Three of these loadouts build Model Context Protocol servers with the official SDKs, in TypeScript, Go and Python: stdout belongs to the protocol, inputs are validated, failures come back as tool errors and every tool is tested through a real client. One moves a server to the stateless 2026-07-28 revision and probes it. One packages MCP servers and skills as Claude Code and Antigravity plugins. The last is for installing other people's servers, skills and plugins: pin them, prove the pin and read every file first.
Every one installs for Claude Code, Codex CLI, Cursor, Gemini CLI, Antigravity CLI and GitHub Copilot unless its card says otherwise.
Rules and an add-a-tool skill for building a Model Context Protocol server in TypeScript that agents can actually use well — clear tools, safe inputs, no stdout noise.
Build one plugin for Claude Code and Antigravity from a single source tree, pin its MCP servers, and pass claude plugin validate --strict before anyone installs.
Build MCP servers in Go with the official go-sdk: typed tools with jsonschema tags, tool errors not panics, stdout kept for the protocol, in-memory client tests.
Build MCP servers on the official Python SDK v2: MCPServer, Annotated schemas, structured output, ToolError for fixable failures and in-memory Client tests.
Move an MCP server to the stateless 2026-07-28 revision on the official SDKs: no sessions or handshake, server/discover, input_required, then a raw stdio probe.
Treat agent skills, plugins and MCP servers as dependencies: pin by SHA or exact version, prove HEAD matches the pin, scan every file, and get a human yes.
Looking for something narrower? Search within MCP or filter everything.