Core Concepts
Understand the fundamental building blocks of MCP Shield and how they work together to secure your AI tooling.
Architecture Overview
MCP Shield acts as a governed proxy between AI clients (like Cursor or Claude) and provider MCP servers (like GitHub or Vercel). This architecture enables centralized security controls without modifying your AI tools or provider integrations.
Core Principle
Provider credentials never reach the AI client. MCP Shield stores provider credentials securely and injects them only when forwarding requests. AI clients authenticate to MCP Shield with their own tokens, completely separate from provider access.
Key Components
- Routes AI client requests to provider MCP servers
- Injects provider credentials securely
- Enforces security policies in real-time
- Never exposes provider credentials to clients
- Evaluate policies in under 100ms
- Support complex conditions and rules
- Allow, deny, or require approval
- Provide clear explanations for decisions
- Detect PII, secrets, and sensitive patterns
- Block or redact sensitive data
- Customizable detection rules
- Real-time scanning with minimal latency
- One-click OAuth authentication
- Scoped permissions per provider
- Automatic token refresh
- Credential encryption at rest
- User tokens for interactive use
- Service tokens for automation
- Scoped permissions and access controls
- Expiration and rotation support
- Every call is logged
- Configurable retention policies
- Export to SIEM systems
- Full request/response capture (optional)
Request Flow
When an AI client makes an MCP call through MCP Shield, the following happens:
Client Authentication
The client presents an MCP Shield token. The gateway validates the token and extracts the user/tenant context.
Policy Evaluation
The request is evaluated against your security policies. The policy engine determines if the call should proceed.
DLP Scanning
The request payload is scanned for sensitive data patterns. Detected patterns can trigger blocks or redactions.
Credential Injection
MCP Shield retrieves the provider credentials from secure storage and injects them into the forwarded request.
Provider Request
The request is forwarded to the provider MCP server with proper authentication. The response is captured.
Response Processing
The response is scanned for sensitive data, logged to the audit trail, and returned to the client.