Model-Context-Protocol (MCP) is one of the most interesting integration mechanism among AI modules(and agents). It is easier to understand, setup and can be implemented on almost all popular programming languages. It is supported by almost all AI modules/agents (Claude, Gemini, Copilot, ChatGPT etc.). But as with most new technology, developments and innovations, there are challenges. It brings a number of limitations or concerns when it comes how the data and tools will be shared, how the underlying communication happens and most importantly how security is implemented (both authentication, authorization). In this talk I want to highlight these concerns and provide a few ways to solve this.