1. Install MCPShield

npm install -g mcpshield

2. Create a Policy File

Create policy.yaml with your allowed operations:

# policy.yaml
default: deny

allow:
  - pattern: "SELECT * FROM products WHERE id = ?"
    description: "Allow reading individual products"

  - pattern: "SELECT * FROM orders WHERE user_id = ?"
    description: "Allow reading user's orders"

  - pattern: "INSERT INTO feedback (user_id, message) VALUES (?, ?)"
    description: "Allow users to submit feedback"

3. Start the Gateway

mcpshield gateway \
  --policy policy.yaml \
  --adapter postgres \
  --connection "postgresql://user:pass@localhost/db" \
  --audit audit.log

MCPShield is now running as an MCP stdio server. Point your AI tool at it!

4. Test It

MCPShield will allow queries matching your policy and block everything else:

# ✅ ALLOWED
SELECT * FROM products WHERE id = 123

# ❌ BLOCKED
DROP TABLE users

Configure Claude MCP Settings

Edit your ~/.claude/mcp_settings.json:

{
  "mcpServers": {
    "postgres-secure": {
      "command": "mcpshield",
      "args": [
        "gateway",
        "--policy", "/path/to/policy.yaml",
        "--adapter", "postgres",
        "--connection", "postgresql://localhost/mydb",
        "--audit", "/path/to/audit.log"
      ]
    }
  }
}

Use in Claude Code

Claude will now route all database requests through MCPShield. You get full policy enforcement and audit logging automatically.

Policy Structure

default: deny  # Start with nothing allowed

allow:
  - pattern: "SELECT * FROM table WHERE id = ?"
    description: "Human-readable description"
    tags: ["read-only", "public-data"]

deny:
  - pattern: "DROP TABLE *"
    description: "Explicitly block dangerous operations"

audit_all: true  # Log both allowed and denied requests

Pattern Matching

• •? matches a single parameter
• •* matches any value
• •Patterns are case-sensitive by default
• •Exact matches are fastest; wildcards are evaluated in order

Audit Log Format

[2026-02-09T12:34:56Z] ALLOWED | SELECT * FROM products WHERE id = 42 | policy_match: allow[0]
[2026-02-09T12:35:01Z] BLOCKED | DROP TABLE users | reason: no_policy_match
[2026-02-09T12:35:12Z] ALLOWED | GET /api/public/status | policy_match: allow[3]

What We Log

• ✓Timestamp (ISO 8601, UTC)
• ✓Decision (ALLOWED or BLOCKED)
• ✓Full operation/query attempted
• ✓Policy rule that matched (if any)
• ✗We do NOT log credentials or sensitive params

Does MCPShield slow down my AI?

Policy evaluation adds <1ms overhead. Network latency to your backend dominates request time.

Can I use MCPShield in production?

Yes! MCPShield is designed for production. Start with read-only policies and expand as you gain confidence.

What if a query doesn't match any policy?

Default deny means it's blocked. MCPShield logs the attempt so you can decide whether to add it to your policy.

How do I write good policies?

Start restrictive (SELECT only, specific tables). Run in audit mode to see what your AI tries to do. Expand policies incrementally.

Is MCPShield open source?

The core gateway is open source (MIT). Enterprise features like SSO and custom adapters are commercial.