What it does
The server exposes five MCP tools. analyze_prompt checks a prompt for injection, jailbreaks, exfiltration and leakage. get_threat_breakdown returns per-signal scores from the last analysis. sanitize_prompt cleans a suspicious prompt while keeping its legitimate content. get_firewall_status reports a health check covering vector database size, model status and uptime. benchmark_firewall runs an adversarial test suite and returns detection statistics.
For each prompt, three agents work in sequence. A retrieval agent runs semantic search against known attack patterns using FAISS and sentence-transformers. A guard agent combines vector similarity, keyword matching and heuristic scoring. A policy agent returns a decision of ALLOW, BLOCK or SANITIZE based on configurable thresholds. Threat signals are weighted 40% vector similarity, 25% keyword match, 20% heuristic and 15% policy weight. It is intended for use from MCP-compatible clients such as Claude Desktop, Cursor, Windsurf, Cline and Roo Code.
It can be installed with pip or run as a Docker container, and is added to a client as a stdio server. Three environment variables control behaviour: FIREWALL_MODE (strict by default, or moderate or permissive), SIMILARITY_THRESHOLD (default 0.50, where a lower value is stricter) and LOG_LEVEL (default INFO). A REST API is also available on localhost port 8000, with OpenAPI documentation at /docs.
Limitations and requirements
The REST API is not included in the base install and requires the optional api extra of the Python package. The server must be run from an MCP-compatible client.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
AI Firewall MCP
Multi-agent LLM security layer detecting prompt injection and jailbreaks.
v1.0.1As shown on registry.modelcontextprotocol.io when we last read it