Technical Context
I wouldn't argue about preferences here. If an agent on a Codex-like pipeline pulls up context7 by default and spends minutes on it before starting, that's a red flag for me. In AI automation, such minor issues later turn into hours of lost time and strange responses.
I dug into the case, and the logic is very familiar: an external MCP server fetches outdated docs, inflates the context, and prevents the model from doing what it can already do itself—search for fresh information via GPT's built-in search. As a result, the agent doesn't fix the task but confidently starts reinventing the wheel. And that's usually where I hit stop.
The problem isn't MCP as an idea. The problem is a poor default. When unnecessary noise is pre-loaded into the context, the model loses focus, thinks longer, more often chooses the wrong path, and struggles to maintain the solution's architecture.
With context7, this is especially noticeable if the source hasn't been cleaned up in a long time: old snippets, questionable examples, duplicates, and junk from documentation. GPT's own search is often simply better for such tasks now: it retrieves fresh data faster and doesn't clog the context window before the first useful token.
Technically, the solution is boring, and that's its advantage: I would just disable context7 in the MCP config or not load this server at runtime for the code agent. Need GitHub or a highly specialized tool? We keep it. Need a general doc search? I'd first let the model use its built-in search, not an external crutch.
What This Changes for Business and Automation
The winners are teams that need a predictable agent, not magic with surprises. Fewer tokens are wasted on garbage, responses come faster, and the risk of the agent rewriting a service instead of making a targeted fix drops significantly.
The only losers are old pipelines where tools were added on the principle of "the more, the smarter." This isn't AI integration; it's system overload. I've seen many times how an excessive toolset breaks quality more than a weak model does.
At Nahornyi AI Lab, we clean up exactly these areas: where an agent needs real tools and where it's better not to interfere. If your code assistant is slow, produces strange solutions, or burns through the budget for no reason, you can safely disassemble the pipeline and build an AI solution development process without this noise. Sometimes, disabling a single context7 provides more benefit than another model upgrade.