A security MCP server for AI coding agents: each language via its native analyzer — Kotlin (detekt), Java (SpotBugs+FindSecBugs), Python (Bandit), C# (Roslyn), JS/TS (ESLint) — plus diff review and secure patterns.
Three tools with clear, domain-specific descriptions (security scanning, diff review, pattern lookup). All tools have input schemas with typed parameters and descriptions. However, output schemas are not formally documented in the code, responses are constructed ad-hoc via helper functions (_findings_payload, _pattern_to_dict). Parameter descriptions are adequate but lack constraint details (e.g., no guidance on path format, diff format, or framework enum values). Error handling is minimal, no recovery guidance or categorization. Tool naming is verb-forward and clear, but descriptions could be more concise (currently 150 - 250 chars, within baseline but verbose for LLM parsing).
Security-review a unified diff: flag issues the change *introduces*. Paste a unified diff (e.g. the output of `git diff`) after editing Kotlin code. Only findings on lines the diff adds are returned, so the agent sees exactly what its change is responsible for — a self-check before committing.
Get the secure way to do a risky Kotlin/JVM task, before writing it. Ask "how do I do X securely" and get vetted before/after snippets and the reason — so the agent writes the safe version the first time.
Scan Kotlin/JVM code for security issues using a 216-rule analyzer. Point this at a file or directory. It returns every security finding the analyzer reports — rule id, message, location, and severity — so the agent can fix issues *while writing*, not after.
Output schemas not formally documented. Responses constructed via helper functions (_findings_payload, _pattern_to_dict) with no visible schema definition. LLMs cannot predict response structure or plan downstream operations.
Parameter 'framework' in secure_pattern lacks enum constraint. Description says 'e.g. spring-webflux, ktor, vertx' but does not declare valid values. LLMs may hallucinate unsupported frameworks.
No error handling or recovery guidance. If a scan fails (e.g., path not found, analyzer not installed), the tool returns no actionable error message. LLM cannot self-correct or retry intelligently.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Parameter 'path' in security_scan lacks format guidance. Description says 'absolute, or relative to the project' but does not specify whether it accepts files, directories, or both, or what happens if the path does not exist.
Result truncation (_MAX_FINDINGS = 500) is not exposed to the LLM. If findings exceed 500, the response is silently capped with no indication that results are incomplete. LLM may believe it has the full picture.