mcpkernel-security

MCPKernel Security Skill

The security kernel for AI agents. MCPKernel provides deterministic, policy-enforced security for every MCP tool call.

What it does

When installed, MCPKernel gives you these security tools:

mcpkernel_scan_tool

Scan any MCP tool's description for prompt injection, poisoning, shadowing, and Unicode obfuscation attacks.

Example:

"Scan the filesystem tool for poisoning attacks"

mcpkernel_validate_policy

Validate MCPKernel YAML policy files for syntax and rule correctness.

Example:

"Validate my policy at policies/strict.yaml"

mcpkernel_discover_configs

Auto-discover MCP server configurations across all installed IDE clients (Claude Desktop, Cursor, VS Code, Windsurf, OpenClaw, Zed, etc.) and flag security issues like exposed secrets.

Example:

"Discover all my MCP server configs and check for security issues"

mcpkernel_check_taint

Analyze text for sensitive data — PII, secrets, API keys, user input.

Example:

"Check this text for sensitive data: my email is [email protected] and my key is sk-abc123"

mcpkernel_scan_skill

Scan OpenClaw SKILL.md files for dangerous shell commands, exfiltration patterns, hidden instructions, and metadata mismatches.

Example:

"Scan the skills/my-skill/SKILL.md for security issues"

mcpkernel_doctor

Run full MCPKernel health diagnostics — check config, dependencies, exposed secrets, file permissions, and available tools.

Example:

"Run MCPKernel doctor to check my setup"

Installation

# Install MCPKernel
pip install mcpkernel

# Install in OpenClaw
mcpkernel install --target openclaw

Or add directly to your OpenClaw config:

{
  "mcp": {
    "servers": {
      "mcpkernel": {
        "command": "mcpkernel",
        "args": ["mcp-serve"]
      }
    }
  }
}

Key Features

• OWASP ASI 2026 compliant policy engine with YAML rules
• Multi-mode taint tracking (PII, secrets, user input, external data)
• Tool poisoning detection with 10+ injection pattern rules
• DLP chain detection (e.g., read_file(.env) → http_post = blocked)
• SARIF v2.1.0 output for GitHub Code Scanning / Azure DevOps
• Multi-backend sandboxing (Docker, Firecracker, WASM, Microsandbox)
• Deterministic Execution Envelopes (Sigstore-signed, replayable)
• Tamper-proof audit logs with SIEM export (CEF, JSONL, CSV)
• eBPF syscall monitoring at kernel level