Six under-documented patterns for production MCP servers that most tutorials skip: a tool name registry (single source of truth to prevent drift), compact/full mode flags per tool to manage token budgets, progress notifications via `_meta.progressToken` for long-running operations, tool registration gating based on context/permissions, explicit cache invalidation tools, and a roots integration priority chain (env → client roots → cwd). Each pattern is illustrated with real production examples from Sentry, GitHub, Cloudflare, and others, with concrete code snippets. The post argues MCP servers are converging on standard product interface concerns: permissions, observability, admin endpoints, and tenancy.

12m read timeFrom evilmartians.com
Post cover image
Table of contents
Tool surface hygieneToken economyLifecycleWorkspace integrationWhat this leaves you with

Sort: