A standard in four months: OpenAI adopts its rival’s connection protocol
In November this blog called the Model Context Protocol boring and important. Last week OpenAI adopted it. That is unusually fast for a standard.
On 26 March 2025 OpenAI said it would support the Model Context Protocol across its products — the standard Anthropic open-sourced on 25 November 2024 for connecting AI systems to the places business data lives.
A competitor adopting a rival’s protocol four months after publication is rare. It usually takes years, or it never happens and everyone keeps their own interface.
Why it went this fast
Because nobody was winning the integration problem, and everybody was paying for it. Every AI vendor was building the same connectors to the same business systems, separately, and every customer was asking the same question: does this work with what we already have?
A shared protocol is cheaper for all of them than fifty proprietary ones. That is the ordinary economics of standards, and it is why the boring news from November turned out to matter more than the model releases around it.
What it means for an SME, concretely
Not much this year, and a lot over three.
The question « will this AI tool work with our ERP? » has been the practical blocker for most SME projects. The answer has been: only if someone builds a connector, and someone usually did not. A standard turns that from a bespoke project into a question of whether both sides support it.
Concretely, add one line to your vendor conversations, alongside the model-swap question from February: how does your tool connect to our systems, and does it use an open standard?
Most vendors will still say no in 2025. The ones who say yes are telling you something about how long they expect to be around.
Closing March
This month produced reasoning on a single graphics card, agent toolkits as standard issue, cheap transcription, a connection standard that stopped being one company’s idea, and Switzerland signing a convention whose effects arrive in 2027.
The recommendation is unchanged, because it has been unaffected by six months of announcements: clean the data for one process, automate that process, decide who checks the output, run it for a month.
Every company that did that in October has something working now. Every company waiting for the technology to settle is exactly where it was.