ENGINEERING FIELD NOTE / mcpguard-lite

Deterministic risk analysis for MCP tool manifests

Accepts a raw MCP tools/list response and returns capability risks, severity, and missing security controls using transparent, explainable rules — it never executes or probes the listed tools.

Start with the declared tool surface

MCPGuard accepts a raw tools/list JSON response and analyzes names, descriptions, and input schemas. It reports capability risks such as command execution, filesystem writes, network access, and secret access without invoking the described tools.

Look for positive control declarations

A risky manifest may need explicit statements about approval, audit logging, filesystem restrictions, and network egress. Wording such as without approval or no audit logging must not count as evidence that a control exists. The regression tests cover these negative declarations.

Keep the conclusion within the evidence

The output describes what the supplied manifest declares. A description can omit a real control or claim one that the runtime does not enforce. A report is a review aid, not a runtime security guarantee. The tool-shaped HTTP endpoint is also distinct from a complete MCP transport.