Wrangling MCP chaos with Kubernetes and paranoia

What it does

Archestra is a self-hosted platform that corrals MCP servers into a centralized Kubernetes orchestrator with a built-in registry, chat interface, and RAG knowledge base. It wraps the whole thing in guardrails meant to stop prompt-injection attacks and data exfiltration, while giving platform teams visibility into who is using what and how much it costs. Think of it as an internal AI app store with a security team and a billing department baked in.

The interesting bit

Instead of trusting a single LLM to police itself, Archestra uses a “Dual LLM” architecture that isolates dangerous tool responses in security sub-agents, and it applies non-probabilistic guardrails to block data leaks rather than hoping the model behaves. The README also claims a dynamic optimizer can cut AI costs by up to 96% by automatically downgrading to cheaper models for simpler tasks.

Key highlights

• Centralized private MCP registry for self-hosted, remote, and third-party servers
• Kubernetes-native orchestrator managing server state, API keys, and OAuth
• Built-in RAG knowledge base with citations—no external vector database required
• “Dual LLM” security sub-agents to isolate potentially compromised tool outputs
• Cost controls with per-team/per-agent limits and dynamic model switching

Verdict

Platform teams at mid-to-large companies that are terrified of shadow IT MCP servers should evaluate this immediately. Solo developers or small teams already comfortable with local MCP servers will find it heavier than necessary.