Tool Calling

Deep workflow for LLM tool/function calling—schema design, validation, permissions, errors, idempotency, testing, and safe orchestration with agents. Use whe...

This skill is a safety-and-design playbook, not executable code—so installing it has minimal direct risk. Before you rely on it in production: (1) ensure any real tool implementations enforce server-side schema validation, authorization, and output filtering as the guide recommends (do not trust the model); (2) avoid giving the model direct filesystem, database, or credential access—use validated server-side adapters and least-privilege principals; (3) if you will allow autonomous agent invocation, pair that with strict allowlists, rate limits, and human approval for dangerous ops; (4) treat the checklist as a design spec to implement securely rather than as an instruction to expose sensitive systems to the model. If you want extra assurance, ask for the concrete tool implementation details you plan to use and have them reviewed for the specific risks (credential handling, logging, and sandboxing).

Type: OpenClaw Skill Name: tool-calling Version: 1.0.0 The skill bundle contains only documentation and procedural instructions (SKILL.md) for an AI agent to guide users through designing secure tool-calling interfaces. It explicitly advocates for security best practices such as least privilege, input validation, and avoiding dangerous anti-patterns like exposing raw shell or SQL access.