I don't think MCP is as simple as dropping an agent into a workflow and expecting it to magically discover all the APIs out there and figure out how to connect to them. You kind of need a standard, which MCP is, to be able to enable that. But even once you have that standard, even once you have MCP servers for a given API, which I think, is difficult to rely on an AI to create those itself, you still need to solve a lot of topics around how the agent connects to the API: - Governance - Logging - Security These are all important considerations to mitigate the unintended side effects of allowing an agent unfettered access to all of your internal APIs, where it could potentially take very destructive actions. My co-founder Simon Yu and I have been working with many organizations facing and struggling with these issues. In building Gram, our MCP server generation product, we've been helping our customers discover the answers to these questions as they embrace the future of API tooling. Simon recently went on the Cloudcast podcast to talk about MCP and the problems our customers. Check it out are in the comments below 👇🏾👇🏾
