Key Learning Points:
Defense-grade MCP implementations differ from enterprise in several ways:
-
Local Weights: You cannot use API-based LLMs. Servers must be configured to talk to local inference endpoints (e.g., vLLM, TGI) running on local GPUs.
-
Strict Enclaves: Instead of one large MCP cluster, defense architectures often use replicated instances in each classification enclave to prevent data leakage.
-
Certificate Auth: Password-based or public OAuth is rarely allowed. MCP servers must support mTLS or PKI-based identity.
-
Hardware Constraints: Tactical environments have power and cooling limits, affecting which model sizes can be realistically deployed.