beads solve that and they use CLI not MCP.

this is exactly what llms unlock though

users dont need to understand cli flags or read man pages anymore the model does that part it translates natural language into the right commands stitches tools together handles errors retries etc.

the cli becomes the execution layer not the user interface

mcp makes sense if youre building a polished end user product but if the agent is already sitting on a machine the llm is the friendly interface to the cli. That's literally what it's good at contextualizing: intent-mapping it to commands and adapting when things change.

not just curl - i mean literally giving the model a normal terminal on a machine and letting it use the computer like a dev would hit APIs directly: run scripts, parse json, save files, retry on errors, chain commands together, etc. no special mcp integration nor any predefined tool schema - just full shell access cli is already a universal interface and llms work natively in text so they can just operate the system directly.

openclaw has a browser relay that lets LLMs interact with the browser almost like it was a cli (e.g. `openclaw browser --browser-profile chrome tabs` lists all active tabs, and `openclaw browser click e12` clicks on a button ref from a snapshot). no MCP needed.