Conventions, community best practices research (Sept 2025 - March 2026), and prioritized improvement backlog for Claude Code workflows. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1.9 KiB
MCP Servers
Why This Matters
MCP (Model Context Protocol) servers extend Claude Code with external tool integrations. Community consensus: "If you're not using MCPs, you're driving a Ferrari in first gear." However, each registered MCP consumes context tokens even when unused, so be selective.
Current State
- No MCP servers configured
Must-Have (Community Consensus)
Context7
The single most-recommended MCP server. Provides real-time, version-specific documentation for any library/framework. Solves the knowledge-cutoff problem — Claude gets accurate docs for the exact version you're using.
Particularly valuable for:
- Helm chart values schemas (change between versions)
- Kubernetes API changes
- Any rapidly-evolving ecosystem
Sequential Thinking
Structured problem-solving for complex architectural decisions. Described as "like having a senior architect who thinks before coding." Useful for planning phases.
Worth Considering
| MCP Server | Purpose | Relevance |
|---|---|---|
| Playwright | Browser automation, UI testing, screenshot capture | Useful if doing web projects |
| GitHub/Gitea | PR/issue management from terminal | Would need a Gitea-compatible MCP |
| Supabase | Database ops, migrations, SQL queries | If using Supabase |
Key Insights
- Start with 2-3 MCPs, not all of them. Each one adds to context overhead.
- MCP Tool Search (lazy loading) reduces context usage by up to 95% — register many, only load what's needed.
- Prefer Skills over MCPs for stateless tools. MCPs are best for stateful environments (browser sessions, database connections, auth boundaries). Stateless CLI wrappers are better as Skills.
Migration Pattern
The community is trending toward moving stateless tools from MCPs to simple CLIs documented in SKILL.md files. The "wrapper pattern" uses commands as thin entry points (~93 tokens) + skills as full implementations loaded on-demand, reducing startup context by ~64%.