[Official FAQ] Corrections and follow-up questions welcome.

Tools let models read files, query systems, or run commands. MCP-style servers standardize tool wiring for compatible clients.

Security first:

• Tools are privileges—sandbox, path allowlists, and deny dangerous commands.
• Begin with read-only tools; add writes with human confirmation.
• Prompt injection can trick local agents too—add guardrails.

Do not expose shell-capable agents on the public internet. Share hardened setups (no secrets).