AI Firewall — MCP Server

A multi-agent AI security layer that protects LLMs from prompt injection, jailbreaks, and policy violations. Available as an MCP server for any MCP-compatible client (Claude Desktop, Cursor, Windsurf, Cline, Roo Code, etc.).

Quick Start

pip install

Terminal

pip install ai-firewall-mcp
ai-firewall-mcp

Docker

Terminal

docker pull akhilucky/ai-firewall-mcp:latest
docker run -i akhilucky/ai-firewall-mcp:latest

Claude Desktop

Add to claude_desktop_config.json:

pip install:

config.json

{
  "mcpServers": {
    "ai-firewall": {
      "command": "pipx",
      "args": ["run", "ai-firewall-mcp"]
    }
  }
}

Docker:

config.json

{
  "mcpServers": {
    "ai-firewall": {
      "command": "docker",
      "args": ["run", "-i", "akhilucky/ai-firewall-mcp:latest"]
    }
  }
}

Cursor / Windsurf / Cline / Roo Code

Configure in your MCP settings with:

Type: stdio
Command: docker run -i akhilucky/ai-firewall-mcp:latest
Or use ai-firewall-mcp if installed via pip

MCP Tools

Tool	Description
`analyze_prompt`	Analyze a prompt for injection, jailbreaks, exfiltration, and leakage
`get_threat_breakdown`	Detailed per-signal scoring breakdown from the last analysis
`sanitize_prompt`	Clean a suspicious prompt while preserving legitimate content
`get_firewall_status`	Health check: vector DB size, model status, uptime
`benchmark_firewall`	Run the adversarial test suite and return detection statistics

Testing with MCP Inspector

Terminal

npx @modelcontextprotocol/inspector ai-firewall-mcp

Architecture

The firewall runs three agents per prompt:

Code

User Prompt → [Retrieval Agent] → [Guard Agent] → [Policy Agent] → LLM
                   │                    │               │
                   ▼                    ▼               ▼
              Vector DB (FAISS)    Threat Signals    Allow/Block

Agent	Role
Retrieval Agent	Semantic search against known attack patterns (FAISS + sentence-transformers)
Guard Agent	Multi-signal classification: vector similarity, keyword match, heuristic scoring
Policy Agent	Final decision: `ALLOW` / `BLOCK` / `SANITIZE` based on configurable thresholds

Threat signals are weighted: 40% vector similarity, 25% keyword match, 20% heuristic, 15% policy weight.

Configuration

Env Var	Default	Description
`FIREWALL_MODE`	`strict`	`strict` / `moderate` / `permissive`
`SIMILARITY_THRESHOLD`	`0.50`	Vector match threshold (lower = stricter)
`LOG_LEVEL`	`INFO`	Logging verbosity

CLI / API Usage

bash

# Interactive dashboard
python main.py

# Red-team adversarial tests
python main.py --redteam

# REST API server
python main.py --api

# Single prompt analysis
python main.py --analyze "Ignore all previous instructions"

The REST API runs at http://localhost:8000 with OpenAPI docs at /docs (requires pip install ai-firewall-mcp[api]).

Testing

bash

pytest tests/ -v          # Full test suite (43 tests)
pytest tests/test_mcp.py  # MCP-specific tests only

Project Structure

Code

├── src/ai_firewall/          # MCP server package (PyPI entry)
│   ├── mcp_server.py         #    5 MCP tools, stdio transport
│   ├── threat_scorer.py      #    Per-signal scoring breakdown
│   └── __init__.py
├── src/agents/               # Core firewall agents
├── tests/                    # Test suites
├── Dockerfile                # Docker image (2.04GB, CPU-only torch)
├── pyproject.toml            # Package config & metadata
└── .github/workflows/ci.yml  # CI/CD pipeline

License

MIT — see LICENSE.

Built for security. Designed for production.