MCPSentry

from fastapi import FastAPI
from mcpsentry import MCPSentry

app = FastAPI()
# ... your existing routes ...

bridge = MCPSentry(app)                    # Your app now speaks MCP.

report = bridge.audit()                    # Pre-flight security audit.
if report.has_blockers():
    report.print_text()                    # 🔴 CRITICAL: /api/admin exposed, /api/auth/login exposed…
    raise SystemExit(1)

bridge.enable_guardrails(policy="block")   # Runtime prompt-injection blocking.

pip install mcpsentry

from fastapi import FastAPI
from mcpsentry import MCPSentry

app = FastAPI(title="My App")

@app.get("/api/users/{user_id}")
async def get_user(user_id: int):
    """Get a user by their ID."""
    return {"id": user_id, "name": "Alice"}

bridge = MCPSentry(app)            # Auto-discovers routes, mounts /mcp
print(bridge.summary())

# Always audit before deploying
report = bridge.audit()
report.print_text()

if report.has_blockers():
    raise SystemExit(1)

🛡️  MCPSentry audit report
   13 tools from 13 routes
   🔴 2 critical · 🟠 4 high · 🟡 3 medium · 🔵 1 low

   🔴 [CRITICAL] POST   /api/auth/login  —  Authentication endpoint exposed to LLM clients
      ↳ Add '/api/auth/login' to exclude_paths
   🔴 [CRITICAL] DELETE /api/admin/users/{user_id}  —  Admin / internal endpoint exposed to LLM clients
      ↳ Exclude this route from MCP exposure
   🟠 [HIGH    ] GET    /api/users/me  —  Response may leak PII fields: email, phone, address
      ↳ Mask PII before returning, or restrict this route from MCP exposure
   🟠 [HIGH    ] POST   /api/internal/run  —  No docstring — an LLM will guess this tool's purpose
      ↳ Add a clear docstring or `summary=` to the route decorator
   🟡 [MEDIUM  ] DELETE /api/recipes/{recipe_id}  —  Destructive DELETE — an LLM call can mutate/delete data
      ↳ Document the consequences in the docstring; consider requiring confirmation
   🔵 [LOW     ] GET    /api/search  —  3 parameters fall back to 'string' — LLMs may send malformed inputs
      ↳ Annotate parameters with explicit types (int, bool, list[str], Pydantic models)

report = bridge.audit()

# Counts
report.critical_count           # → 2
report.high_count               # → 4

# Structured access (CI/CD, dashboards, reporting)
import json
print(json.dumps(report.to_dict(), indent=2))

# Iterate
for finding in report.findings:
    if finding.severity == Severity.CRITICAL:
        send_to_security_team(finding)

# .github/workflows/mcp-audit.yml
- run: python -c "from myapp import bridge; r = bridge.audit(); r.print_text(); exit(1 if r.has_blockers() else 0)"

bridge = MCPSentry(
    app,
    name="My Recipe Manager",
    description="Search recipes, plan meals",
    include_paths=["/api/*"],                          # Only expose API routes
    exclude_paths=["/api/admin/*", "/api/auth/*"],     # Hide sensitive endpoints
    max_tools=30,                                      # Limit context window usage
    mcp_endpoint="/mcp",
)

# Refine specific tools for better LLM understanding
bridge.tool("/api/recipes/search", description="Find recipes by name or ingredients")
bridge.exclude("/api/internal/*")

git clone https://github.com/miloudbelarebia/mcpsentry
cd mcpsentry
pip install -e .
python examples/mealie_demo/app.py

{
  "mcpServers": {
    "mealie": { "url": "http://localhost:9925/mcp" }
  }
}

┌─────────────────────────────────────────────────────┐
│                  Your FastAPI app                    │
│                                                     │
│   @app.get("/api/recipes")     ← Existing routes    │
│   @app.post("/api/recipes")                         │
│   @app.delete("/api/admin/...")  ← BAD              │
│                                                     │
│   ┌─────────────────────────────────────────────┐   │
│   │            MCPSentry (embedded)              │   │
│   │                                             │   │
│   │  1. Introspect routes at startup            │   │
│   │  2. Extract Pydantic schemas + type hints   │   │
│   │  3. ⚡ Pre-flight security audit             │   │
│   │     ↳ refuse to start on CRITICAL findings  │   │
│   │  4. Generate MCP tool definitions           │   │
│   │  5. Serve MCP JSON-RPC on /mcp              │   │
│   └─────────────────────────────────────────────┘   │
│                                                     │
│   POST /mcp  ← LLM clients connect here             │
└─────────────────────────────────────────────────────┘
         │
         ▼
   ┌───────────┐ ┌───────────┐ ┌───────────┐ ┌─────────┐
   │  Claude   │ │  ChatGPT  │ │  Gemini   │ │ Cursor  │
   └───────────┘ └───────────┘ └───────────┘ └─────────┘

bridge = MCPSentry(app)
bridge.audit()                         # pre-flight
bridge.enable_guardrails(policy="block")   # runtime

bridge.enable_guardrails(policy="block")   # refuse the call (default)
bridge.enable_guardrails(policy="alert")   # let it through, log loudly + on_alert callback
bridge.enable_guardrails(policy="log")     # shadow mode — just observe

def to_security_team(decision):
    slack.post(f"⚠️ MCPSentry blocked {decision.tool_name}: {decision.reason}")

bridge.enable_guardrails(policy="block", on_block=to_security_team)

bridge.guardrail.stats()
# → {"total": 1284, "blocked": 7, "alerted": 23, "clean": 1254}

for entry in bridge.guardrail.recent(10):
    print(entry.tool_name, entry.decision.allowed, entry.decision.reason)

{
  "isError": true,
  "content": [{
    "type": "text",
    "text": "🛡️ Blocked by MCPSentry runtime guardrail.\nReason: Prompt-injection detected (HIGH:1 MEDIUM:0 LOW:0)\nTop matches: high instruction_override @ arguments.query"
  }]
}

git clone https://github.com/miloudbelarebia/mcpsentry
cd mcpsentry
pip install -e ".[dev]"
pytest