# AI Interoperability The principle that AI skills, agents, and context should work transparently across models, providers, and platforms. Switching from Claude to GPT to Gemini should not break your setup. Switching from Claude Code to Cursor should not require rewriting everything. The investment you make in your AI infrastructure should survive vendor changes. ## Dimensions of interoperability - **Model interoperability**: skills and agents work regardless of which LLM runs them (Claude, GPT-4, Gemini, Mistral, local models via [[Ollama]]) - **Provider interoperability**: switching API providers (Anthropic, OpenAI, Google) is transparent; no provider-specific assumptions baked in - **Platform interoperability**: skills written in one AI tool ([[Claude Code]], [[Cursor.com]], [[Cline]], [[Windsurf]]) work in others ([[AI Skill Portability]]) - **Machine interoperability**: your setup works identically on all your machines, not just the one where you built it - **Team interoperability**: skills and agents work for every developer on the team, not just the author - **Org interoperability**: context management scales from individual to team to enterprise without breaking ([[Context Inheritance]]) ## Why interoperability matters The AI landscape moves fast. Today's best model is tomorrow's second choice. Today's best tool might not exist next year. If your skills, agents, and context are tightly coupled to one model, one provider, or one platform, you're building on sand. Organizations investing in [[Enterprise Context Management (ECM)]] need interoperability guarantees. A company can't standardize on AI infrastructure that locks them to a single vendor. ## What breaks interoperability - **Model-specific prompt tricks**: techniques that only work on one model family - **Provider-specific API calls**: hardcoded to one provider's tool naming or response format - **Platform-specific syntax**: SKILL.md frontmatter that only one tool understands - **Hardcoded paths**: skills that assume specific file locations, folder structures, or OS conventions (see [[AI Skill Resilience]]) - **Capability assumptions**: relying on features only one model/tool supports (e.g., extended thinking, specific tool names) ## Toward interoperability - [[Model Context Protocol (MCP)]] provides a standard for tool integration across platforms - [[Context-as-Code]] (CLAUDE.md, AGENTS.md) uses plain Markdown, readable by any tool - [[AGENTS.md (File Convention)]] is emerging as a cross-tool agent definition standard - Skill bodies written as tool-agnostic instructions (describe what to do, not which tool to call) - Abstraction layers that map generic operations to platform-specific implementations ## The tradeoff Full interoperability means giving up platform-specific optimizations. A skill that uses Claude Code's specific tool names will be more precise on Claude Code but break everywhere else. The pragmatic approach: keep the core intent portable, wrap platform-specific details in thin adapters. ## References - ## Related - [[AI Skill Portability]] - [[AI Agent Portability]] - [[AI Skill Resilience]] - [[AI Skill Distribution]] - [[AI Agent Distribution]] - [[Model Context Protocol (MCP)]] - [[AGENTS.md (File Convention)]] - [[Context-as-Code]] - [[Context Inheritance]] - [[Enterprise Context Management (ECM)]]