AI Agent That Runs Terminal Commands

Highlights

• Real shell execution

  Bash, zsh, and PowerShell commands run on your machine. The agent does not simulate output — it captures actual stdout and stderr.

• Per-action permission

  Every shell invocation requires explicit user approval. You see the command before it runs and can edit or deny it.

• Output-aware reasoning

  After a command, the agent reasons about the output and decides what to do next. Errors trigger debugging steps; success advances the workflow.

• Cross-platform

  The same agent loop runs on macOS (zsh, bash) and Windows (PowerShell, cmd). Skills written on one platform port to the other where the underlying commands exist.

How shell execution works

When a step requires a command, Lapu AI shows the command, the working directory, and the rationale. You approve or edit it. The command runs locally, and the agent receives the captured output. If the command fails, the agent attempts to debug — reading error messages, trying alternative invocations, or asking you for clarification — instead of pretending to succeed.

What this enables

Anything you do at the shell, the agent can do with your approval. Run a test suite, kick off a build, search a codebase with grep, fetch and parse JSON with curl and jq, manage virtual environments, install packages. Combined with the file tools and desktop automation, the shell is the workhorse for technical workflows.

Frequently asked questions

Can the agent run commands without my approval?

No. Every shell command surfaces a per-action approval prompt before executing. You see the command and can deny or modify it.

What shells are supported?

On macOS, zsh and bash. On Windows, PowerShell and the Command Prompt. The agent picks the appropriate shell based on the platform and the command.

Can it run sudo commands?

Sudo prompts for the user password, which Lapu AI does not provide automatically. You can approve a command and enter your password manually if appropriate.

Does it read environment variables?

Commands inherit the environment of the shell session the agent runs them in. Be mindful of secrets in environment variables when scoping what the agent can run.

Is there a sandbox?

Commands run in your real shell so they have the same effect as if you typed them. The safety layer is the per-action approval prompt, not a sandbox. Plan permissions accordingly.

Related

• AI Agent That Reads and Edits Files

  An AI agent that reads, edits, and writes files on your machine. Lapu AI ships local file tools — read, write, edit, grep — with per-action approvals and a visible audit trail.

• Desktop AI Agent for Developers

  A desktop AI agent for developers. Lapu AI runs locally with shell, grep, file edit, and desktop automation — across macOS and Windows. Permission-based execution.

• Desktop AI Agent for Operations Teams

  A desktop AI agent for ops, IT, and BizOps teams. Lapu AI automates repetitive desktop work across files, shell, and apps — with permission gates and reusable skills.

• Local Desktop AI Agent

  A local-first desktop AI agent for macOS and Windows. Files stay on your machine. Lapu AI ships permission-based local tools with frontier AI through Lapu AI infrastructure.

• Compare against alternatives