Tools (Concept)

Tools let the agent perform actions (browse, run commands, fetch web content, automate workflows). The key is to keep boundaries explicit so “can do” never silently becomes “did do”.

In OpenClaw, three layers matter:

1. Sandboxing (where code runs; host vs container isolation)
2. Tool policy (what tools are allowed/denied)
3. Approvals (who must confirm side effects)

This separation makes security posture understandable and debuggable.

Next steps:

• Tool docs: /tools
• Sandbox
• Approvals
• Sandbox vs tool policy vs elevated