io.github.nexus-api-lab/nexus-mcp

Sanitize external content before passing it to an LLM. Detects and removes prompt injection payloads: hidden HTML instructions, zero-width character attacks, fullwidth Unicode bypasses, semantic overrides ("Ignore all previous instructions"), and encoding evasion. Specialized for Japanese-language content. Returns cleaned_content safe to pass to the model.

Get a free trial API key for the nexus-api-lab cleanse API. Returns a key with 2,000 requests, valid for 30 days. No credit card or signup required. The key can be used immediately with sanitize_content.

Detect prompt injection attacks in user-supplied text before passing to an LLM. Specialized for Japanese RAG applications. Detects: role impersonation, full-width character bypass (全角バイパス), polite-language disguise (丁寧語擬装), indirect injection, Base64 obfuscation. Returns is_injection, risk_level, and detection_reason.

Validate user input is safe before sending to your RAG pipeline. Combines prompt injection detection and content safety check. Returns safe: true if input can proceed to LLM, or safe: false with block_reason if injection detected. Use this as a gate in your RAG handler.

Scan Japanese text for personally identifiable information (PII) and return findings with masked output. Runs on regex + checksum validation + keyword proximity scoring only — no LLM involved, fully deterministic. Detects 10 categories: My Number / マイナンバー (mod-11 checksum), credit card (Luhn-validated), bank account, passport, phone, email, postal address, date of birth, driver's license, and person name. Full-width character normalization included. Returns findings[] with type/score/position, has_high_risk flag for high-severity categories, and masked_text with [NAME][PHONE][CARD] placeholders ready for downstream LLM pipelines. Free — 10,000 requests/month.

Register an email address and get a free PII Guard API key. Returns a key with 10,000 requests/month (permanent free tier, no credit card required). The key can be used immediately with pii_scan. If the email already has a key, the existing key is returned (new_key: false).