This week, Reddit's developer community surfaced an awkward reality: teams building enterprise-grade "AI computer-use" systems can hardly bypass OpenAI's Codex—its capability comes from a protocol built by a small company called Sky, and open-source alternatives still lag behind. This matters because it exposes the real bottleneck in Agent deployment: not that models are too weak, but that critical modules are locked behind a single vendor.
What this is
"Computer Use" refers to AI operating computers like a human: moving the mouse, clicking buttons, filling forms. MCP (Model Context Protocol) is an open standard proposed by Anthropic in 2024 that lets AI models talk to external tools and data sources—the "USB port of the AI world."
Codex's Computer Use MCP was built by Sky and is widely regarded by developers as the best-performing implementation. But this also means two things: the underlying capability depends on a single vendor, and the code isn't open—the original post bluntly states, "open source MCPs don't come as close as Codex."
How the industry sees it
Supporters argue the division of labor—"vertical small players build critical modules, big players build models"—is reasonable: specialists do specialist jobs, and market efficiency is highest. But critics are sharper: Computer Use is a core scenario for enterprise RPA (Robotic Process Automation), and once locked to a single vendor with no source code released, data compliance and long-term costs both become uncontrollable.
A more practical risk: if companies like Sky are tied to a single buyer, the open-source community's catch-up effort may fall behind permanently. Some developers under the post already plan to "build one from scratch," but the engineering effort likely far exceeds what any individual expects—and that is precisely the real threshold for enterprise-grade AI tools.
What this means for regular people
For enterprise IT: if you want AI to operate legacy internal systems for employees, you'll most likely have to choose within the OpenAI ecosystem or invest in in-house development—both options are expensive.
For individual careers: once Computer Use matures, roles built on repetitive form-filling and cross-software data shuttling will be hit first, but we're still in the "works but doesn't work well" stage.
For consumer markets: everyday users won't feel much in the short term; real adoption depends on whether enterprise-side software vendors are willing to plug into this new protocol.