Big question in my mind was if OpenAI was going to formally endorse this (since it was created by Anthropic) but we have our answer.
MCP is now the industry standard for connecting LLMs to external tools.
MCP is now the industry standard for connecting LLMs to external tools.
No it isn't. We're still five or ten years away from real standards.
(Anyways it's clear that the future is smaller self-hosted LLMs, so whatever standard eventually emerges will be based on that paradigm.)
2) For me it is not clear the future is smaller self-hosted LLMs. As of today, most useful for me is to use best models, and those are not self-hosted.
They'll be used for particular language classification and search tasks, and for that you'd want several lighter, faster, cheaper and more specialized models. Not one with an arbitrary "best" score that's based on tricking the Turing test.
Standards are already emerging, including MCP, and to say that simply because they’ll evolve and be replaced over time means they’re not ‘real’ now is ridiculous. Look at the early internet and web as examples.
Local models, even accounting for reasonable progress and device performance improvements, will always, inherently, be behind the eight ball compared to SOTA models. While they may be sufficient for the low hanging fruit, I’d not bet against the frontier models being every bit as compelling relatively speaking.
Using weasel words like ‘real’ and ‘x is the future’ is astoundingly arrogant, and anyone claiming with confidence that they’ve got any idea where we’re heading is almost assuredly wrong.
We've had tool call frameworks before, but we haven't had a way of making tools that our clients actually talked to. There was no way to build tools that other people could download and run locally without getting them to switch to a client that baked those tools in. It's like the difference between static linking and dynamic linking, but for tool calls.
The VCs that were invested in AI companies (like Cursor) would of course need to hype something up like MCPs to get us to build them as little of those tools did not exist.
Cursor already makes $100M+. So why not get behind and integrate this chosen standard to make even more money with more MCP servers.
The last ingredient is to hype it all up on the internet to get everyone building.
A win for the VCs regardless even though it was suspiciously orchestrated as soon as Cursor integrated it.
I'd like some more of that getting everyone building please.
Example: JetBrains released an MCP server plugin for their IDEs recently and it enables Claude Desktop to seamlessly use the IDE for reading and writing files and even compiling or running tests against it.
And that's just one example; there are MCP servers for anything from VSCode to Emacs already. That means almost anyone can now use LLMs to generate code in their IDE of choice, just as they would do with Cursor.
In fact, MCPs are posing a threat to many startups, especially the "wrapper" ones.
MCP let's you handle the processing outside of the agent context, redistribute tools independently, and provide them as either private or public hosted services.
Basically, add an HTTP layer on the existing concept of tools. It is not a replacement for a framework, it is an enhancement of a well established pattern.
VCs invested in AI and agentic companies needed a way to get you guys to accelerate agents this year.
So why not "create" artificial hype for MCPs on the internet, since there were little to no MCP servers for LLMs to use despite it being several months old (November 2024) until Cursor integrated it.
This is the true reason why you see them screaming about MCPs everywhere.
Anthropic subsidizes an open standard, but proprietary extensions later emerge that lock you into servers from their marketplace or something? I need a sociopath to help me understand this.
I could see a future where companies have their developer portal where now they have their APIs document, the pretty in swagger, the samples, etc, but they'll have a MCP endpoint (potentially), where they're safely exposing the data to an LLM. Your langchain node step to to get context, could call out to some of these hosted/shared mcps where you do standard stuff, like post to a Slack channel, grab some data from a SFDC instance, etc....