Krypton MCP – Securing Local LLMs and MCP Tools with Zero-Trust Proxy in Go

Connecting local or cloud Large Language Model (LLM) applications directly to databases, file systems, and internal APIs via the Model Context Protocol (MCP) exposes a significant security risk. Prompt injection attacks and credential leakage from model context present vulnerabilities that are often unguarded. Over the past few weeks, I have developed Krypton MCP, a high-performance security proxy written in Go, serving as an interceptor between LLM clients and MCP servers.

The primary purpose of Krypton is to act as a zero-trust proxy to inspect, sanitize, and log all interactions in real-time. Most MCP setups implicitly trust the LLM, allowing the model to directly invoke tools without any inspection. Krypton intercepts tool calls and context windows before execution to identify and block malicious payloads.

The proxy also performs in-flight data masking, tokenizing sensitive information like API keys and passwords, preventing them from appearing in logs or model context. Krypton implements Merkle Tree audit logs for every interaction, enabling mathematical proof of the model's actions. The proxy isolates upstream database and cache credentials, eliminating the need to expose raw connection strings to the agent layer.

Krypton MCP is implemented in Go, designed to have minimal overhead for sub-millisecond proxy latency in local tool-calling scenarios. The proxy handles Server-Sent Events (SSE) transport, JSON-RPC framing, and native downstream proxying. The codebase is available on GitHub (https://github.com/muhammetemirerkut/krypton-mcp) and accompanied by comprehensive documentation at (https://muhammetemirerkut.com/krypton/).

I invite contributions, benchmark feedback, and discussion on security edge cases to enhance the security of local MCP setups. Open-source collaboration is encouraged to refine the security approach for local model and tool interactions.