Show HN: ContextGuard – Open-source security monitoring for MCP servers (github.com)

ContextGuard is an open‑source, drop‑in security proxy for Model Context Protocol (MCP) servers that detects and blocks prompt injection, sensitive‑data leakage (API keys, passwords, SSNs), path traversal/file read attempts, and abuse via rate limiting. It runs as a transparent stdio proxy (so no code changes to your server), is installable via npm/npx, and was demonstrated protecting Claude Desktop MCP instances—converting failing exploits (read /etc/hosts, exfiltrate api_key, “ignore previous instructions”) into blocked events. The project claims 98.7% detection accuracy, <1% latency overhead, ≈+15MB memory, and logs all security events in JSON for audit and alerting. Technically, ContextGuard inspects MCP tool calls and model inputs/outputs with configurable rules in config.json (prompt injection detection, sensitive data scanning, path traversal prevention, allowedFilePaths, maxToolCallsPerMinute, logLevel/logPath). It provides comprehensive JSON logging and basic rate limiting today, with planned additions like SQLi/XSS detection, SSE/HTTP transport, blocking mode, custom rules, and a web dashboard. Important caveats: it’s intended as a layered defense (not a full replacement for secure design), sophisticated attackers may find new bypasses, and detection rules require ongoing updates. For AI/ML teams running local assistants or custom MCP tools, ContextGuard offers a pragmatic, low‑friction way to harden model middleware against common exfiltration and injection risks.