Shell — A Real Linux Environment

The agent runs real shell commands in a Linux environment (Alpine-based, via proot) right on your phone. This isn't a simulation — it's a working shell with a filesystem, package manager, and scripting.

What it can do

• Run commands and scripts — one-liners, pipelines, shell scripts, cron-style automation.
• Manage the workspace — create, edit, move, organize, and search files.
• Install packages — apk packages, Python packages, npm tools, CLIs.
• Build and test — compile code, run test suites, process data, convert files.
• Network — curl APIs, download files, probe endpoints.

Environment rules (Alpine)

The shell is Alpine Linux, which is minimal by design:

1. First command in a fresh environment: run apk update before installing anything.
2. Python packages: prefer pure-Python or wheels-only packages — there's no compiler toolchain, so anything needing gcc won't build.
3. Long-running commands: start them detached (background) and poll the output file instead of blocking the chat.

Secrets are redacted automatically

API keys and tokens in the environment are never shown to the model — tool results replace them with [REDACTED_SECRET]. The agent can use secrets (e.g. call an API) without ever seeing their values. This is enforced centrally, not per-tool.

Practical examples

• "Download this CSV and tell me the top 10 rows by revenue" — the agent fetches, parses, and summarizes.
• "Clone this repo and run its tests" — full dev workflow on your phone.
• "Every morning, check this site and summarize changes" — combine with scheduled tasks.

What's next

• MCP — plug in external tool servers
• Skills — save reusable shell workflows as skills