MCP isn’t the cool kid anymore

Context : MCP is a protocol devised the peace and love Anthropic by to standardize how LLM can access tools. Once you connect a dedicated MCP server, your LLM will be able to identify your prompt with a feature available on that server and call the appropriate function. Let’s say you have a weather question and your Claude is connected to a weather mcp server, Claude will decide to call on its own the “get-weather()” function with the appropriate parameters and return the results.

Nice right? The right answer is “OH MY GOD” because with marginal effort at understanding context (and they get better at this every generation) you can plug tools as elaborate and let your AI use them.

Even better, you can think of an entirely new ECONOMY with companies building/selling/renting MCP server with dedidicated tasks. You keep your expertise hidden on your private MCP server and sell the access to the everyday LLM you have (let’s be honest folks) no way of competing with.

The final piece to make it one of the most glorious piece of AI engineering was the ability to plug website frontend as MCP interface : your API/webform on your website will then be able to be filled autonomously by an AI. Spoiler alert, it is called WebMCP.

So Oh sadness was my thoughts when I read about Perplexity and Cloudlfare deciding to kill the prodigal son (well one announced migrating to the old fashion REST api whereas the other wrote a scathing technical demonstration). To summerize the excellent article, each MCP needs to be logged into and swimming into all the context information eats a lot of token. Manageable on your debugging agent with one MCP server, a bit expensive and slow when you build your killer app requesting hundreds of them.

Let’s watch closely how the other big players react to this. Anthropic any comments?