I built Krypton MCP after realizing how much blind trust we put into LLMs once we hook them up to actual tools via MCP. If an agent gets tricked by prompt injection or pulls sensitive credentials into context, there's usually zero boundary sitting in between to block or log it. So I put together a lightweight proxy written in Go that intercepts traffic between LLM clients and MCP servers. A few core parts under the hood: - Prompt Injection Defense: Catches malicious payloads before they hit downstream tools. - In-Flight Masking: Tokenizes API keys and PII in real-time so they don't leak into logs or LLM context. - Merkle Tree Auditing: Generates tamper-proof audit logs so you can mathematically verify past execution histories. Kept it in Go to avoid introducing latency to local tool-calling loops. The project is completely open source: https://github.com/muhammetemire... Curious to see how you guys are handling security boundaries in your current agent setups. Any feedback or edge cases are appreciated!

Krypton MCP

Open-source security gateway for Model Context Protocol

Open-source security gateway for Model Context Protocol