An engineer who worked on the enterprise side of ChatGPT has reversed course of the Model Context Protocol, the standard that lets AI models reach outside services. The engineer dismissed it as a pointless invention in late 2024. After two years of changing their mind more than once, they now calls it a good protocol that earned its place.
Why it looked unnecessary at first
When rumors of a new standard for exposing services to language models began circulating in late 2024, the reaction was blunt. OpenAPI, also known as Swagger, already existed and worked fine. Why not hand the model an HTTP client and be done with it?
That skepticism flipped once the engineer started building for business customers.
What big companies saw in it
The engineer points to two practical effects. The first was attention. Having an MCP server briefly became shorthand for “we know how to ride the AI wave,” and that made large companies stop and rethink how they approached AI.
The second was control. Company administrators wanted to decide what employees using ChatGPT could and could not do. With the Gmail connector, for example, an admin could allow read actions and block write actions. The protocol’s vocabulary matched how the people in charge of AI rollouts already thought about permissions.
The rough early implementations
Closer inspection changed their mind again. In that telling, MCP at that point was a Wild West, with every team implementing it their own way. Tool descriptions were so long that they strained context windows. Models were poor at calling tools, which pushed developers to write even longer descriptions.
The protocol was also stateful, which made it hard to run reliably for an audience approaching a billion people.
Why command-line tools gained ground
Command-line interfaces looked like the better answer for a while. They sidestepped the context problem and made it easy to chain steps together. Models had already been trained in POSIX-style environments, so they could grep, sort, awk and pipe commands into a working solution without dumping all the data into their own context.
Code mode shifts the balance
Then came code mode, an idea pioneered by Cloudflare. It later reached research labs, where post-training taught models to write JavaScript rather than bash when composing tools and transforming data. That proved more reliable and used fewer tokens.
It also made MCP easy to fit in. Tools exposed through the protocol simply became functions the code could call, and models got better at exploring and picking the right ones.
Command-line tools lost ground at the same time, for two reasons:
- There was no clean way to keep authentication tokens away from the model. Even inside a sandbox, the CLI needs a readable token, so the model can read it too and do things the CLI never exposed.
- CLIs need Linux, and running a Linux container for a billion people is hard.
What changed in the protocol itself
The engineer says the Model Context Protocol improved a lot in those two years. It can now run over stateless HTTP, supports events, and works with authentication gateways, along with other features that help when AI runs at scale. Even the core agentic building blocks can be expressed as MCP servers.
The summary is that MCP is a case of “worse is better.” It began rough, gathered enough momentum to become popular, and that popularity brought the work needed to make it good.
The lesson about forecasting AI
The engineer also admits it is humbling to look back at how often their own assessment shifted. Each time, the change followed a new development: enterprise admin needs, the weaknesses of early implementations, the rise of command-line tools, and then code mode.
The takeaway is that it remains hard to predict how quickly models will learn things that look out of reach. A limitation that seems fixed in one year, such as weak tool calling, can be gone the next.
What to watch in MCP’s next stretch
For developers choosing between a protocol and a command-line tool, the argument is practical. Admin controls, token handling, and the ability to scale without a Linux container per user all favor the Model Context Protocol today. Whether that holds will depend on how quickly models keep improving at tool use and code.
What looks settled for now may not stay that way. By that account, a long record of reversals suggests treating any firm verdict on AI infrastructure with caution.
Source: frantic.im