You connect Claude Code to your database through the Postgres MCP server every list recommends, assuming read-only means read-only. That server is Anthropic’s own reference implementation, it’s archived, and it has a documented read-only bypass. The same lists rank a “1.7k-star” MongoDB integration above an official server that ships weekly.
We ran the same maintenance check from our first curation guide against the database and backend category. Two names got cut for being stale or unsafe; five made it through.
What the full guide covers:
The commands, our own test notes, and the troubleshooting steps are in the full guide, which we keep up to date:
Top comments (2)
Good call flagging that archived reference servers still run without patches. npm has the same problem: of the 30 most-downloaded MCP server packages I checked (the run is written up here), 6 are deprecated yet still get around 455K weekly downloads, with READMEs still saying
npx -y. Several quick-start lines here are unpinned too (@latest, barenpx -y), so the tool surface can change on any restart. Would you consider pinning versions in the install commands?Cảm ơn bài viết tổng hợp rất kịp thời — việc Anthropic archive reference implementation kèm read-only bypass là red flag lớn cho ai đang dựng production MCP server. Mấy cái list "awesome-mcp" trên GitHub vẫn còn để server này ở top, dẫn dắt dev mới lầm tưởng nó production-ready.
Hai điểm mình thấy critical khi eval các alternative:
Connection pooling & transaction boundary — Hầu hết server MCP database đều mở connection per request, không có pool sẵn. Dưới load thực tế (cùng lúc 50+ agent call tool) sẽ exhaust DB connection limit. Cần check xem server nào support pgBouncer / built-in pool, hoặc ít nhất expose config để inject pooler ngoài.
Schema introspection caching — Tool
list_tables/describe_tablegọiinformation_schemamỗi lần. Với schema vài trăm table, latency rõ rệt. Các server mature (như Postgres-MCP mới hơn) đã có TTL cache layer cho metadata — tiết kiệm round-trip đáng kể.Mình đang test server thứ 3 trong danh sách (viết bằng Go, có embedded pooler) — initial benchmark cho thấy p95 latency ổn định ~12ms vs 45ms của reference implementation khi chạy concurrent 100 requests. Trade-off là config hơi verbose, cần define resource mapping thủ công thay vì auto-discover.
Ai đã chạy load test trên các option còn lại? Share numbers cho anh em tham khảo PS: the tool I meant is on labagent .tech