Skillguard

What skillguard MCP server does

The skillguard MCP server brings SkillGuard's local security scanner into an agent workflow. It examines a skill directory containing a SKILL.md file and related hooks or scripts, then reports findings with rule identifiers, severity information, and file-and-line citations. The scanner checks for ten documented categories of attack patterns, including network mismatches, remote code execution, file-scope escalation, hook supply-chain risks, obfuscated payloads, credential harvesting, frontmatter-scope spoofing, embedded prompt injection, cross-skill privilege chaining, and marketplace typosquatting.

The MCP interface is intended for checking a skill before an agent installs or executes it. SkillGuard does not need an account, API token, or scan-time network connection. Its rule packs and pattern-matching engine are included in the package.

How it works

Run the package's mcp command to start a stdio MCP server. An MCP-compatible client can then discover and call the exposed scan_skill tool. The scan operates on local content rather than executing the skill. Results identify detected patterns and can be used by the calling agent to decide whether to continue.

The same project also provides command-line and library interfaces. The CLI supports scan for one skill and scan-set for a directory whose immediate children are skill directories. The set operation adds cross-skill analysis, which can identify combinations such as one skill reading sensitive files while another provides network egress. CLI output can be human-readable, JSON, or SARIF, with severity thresholds and timeout settings.

Setup and configuration

Install the JavaScript package with npm install --save-dev skillguard-cli, then start the MCP process with:

npx skillguard-cli mcp

No environment variables, account, or API token are required. The package includes its rule packs, so scanning does not fetch rules or submit scanned content to a remote service. Configure the MCP client to launch the command as a stdio server, then provide a local skill path when calling scan_skill.

For direct CLI use, npx skillguard-cli scan ./my-skill scans one directory. The Python distribution, installed with pip install skillguard-cli, provides a separate implementation with matching rule-pack behavior and library functions.

Tools and capabilities

• scan_skill scans a local skill directory through MCP.
• scan checks one skill from the command line.
• scan-set checks multiple skills and evaluates cross-skill privilege combinations.
• Structured results include findings, warnings, timeouts, and an exit code in the library API.
• CLI scans support human, JSON, and SARIF output.

Limitations and notes

The MCP command documented in the repository exposes scan_skill; the multi-skill scan-set operation is documented for the CLI and library interfaces. A clean scan returns exit code 0, while a finding at or above the selected severity threshold returns exit code 1. A missing path, or a single-skill scan with no skill files, returns exit code 2. Inline suppressions are disabled by default, and suppression files must be passed explicitly rather than being loaded automatically from the scan target.