Monitors interact with your system via execute_command. This tool is the bridge between the AI’s logic and your terminal environment.

execute_command

Executes a terminal command within the block’s current working directory.

• Parameters: command (string), timeout (number, default: 120s)
• Approval: Dynamic (Safe commands auto-execute; others require human approval)
• Timeout: 120 seconds default; configurable per-block in config.json

Safety first

The memoryblock shell is sandboxed by default. Commands are categorized as Safe or Potentially Dangerous.

🟢 Safe Commands (Zero Approval)

The system knows these commands are read-only or harmless. They always execute instantly:

• Listing: ls, find, which, pwd, echo
• Reading: cat, head, tail, wc, grep
• Environment: node --version, bun --version, pnpm --version
• Git (Read): git status, git log, git diff, git branch
• Dev-Run: tsc --noEmit, npx eslint, pnpm lint, pnpm build, pnpm test

🟡 Approval Required

Any command not on the SAFE_PREFIXES list will trigger an approval request. The monitor will pause, and you must explicitly allow or deny the action on your active channel.

Examples of commands requiring approval:

• rm -rf
• git push
• npm install
• curl / wget

Permission Scopes

The scope (block, workspace, or system) restricts where a command can explore your files. By default, the block scope is isolated to its own directory.

To elevate a block’s permissions, use the CLI:

mblk permissions my-block --allow-shell --scope workspace

Tips for Token Efficiency

Avoid running commands that produce massive output. The system caps shell output at 50,000 characters. For large outputs, use grep or tail to filter the data before it reaches the monitor.