MCP is a fad

Original Article ↗ Hacker News Discussion ↗

Perceived role and value of MCP

  • Many see MCP as a small, boring integration layer: a standardized way for agents to call tools and resources, especially across different AI clients (Claude, ChatGPT, IDEs).
  • Supporters emphasize interoperability and “write once, use in many agents,” likening it to LSP or USB for AI tools rather than just a local scripting mechanism.
  • Critics argue the article focuses too much on local filesystem use and misses broader agent-to-service scenarios, including async and long‑running operations, generative UI, and SaaS integration.

Comparison with Skills, CLIs, and OpenAPI/HTTP

  • Several argue Claude Skills (markdown + front matter) are simpler and often sufficient; some think useful commands/docs should live in human‑oriented files and be “taught” to the AI, not moved into AI‑specific configs.
  • A recurring claim: almost everything MCP does can be done with CLIs plus a shell and tools like just/make, or via existing HTTP/OpenAPI APIs.
  • Others counter that MCP’s structured tool schemas, resource mounting, and stateful handles provide more predictable, testable flows than agents dynamically generating glue code or scripts.

Security, lifecycle, and operational concerns

  • Strong skepticism around security: MCP is seen as an easy data‑exfiltration vector, especially if people casually add third‑party servers.
  • Some argue MCP is “just the protocol” and security is an implementation concern; others reply that in practice bad ops and weak curation are common, so the risk is real.
  • Process lifetime and resource usage are highlighted: one‑process‑per‑server can lead to many heavy apps idling, especially with multiple coding agents.
  • There is debate over whether MCP meaningfully improves sandboxing vs. running tools in containers/VMs or via safer gateways.

Interoperability, auth, and enterprise use

  • Pro‑MCP voices stress OAuth-based auth, auditability, permission prompts, and approval workflows as key for exposing enterprise SaaS/APIs to agents (including web UIs like ChatGPT/Claude).
  • Others ask why not just expose OpenAPI specs and treat AI calls as normal RPC, avoiding a parallel ecosystem.

Broader AI‑for‑coding and “fad” discourse

  • Thread widens into whether AI coding and tool‑calling are fads: some report disastrous experiences and see LLMs as wasteful slop generators; others say latest models, used well, dramatically speed up meaningful work.
  • There’s tension between those prioritizing code quality and domain expertise vs. those emphasizing speed, delegation, and acceptance of “good enough” outputs.