CodeSlick MCP Server: 6 Security Tools for Cursor and Claude Desktop

The Problem: AI Tools Don't Know Your Code Is Vulnerable

When you write code in Cursor or iterate with Claude Desktop, the AI can suggest fixes, refactor logic, and explain patterns. What it cannot do — by default — is run a proper security scanner on what you just wrote. It can guess. It can pattern-match from training data. But it cannot run 306 deterministic security checks with CVSS scores, CWE mappings, and verified remediation examples.

The CodeSlick MCP Server closes that gap. It connects CodeSlick's static analysis engine directly to your AI environment. When you ask your AI to check code for vulnerabilities, it now has a real security tool to call — not just its own pattern memory.

The Six Tools

Why AI-BOM Is Different from SBOM

Every compliance framework that mentions software inventory was written before LLMs were embedded in production applications. SBOM tells you which open-source libraries your code links against. It says nothing about which AI model is processing your users' data, which agent framework is making autonomous decisions, or whether your API keys are referenced in source code or pulled from a secrets manager.

The EU AI Act, the NIST AI RMF, and emerging enterprise security policies are all starting to require AI component inventories. “What AI systems does this application use?” is now a question that compliance teams, security auditors, and procurement departments ask — and most engineering teams cannot answer it systematically.

generate_ai_bom answers it. Point it at any codebase and get a structured inventory of every AI component: providers, models, tools, frameworks, and key references. It is the only tool in any MCP server that does this.

Local Analysis. No Account. No Telemetry.

Every analysis runs on your machine. The MCP server shells out to the locally installed CodeSlick CLI — the same binary used in pre-commit hooks and CI pipelines. Your source code never leaves your environment.

Performance in Practice

We ran all five tools against test payloads during integration testing in both Cursor and Claude Desktop. Response times on a MacBook Air M1:

The CLI-based tools (analyze_code, detect_secrets) write a temp file and shell out to the CodeSlick binary — the 1–2 second overhead is the CLI startup cost, not the analysis itself. The in-process tools (SBOM, dependencies, AI-BOM) parse manifests directly and return sub-second.

Installation

The MCP server requires the CodeSlick CLI installed globally. One-time setup:

npm install -g codeslick

Cursor — add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "codeslick": {
      "command": "npx",
      "args": ["-y", "codeslick-mcp-server@latest"],
      "env": {}
    }
  }
}

Claude Desktop (macOS) — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "codeslick": {
      "command": "npx",
      "args": ["-y", "codeslick-mcp-server@latest"]
    }
  }
}

Claude Code (CLI):

claude mcp add codeslick -- npx -y codeslick-mcp-server@latest

Restart your AI tool after updating the config. The five tools will appear in the MCP tools panel (hammer icon in Claude Desktop, MCP panel in Cursor).

What's Next

The next release adds check_security_policy — a tool that reads a .codeslick.ymlfile and validates code against your team's defined security thresholds. The same policy file that blocks a CI pipeline will be enforceable from inside Cursor and Claude Desktop. Define it once, enforce it everywhere.

We are also expanding the MCP-specific check set. Our ecosystem scan of 32 community MCP servers identified recurring patterns — prompt injection vectors, missing tool parameter validation, unsafe data pass-through — that warrant dedicated detection rules beyond the 12 checks in v1.4.