How to Register SkillSpector with Claude Code: A Complete MCP Setup Guide

Register SkillSpector with Claude Code by installing the MCP extras, launching the server with skillspector mcp, and executing claude mcp add skillspector -- skillspector mcp to enable the scan_skill security guardrail.

SkillSpector is NVIDIA’s open-source security scanner for AI agent skills. When you register SkillSpector with Claude Code, you create a Model-Context Protocol (MCP) bridge that intercepts potentially unsafe skill deployments before they execute. This integration runs over stdio transport and requires no external API keys when configured with the Claude CLI provider.

Install SkillSpector with MCP Support

Before registering, install the package with the optional MCP dependencies. The pyproject.toml in the repository defines an mcp extra that includes FastMCP and related server components.

uv tool install --force 'skillspector[mcp] @ git+https://github.com/NVIDIA/SkillSpector.git'

This command installs the CLI entry point and the server implementation found in src/skillspector/mcp_server.py.

Start the SkillSpector MCP Server

The server exposes a single tool, scan_skill(target, use_llm=true, output_format="json"), which performs static analysis and optional LLM-driven semantic validation. The implementation in src/skillspector/mcp_server.py defines a FastMCP-compatible server that handles RPC calls over stdio.

Launch the server using the CLI entry point defined in src/skillspector/cli.py (lines 466-499):

# Configure the Claude CLI provider (uses existing `claude` binary, no API key required)

export SKILLSPECTOR_PROVIDER=claude_cli

# Start the MCP server over stdio (fastest transport for local agents)

skillspector mcp

The stdio transport provides the lowest latency for Claude Code interactions, as the server runs as a child process and communicates through standard input/output streams.

Register the Tool with Claude Code

With the server running in one terminal, register it with Claude Code using the MCP discovery command. The repository’s README.md documents this exact registration syntax:

claude mcp add skillspector -- skillspector mcp

This one-liner tells Claude Code to add a tool named skillspector whose implementation is the skillspector mcp command. After registration, Claude Code can invoke skillspector.scan_skill(...) from within any skill’s execution flow.

The server returns a JSON object containing risk_score, severity, recommendation, and safe_to_install flags. Claude Code uses this verdict to abort or modify execution based on the security assessment.

Provider Architecture and Configuration

SkillSpector’s MCP implementation is provider-agnostic. When use_llm=true is passed to scan_skill, the server routes the request to the configured provider without requiring specific LLM credentials.

The Claude CLI provider, implemented in src/skillspector/providers/claude_cli/provider.py, wraps the local claude binary for model inference. This provider leverages your existing claude auth login session and requires no separate API key.

Set the provider explicitly before starting the server:

export SKILLSPECTOR_PROVIDER=claude_cli

When the MCP server receives a scan request, it performs static analysis first, then optionally calls the provider for semantic analysis if use_llm is enabled.

Complete Workflow Example

Follow these steps to implement SkillSpector as a Claude Code guardrail:

1. Install and configure the environment:

   uv tool install --force 'skillspector[mcp] @ git+https://github.com/NVIDIA/SkillSpector.git'
   export SKILLSPECTOR_PROVIDER=claude_cli

2. Start the MCP server in a dedicated terminal:

   skillspector mcp

3. Register with Claude Code in another terminal:

   claude mcp add skillspector -- skillspector mcp

4. Invoke from a Claude Code skill (pseudo-code):

   result = await skillspector.scan_skill(
       target="https://github.com/example/my-skill",
       use_llm=True,
       output_format="json"
   )
   
   if not result["safe_to_install"]:
       raise RuntimeError(f"Skill rejected: {result['recommendation']}")

Summary

• Register SkillSpector with Claude Code using the single command claude mcp add skillspector -- skillspector mcp after starting the MCP server with skillspector mcp.
• The server implementation in src/skillspector/mcp_server.py exposes the scan_skill tool that performs both static and LLM-driven security analysis.
• Configure the Claude CLI provider via SKILLSPECTOR_PROVIDER=claude_cli to use the claude binary at src/skillspector/providers/claude_cli/provider.py, eliminating the need for API keys.
• The stdio transport provides fast local execution, making it ideal for guardrails that must evaluate skills before allowing Claude Code to execute them.

Frequently Asked Questions

What is the exact command to register SkillSpector with Claude Code?

Run claude mcp add skillspector -- skillspector mcp in your terminal. This tells Claude Code to add a tool named skillspector that executes the skillspector mcp command to start the MCP server. The registration persists in Claude Code’s configuration until you remove it with claude mcp remove skillspector.

Does SkillSpector require an API key when used with Claude Code?

No. When you set SKILLSPECTOR_PROVIDER=claude_cli, SkillSpector uses the Claude CLI provider implemented in src/skillspector/providers/claude_cli/provider.py, which wraps your local claude binary. This provider uses your existing claude auth login session and requires no separate API key for LLM analysis.

What parameters does the scan_skill tool accept?

The scan_skill tool accepts three parameters: target (the skill path or URL to scan), use_llm (boolean, defaults to true for semantic analysis), and output_format (string, defaults to "json"). These are defined in src/skillspector/mcp_server.py and return a JSON object containing risk scores, severity levels, recommendations, and a safe_to_install boolean.

Can I use SkillSpector with MCP servers other than Claude Code?

Yes. While the claude_cli provider is optimized for Claude Code integration, the MCP server in src/skillspector/mcp_server.py is provider-agnostic. It supports any MCP client over stdio or HTTP/SSE transports, and can interface with Bedrock, OpenAI, Anthropic, or other LLM providers by changing the SKILLSPECTOR_PROVIDER environment variable.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}