AgentShield — AI Skill & MCP Security Scanner

Running untrusted AI skills or MCP servers on your local machine is a significant security risk. This skill performs defensive static analysis to identify vulnerabilities, over-privileged permissions, and malicious patterns before you install new agent capabilities. ## What it does - **Permission Auditing** identifies excessive filesystem, network, and environment variable access. - **Secret Detection** flags hardcoded credentials or dangerous secret-handling patterns without exposing the values. - **Injection Analysis** scans for prompt injection indicators and suspicious instructions inside agent assets. - **Execution Risk** detects shell command execution, hidden scripts, and arbitrary code patterns. - **Supply Chain Review** evaluates tool schemas and MCP configurations for external dependencies and risks. ## How it works 1. Point the agent to an AI skill directory, MCP config file, or tool schema. 2. The skill performs a static scan of the assets, scripts, and configuration files. 3. It categorizes findings by severity from Low to Critical with specific evidence. 4. You receive a structured security report including a permissions summary and remediation steps. ## Frameworks & tools This skill works with MCP (Model Context Protocol) servers, JSON tool schemas, Python/Node.js agent scripts, and standard configuration files like `mcp-config.json`. ## Why this beats prompting it yourself Manual inspection often misses subtle permission overlaps or obfuscated shell commands. This skill uses a structured risk model and mandatory reporting format to ensure every capability is explicitly accounted for and categorized by threat level. ## Use cases - Auditing a third-party MCP server before adding it to your Claude Desktop config. - Reviewing a downloaded AI skill from a marketplace for hidden network calls. - Checking local agent scripts for hardcoded API keys or wide file-system write access. - Benchmarking a new version of a tool against a known secure baseline. ## Known limitations Static analysis cannot guarantee the detection of all runtime malware or dynamic execution paths. It does not execute code or connect to endpoints to verify behavior.