12 MCP patterns for production agents: why wrapping one tool per API endpoint fails
bibryam · x · 2026-09-18
A new article lays out 12 MCP patterns behind production agents, arguing that the easiest approach — wrapping an API with one tool per endpoint — is also one of the fastest ways to build a poor agent interface.
Key points:
- Production MCP design is not just protocol plumbing; it's designing an agent-facing product surface: which capabilities are visible, how users provide input, where credentials live, what reaches the model's context, and how the integration is packaged
- The twelve patterns are organized into five areas, starting with the tool surface
- Pattern: Remote-first server — local stdio servers only work when the agent can spawn a process on the same machine, excluding browser, mobile, and many cloud agents; production should prefer remotely hosted servers with standard transport and auth, keeping local deployment for dev, offline, or machine-local capabilities
More from coding & agent
- NoSpoon microdrama agent writes, edits and casts short dramas autonomously in ~10 minutes — Kyrannio · 2026-09-18
- Alchemy PR adds Neon backend resources, Effect runtimes and 13 website constructors — samgoodwin89 · 2026-09-18
- After 19 full runs, this dev ditched 100% as the bar for coding-agent evals — oliver-zehentleitner · 2026-09-18
- Showly launches: turn your coding agent's output into shareable web pages — HeyAmit_ · 2026-09-18
- Pydantic AI agents now run on TypeSafe's Jev with per-field confidence outputs — Paimaamu · 2026-09-18
- Demo: GPT-6 Astra edits video in DaVinci via MCP at impressive speed — toolstelegraph · 2026-09-18