Use the Same MCP Server Across AI Clients

Portability is strongest when the server stays close to the MCP standard and treats host-specific features as optional enhancements.

Use core primitives first

Tools, resources, prompts, standard schemas, stdio, and Streamable HTTP have broad support. Add extensions only when both sides negotiate them.

Progressive enhancement

An MCP App should still return useful ordinary content. Enterprise auth should have a documented deployment requirement. Do not make a basic read tool depend on a client-specific UI feature unnecessarily.

Do not bake model-vendor language into tool semantics

A tool named search_repository is portable. A tool whose contract assumes a specific vendor's hidden runtime behavior is harder to reuse.

Maintain a compatibility matrix

Test the server against the actual hosts your users care about. MCP compatibility does not guarantee every optional extension or UX behaves identically.

One server, different adapters

OpenAI may consume it through its Agents SDK, Claude through its MCP surfaces, LangChain through adapters, and editors through configuration files. The server's business capability should remain the same.