Summary
A guard rail you can put in front of a RAG pipeline in an afternoon.
It covers the two checks that belong at a RAG entry point — is this input trying to hijack the model, and does this text carry personal data — with Japanese-language coverage that English-first filters tend to miss.
What it is
A hosted MCP server fronting the jpi-guard and PII Guard security APIs, built for Japanese-language applications. It speaks Streamable HTTP over JSON-RPC and issues its own trial keys on first use.
What you get
- Text checked for prompt injection, returning is_injection, risk_level and detection_reason
- External content cleaned of injection payloads before it enters model context
- Text scanned for personal data, returning findings with type, score and position
- A masked version of the text, with placeholders in place of names, phone numbers and card numbers
- A single gate for RAG input that combines the checks and returns a verdict
- Free trial keys fetched by the server itself
Requirements
Nothing to start with — connect over HTTP and the server fetches a trial key on its own before the first check. Detection is tuned for Japanese text.
Setup effort
One command — claude mcp add --transport http nexus https://mcp.nexus-api-lab.com/