ai-tool-integrationlisted

**CONSUMES**: User tool integration task + target API/CLI/service description + optional existing MCP server code or tool configs **PRODUCES**: Applied tool integration judgment rules + MCP server code + CLI wrapping decisions + API integration configs + schema reviews + permission models + test plans + tool documentation # AI Tool Integration Capability Pack **Version**: 0.1.0 **Compatibility**: Claude Code (Phase 1); Codex / Cursor / Gemini in Phase 3 **License**: Apache 2.0 --- ## What This Pack Does AI agents build MCP servers by copying tutorial scaffolds. They wrap every CLI tool as an MCP server even when code execution would cut a 150,000-token Drive-to-Salesforce workflow to 2,000 tokens (98.7% reduction, Anthropic engineering 2026-06-13). They skip schema constraints, letting models hallucinate parameter values. They put read and write operations on the same server with the same IAM role. They never test with MCP Inspector. They write tool descriptions that say "use this tool" without examples, error guidance, or "when NOT to use." They teach the deprecated 4-arg `server.tool()` API instead of the current `server.registerTool()` (SDK v1.29.0). They ignore tool poisoning (CVE-2025-54136 / CVE-2025-54135), the OWASP LLM01 #1 risk. This pack embeds the judgment rules that tool integration engineers apply automatically -- rules from the MCP TypeScript SDK (v1.29.0), the MCP 2025-06-18 spec, Anthropic's measured tool-use research, OAuth 2.1 (RFC 9728 / RFC 8707), a