A Model Context Protocol server that exposes dephawk to an AI coding agent, allowing the agent to watch what a command's dependencies actually do at runtime — the surface it otherwise installs and runs blind.
Three tools with complete schemas and strong descriptions. All tools have clear action verbs (list_, audit_*) and detailed, LLM-optimized descriptions (100-300 chars). Input schemas are properly typed with required fields and additionalProperties constraints. However, output schemas are entirely undocumented, LLMs cannot plan downstream calls or extract structured data. Error handling and recovery guidance are absent. No parameter descriptions for 'cwd' in audit_command. Tool composition is sound (each does one thing), but lack of output documentation is a significant gap for agent reasoning.
Run a command (e.g. an install or a build/test script) under dephawk and report what its npm dependencies actually did at runtime — files read/written, network reached, processes spawned, secrets touched, and any recognised attack technique. Runs in OBSERVE mode: it records, it does not block (use `dephawk run --enforce` to block). Use this before trusting freshly-installed packages, or to investigate a suspicious dependency.
Vet a single npm package BEFORE trusting it: install it in a throwaway sandbox, run it (install scripts + import) under dephawk, and report what it actually does — secrets read, hosts reached, processes spawned, recognised attack techniques, likely exfiltration. Call this before adding a dependency you are unsure about. Accepts a registry name, name@version, or a local path/tarball.
List the supply-chain attack techniques dephawk recognises at runtime (cloud-metadata SSRF, dead-drop C2, CI/git/editor-hook persistence, registry self-publish, credential exfiltration), each with a plain-English gloss. No arguments. Use it to explain what a dephawk finding means.
Output schemas completely undocumented. LLMs cannot infer what fields audit_package and audit_command return, preventing structured downstream reasoning and chaining.
audit_command 'cwd' parameter lacks description. LLMs cannot determine if it accepts absolute paths, relative paths, or both, or what the default behavior is.
No error handling guidance. Tools do not document what errors are retryable, what indicates user misconfiguration, or how to recover from failures (e.g., package not found, command timeout).
audit_package accepts 'registry name, name@version, or local path/tarball' but no enum or format constraint. LLMs may pass invalid specs; validation rules should be explicit.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
No pagination or result-limiting guidance. audit_command may return unbounded output (files, network connections, processes); no mention of limits or how to handle large result sets.