Code Execution

Agent Nine can run shell commands on your machine (Code mode) or in a cloud sandbox (Chat mode).

Code Mode (Local)

In Code mode, commands run directly on your machine with full access to your project files, tools, and environment.

Bash (Linux / macOS)

Run the tests

Check if the project compiles

Install the dependencies and start the dev server

PowerShell (Windows)

Get-ChildItem -Recurse *.ts | Measure-Object

Run npm test

Same security protections apply for both shells.

Chat Mode (Cloud Sandbox)

In Chat mode, code runs in an isolated cloud sandbox:

• Separate container for each session
• Pre-installed runtimes: Python, Node.js, Go, Rust, Java
• No access to your local files
• Safe for experimentation — nothing affects your machine

Write a Python script that generates a Fibonacci sequence

Create a React component and show me the result

Security

Agent Nine protects you from dangerous commands:

Category	Examples	Behavior
Destructive	rm -rf /, git reset --hard, DROP TABLE	Blocked with warning
Irreversible	git push --force, npm publish	Requires explicit approval
Safe / read-only	git status, ls, cat, grep	Auto-approved

Dangerous patterns are detected even when disguised:

• Encoded commands (base64, hex)
• Commands inside scripts
• Piped sequences that end in destructive operations

Timeouts

	Duration
Default	2 minutes
Maximum	10 minutes
Background	No timeout (you'll be notified when done)

Long-running commands can be started in the background — Agent Nine will continue chatting and notify you when the command finishes.

Permissions

Mode	Behavior	Best for
Default	Asks before running every command	Beginners, sensitive projects
Session-approved	Approve once, auto-run similar commands	Regular development
Auto	Auto-run everything except destructive commands	Power users

Read-only commands (git status, ls, cat, grep, find) are always auto-approved regardless of permission mode.

Tips

• Be specific — "run npm test in the backend folder" is better than "test it"
• Chain tasks — "install dependencies, run tests, then fix any failures"
• Background execution — for long builds or deployments, Agent Nine runs them in the background
• Review output — Agent Nine shows you the full command output and explains errors