Mel – the agentic terminal
The AI terminal for teams.
Learn · natural language to shell commands
Natural language to shell commands: how it works
Updated
Natural language to shell commands means you describe what you want in plain English, such as "find every file over 100 MB in this folder", and an AI turns it into the command, find . -size +100M, or runs the steps for you. The safe versions show you the command before anything runs.
Mel – the agentic terminal does this in two ways: an agent that works out and runs the commands under your approvals, and AI completion of the command you are typing. Shell syntax is exact and unforgiving. Natural language closes the gap between knowing what you want and remembering which flag does it.
The three kinds of tool
Translate. You type a sentence, the tool prints one command, and you decide whether to run it. Good for one-off commands you would otherwise look up.
Complete. You start typing a command and the AI suggests the rest, like autocomplete that knows the flags.
Do. You describe an outcome ("upgrade the dependencies and make sure the tests still pass") and an agent runs as many commands as it takes, reads the output and keeps going. This is where the biggest time savings are, and where approvals matter most.
The risks, and how to handle them
- A wrong command can delete things. Read anything that deletes, moves, overwrites or uses
sudobefore it runs. - Plain English is vague. "Clean up the folder" can mean many things. Name the folder, the files and the result.
- Platforms differ. The same request needs different commands in bash, zsh and Command Prompt. A good tool knows which shell it is writing for.
- Approvals are the safety net. Choose which commands need your yes, and keep that on for anything that cannot be undone.
How Mel does it
- Ask the agent. Describe the task in plain English. Mel's agent works out the commands, runs them in your project and reads the results. Each command it wants to run can need your approval, and you see exactly what will run.
- Commands still run as commands. In an agent conversation, a line that is already a terminal command runs in the shell. A leading
!forces the shell. - AI completion while you type. When your history and files have no match, Mel can ask the AI to finish the command you are typing, shown as grey ghost text you can take or ignore.
- Errors explained. Attach a failed command's block and ask what went wrong.
- Destructive commands always ask. A command that cannot be undone needs a person, whatever the approval setting.
Examples
| You type | A command it could become |
|---|---|
| find files over 100 MB here | find . -size +100M |
| which process is using port 3000 | lsof -i :3000 on Mac and Linux, netstat -ano on Windows |
| show what changed since the last commit | git diff HEAD |
| count the lines in every Python file | find . -name "*.py" -exec wc -l {} + |
FAQ
What does natural language to shell commands mean? Typing what you want in plain English and getting the matching command, or having an agent run the steps for you.
Is it safe to let AI run shell commands? It is when you can see each command first and control which ones run. Mel's configurable approvals decide which commands need your yes.
Does it work on Windows? Yes. Mel runs on Windows 10 and 11, where commands run through Command Prompt by default.
Do I need to pay for it? No. Mel is free with your own API key. Pro and Max add hosted credits.
Mel – the agentic terminal · openmel.dev. A multi-agent, multi-model AI terminal for Mac, Windows, Linux and VS Code. Free with your own API key.