcp-testbench
pip install -e .

GitHub Action

Add to .github/workflows/security.yml:

- name: MCP Security Scan
  uses: ubermorgenland/mcp-testbench@v1
  with:
    target: 'http://localhost:8000'

See GITHUB_ACTION.md for complete guide.

Quick Start

Testing Stdio MCP Servers (Local)

Most MCP servers use stdio transport (npx commands). Test any stdio server:

# Test time-mcp
mcp-testbench run --stdio "npx time-mcp"

# Test docker-mcp
mcp-testbench run --stdio "npx @edjl/docker-mcp"

# Test GitHub MCP server
mcp-testbench run --stdio "npx @modelcontextprotocol/server-github"

# With verbose output
mcp-testbench run --stdio "npx time-mcp" --verbose

# Custom output directory
mcp-testbench run --stdio "npx time-mcp" --output ./security-reports

Real Results: We tested popular MCP servers:

• time-mcp: F grade (5 crashes, 36% failure rate)
• docker-mcp: F grade (12 crashes, 86% failure rate)

Both fail on basic input validation. See detailed crash analysis.

Testing HTTP MCP Servers (Remote)

Test remote MCP servers over HTTP:

# Test local HTTP server
mcp-testbench run http://localhost:8000

# Test production deployment
mcp-testbench run https://mcp.yourcompany.com/

# With custom output
mcp-testbench run http://localhost:8000 --output ./security-reports

Docker Isolated Testing (Recommended)

For maximum security, run tests in Docker:

mcp-testbench run --docker --docker-path /path/to/mcp-server

This mounts your MCP server in an isolated container with:

• --network none (no network access)
• --cpus 2 (CPU limit)
• --memory 2g (memory limit)

Output

MCP Testbench generates:

1. JSON Report: Detailed test results with vulnerability details
2. Security Badge: shields.io markdown badge for your README
3. Terminal Summary: Human-readable results with risk levels

Example Badge

Add this to your MCP server's README:

![MCP Security Score](./mcp_testbench_report/SECURITY_BADGE.md)

Security Tests

CVE Scanner

• Checks for CVE-2025-6514 (mcp-remote RCE)
• Checks for CVE-2025-49596 (MCP Inspector RCE)
• Identifies vulnerable server versions

Fuzzing Engine

• Invalid JSON payloads
• Oversized inputs (100KB+ strings)
• Deeply nested objects (1000+ levels)
• Unicode exploits and null bytes
• Type confusion attacks
• Protocol conformance testing

Injection Testing

• Prompt Injection: Tool poisoning attacks
• Command Injection: Shell command execution attempts
• Path Traversal: File system access tests
• SQL Injection: Database query manipulation
• XSS: Cross-site scripting payloads

Architecture

CLI → TestEngine → PluginRegistry → [CVEScanner, Fuzzer, PromptInjection]
                                                    ↓
                                              Reporter → JSON + Badge

Plugin System

MCP Testbench uses a plugin architecture for extensibility:

from engine import Plugin
import httpx

class CustomPlugin(Plugin):
    async def run(self, client: httpx.AsyncClient) -> dict:
        # Your test logic here
        response = await client.get("/")
        return {
            "status": "completed",
            "custom_metric": response.status_code
        }

Save to plugins/custom_plugin.py and it will be automatically discovered.

Security Scoring

• A (Bright Green): No vulnerabilities, passes all tests
• B (Green): Minor issues, no critical vulnerabilities
• C (Yellow): Moderate risk, some tests failed
• D (Orange): High risk, multiple vulnerabilities or plugin errors
• F (Red): Critical vulnerabilities detected

Docker Isolation

MCP Testbench runs tests in isolated containers with:

docker run --network none --cpus 2 --memory 2g

This prevents malicious MCP servers from:

• Making network requests
• Consuming excessive resources
• Accessing host filesystem
• Compromising the testing environment

Roadmap

• GitHub Action for CI/CD ✅
• Docker isolation mode ✅
• Rate limiting tests
• Log sanitization checks
• Input validation tests
• Real-time CVE feed integration
• Observability metrics
• Registry signing verification

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

License

MIT License - see LICENSE for details.

References

• MCP Security is still Broken (Reddit) - 345 upvotes
• CVE-2025-6514: Critical RCE in mcp-remote (CVSS 9.6)
• CVE-2025-49596: RCE in MCP Inspector (CVSS 9.4)
• Model Context Protocol Specification

Built by the ApInference Team | Website | Twitter