MCP skepticism grows as real-world usage hits friction
A thread titled 'MCP was always a bad idea?' generated substantial discussion. The core complaints: credential exposure to the model, lack of shell access in some agent contexts, and the overhead of wrapping every tool. Commenters pushed back noting that the alternatives (CLI-only modes) have their own unsolved problems around credential hiding. One builder described a tool wrapper that captures and smartly truncates output to save tokens, essentially reinventing parts of what MCP was supposed to handle.
The key tension here is between the simplicity MCP promises in theory and the rough edges that appear in practice. The thread isn't a consensus takedown of MCP, but it reflects a moment where early adopters are hitting real limitations and starting to question architectural choices made early in the agentic wave. Google's AX open agentic orchestrator also launched to notable skepticism, with multiple commenters saying they 'don't pay attention to news like this from Google' and citing Google's track record of abandoning open source projects.
The two threads together suggest the agentic tooling layer is still genuinely unsettled. Builders are shipping workarounds, questioning standards, and treating even Google's contributions with suspicion.
So what?
If you're building on MCP or planning to, the credential exposure problem is the most concrete risk to address now. The broader pattern is that the agentic tooling layer has no settled winner yet, which is an opportunity for founders building in this space but also a risk for anyone who bet heavily on a specific standard. Treat the current protocols as provisional and keep your abstraction layer thin.