SMCP

MCP connects agents to tools. It has no security.

A proof of concept for Secure MCP: authentication, encryption and audit, fully backward compatible.

Kevin Keller · kevinkeller.org

Why it matters

Most MCP deployments trust the transport and nothing else

  • No authentication and no authorisation at the protocol level.
  • No message-level encryption.
  • No audit trail of which tools an agent used or what data it retrieved.

Acceptable for personal use. Not for enterprise production.

The gap

SMCP adds what standard MCP leaves to the developer

CapabilityStandard MCPSMCP
AuthenticationNoneAPI keys, JWT, OAuth2/SAML
AuthorisationNoneRole-based, per tool
EncryptionNoneECDH + AES-256-GCM
Audit trailNoneFull operation logging
Multi-agentNot supportedDiscovery and orchestration

The idea

Choose the security tier that matches your threat model

Security is progressive: start simple in development, add encryption and audit as the stakes rise.

Four tiers

From a local API key to regulated-industry audit

  1. Simple (development): API key, no encryption.
  2. Basic (staging): JWT over HTTPS with token refresh.
  3. Encrypted (production): ECDH key exchange, AES-256-GCM, HMAC.
  4. Enterprise (regulated): OAuth2/SAML with full audit trails and compliance reporting.

Encrypted mode

Messages stay protected even if TLS is compromised

  • Client and server exchange ECDH public keys once per session.
  • Both derive a shared AES-256-GCM key.
  • Every request and response is encrypted and HMAC-authenticated.

Encryption is independent of the transport layer.

Beyond security

Agents can find each other and share the work

  • Agents register with a discovery service and advertise capabilities.
  • An orchestrator splits requests into parallel subtasks.
  • Built-in load balancing spreads work across instances.
  • DuckDB and filesystem connectors inherit the requesting agent's security context.

Performance

Security adds overhead, but very little

OperationOverhead
ECDH key exchange< 100ms, once per session
Agent discovery< 50ms
AES-256-GCM encryption< 10ms per message
JWT validation< 5ms

The risks it addresses

Unsecured agent tooling opens four doors

  • Data exfiltration: tools read anything the host can
  • Prompt injection: manipulated responses leave no trail
  • Lateral movement: agents impersonate other agents
  • Compliance: regulated industries need an audit trail for every AI action

The takeaway

Security can be layered onto MCP without breaking compatibility.

Read the full article

Kevin Keller · Proof of concept, not a production framework