MCP server for PhishClean security analysis tools — check URLs, passwords, emails, headers, JWTs, and source code for security issues
PhishClean provides 7 security-focused tools with complete input schemas and reasonable descriptions. All tools follow verb_noun naming (check_*, scan_*, analyze_*, decode_*) and have 10 - 1024 character descriptions. However, output schemas are entirely undocumented, LLMs cannot predict response structure or plan downstream calls. Parameter descriptions are minimal (e.g., 'The URL to analyze' lacks format/constraint guidance). No error handling guidance, no pagination support, and no indication of which tools are idempotent or safe to retry. Tools are read-only (good for security), but descriptions do not explicitly state this, forcing LLMs to infer safety.
Analyze raw email headers to check SPF/DKIM/DMARC authentication, trace relay hops, and detect email spoofing indicators.
Analyze email body text for phishing indicators: urgency language, threats, credential requests, suspicious links, generic greetings, and attachment references.
Check a URL for phishing signals including domain spoofing, suspicious TLDs, IP addresses, HTTPS downgrades, token leaks, and more. Returns risk level and detailed signals.
Analyze password strength. Checks length, character variety, common passwords, keyboard walks, dictionary words, date patterns, and calculates entropy and estimated crack time.
Decode a JSON Web Token and check for security issues: expired tokens, algorithm 'none', weak algorithms, sensitive data in claims, very long expiration, missing standard claims.
Output schemas completely undocumented. LLMs cannot predict response structure (e.g., does check_link return {risk_level, signals} or {score, details}?). Forces agents to guess field names and risks context loss when chaining tools.
Parameter descriptions lack format/constraint guidance. 'The URL to analyze' does not specify: must be http/https? Max length? What happens if invalid? LLMs cannot validate inputs before calling.
No error handling guidance. Tool descriptions do not explain what errors are possible, whether they are retryable, or what the LLM should do next. E.g., 'Invalid JWT', should the agent retry, ask the user, or move on?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Full security page scan — analyzes a web page for 12 phishing/security signals including password fields, domain mismatches, hidden iframes, JWT in URL, hardcoded secrets, and more. Returns a risk score.
Scan source code for leaked secrets: AWS keys, Stripe keys, GitHub tokens, Google API keys, Slack tokens, Twilio keys, SendGrid keys, private keys, bearer tokens, and more.
Descriptions do not explicitly state tools are read-only/safe to retry. LLMs cannot infer idempotency, they may avoid retrying on transient failures, reducing reliability.
No pagination or result-limiting guidance. If scan_secrets or scan_page return hundreds of findings, no indication of max results or how to paginate. Large responses waste tokens and risk context exhaustion.