Forget MCP, Bash Is All You Need

7 days ago

LLMs require tools to interact with the real world, with MCP (Model Context Protocol) being the industry's current solution.
MCP is criticized for reinventing operating system features like process invocation and IPC, which Unix and Linux already handle efficiently.
Key issues with MCP include high token costs for tool definitions and the inefficiency of requiring model inference for each tool call.
Successful AI agents like Claude Code, Cowork, and OpenClaw bypass MCP, leveraging the operating system's native capabilities through bash.
Linux provides essential runtime features for AI agents, such as process management, IPC, and composition, making it an ideal foundation.
Bash serves as a powerful interface to the OS, enabling deterministic, fast, and free script execution without continuous model inference.
MCP's primary advantage is streamlined authentication in chat UIs, but CLI tools already handle OAuth effectively.
The future of AI agents lies in orchestrating the OS directly, utilizing its robust, battle-tested infrastructure rather than building atop it.

Hasty Briefsbeta