MCP explained without hype or fluff
23 points by mogambo1
23 points by mogambo1
It’s worth noting that this is not a mature protocol—it is continuously evolving. But the adoption has been fantastic—I opened the first random MCP aggregating website and it lists over 4000 servers coming from various organisations and individuals. I’d estimate there’s a lot more out there.
This is a somewhat overly optimistic take on the situation. We’ve done some experimental integrations of these tools and our team’s observation is that they are very much unstable and quite flaky. I don’t get the sense that most of these tools are anywhere near ready for production usage.
How is that overly optimistic? It sounds like exactly what the author is saying with “this is not a mature protocol”
I was speaking specifically about the adoption being fantastic wording. I don’t consider really poor implementations that use the protocol is what I would call fantastic adoption. Most of them are pretty close to unusable.
My understanding after reading this is that a MCP client will define a bunch of things via RPC that it can do. This is fed into the LLM so that it knows what tools it can use. The LLM (which is synonymous with MCP server) will make calls RPC calls to the client and the results will be fed back into its context.
AFAIK, most applications already do this, but they would make their own adhoc protocol that the LLM should follow. Using MCP should allow LLM providers to fine-tune their models to work well with this this particular protocol.