Skip to main content
NNextGen AI Learn
All news
EcosystemMCPagentstooling

MCP one year on — what stuck, what didn't

Model Context Protocol crossed 1k public servers in early 2026. The "USB-C for agents" pitch is mostly working. Where the rough edges remain.

The good

  • Connector ecosystem. Slack, Notion, Linear, Postgres, Cloudflare, Sentry, Stripe — most major SaaS APIs have community or vendor MCP servers. Plug Claude Desktop into your stack in minutes.
  • Cross-client compatibility. Cursor, Zed, Continue, Claude Desktop, several agent frameworks all speak MCP. A tool you write once works everywhere.
  • Discovery is improving. The MCP server registries (community-run + Anthropic's) make finding servers easier than it was at launch.

The not-so-good

  • Auth is still gnarly. OAuth flows for MCP servers vary. Local-only stdio transport works fine; remote auth is messier.
  • Permissions feel coarse. MCP doesn't natively express "this tool is read-only" vs "this is destructive". Clients build their own scopes.
  • Discovery + ranking. With 1k+ servers, finding the right connector is harder than it should be.

What to do

If you build internal tools — strongly consider exposing them as MCP. Future-you will thank you when you switch agent frameworks. If you build product-specific business logic — inline tool definitions stay simpler.

Want the deep dive?

The lessons that ground this news in mechanics — not opinion.

Browse courses