MCP's 18-Month Surge to Production Readiness

MCP evolved from a local-only spec with basic tools 18 months ago—mostly Claude-generated SDKs—to a standard with remote capabilities, centralized auth, elicitation, tasks, and experimental apps. Ecosystem hit 110 million monthly downloads, outpacing React (which took double the time), fueled by integrations in OpenAI's agent SDK, Google's ADK, LangChain, and thousands of frameworks. Servers range from toys (WhatsApp, Blender) to SaaS (Linear, Slack, Notion), but most connect internal company systems to agents. 2025 focused on coding agents (ideal: local, verifiable, UI-displayable); 2026 shifts to general knowledge worker agents (e.g., financial analysis, marketing) needing multi-SaaS/shared drive connectivity. MCP servers ship portable apps/tools with UI for humans and tool interfaces for models, deployable to cloud/ChatGPT/VS Code/Cursor without client-side rendering.

Connectivity Stack: Skills + CLI + MCP for Versatile Agents

No single solution fits all—reject one-size-fits-all claims. Use three layers:

  • Skills: Simple, reusable domain files (minor platform variances) for specific capabilities.
  • CLI/Computer Use: Auto-discoverable for local coding agents; excels with pre-trained tools (GitHub CLI, Git) in sandboxed code environments.
  • MCP: For rich semantics (resources, long-running tasks UI), platform independence, auth/governance/policies, and experiments (apps, skills over MCP). Top agents in 2026 seamlessly blend all three, e.g., CLI for local verification, MCP for enterprise decoupling.

Client Improvements: Slash Context and Latency

Agents fail from poor harnesses. Fix with:

  • Progressive Discovery: Defer tool loading via tool search (Anthropic API or custom)—model requests tools on-demand. Claude Code saw massive context reduction post-implementation (left: before, right: after in demo).
  • Programmatic Tool Calling (Code Mode): Model writes/executes scripts (V8 isolate, Monty, Lua) composing tools/APIs/CLIs, avoiding serial inference (latency/token waste). Leverage MCP structured outputs for type info; fallback: cheap model extraction. Example: Single call filters JSON vs. multiple tool hops. Compile with executables for efficiency.

Server Best Practices and MCP Roadmap

Ditch 1:1 REST-to-MCP wrappers (produces "horrible" tools). Design for agents/humans: Provide execution envs (e.g., Cloudflare MCP) for model orchestration. Exploit MCP uniques: apps, tasks, elicitations, skills over MCP (ship updated domain knowledge sans registries).

Roadmap targets scale:

  • Core: Stateless transport (Google proposal, June) for hyperscaler deployment (Cloud Run/K8s); async tasks for agent-agent comms; TS/Python SDK v2 (lessons learned; fast-mcp outperforms current Python).
  • Integrations: Cross-app access (single IdP login: Google/Okta); server discovery (well-known URLs for crawlers/agents).
  • Extensions: Skills over MCP (large servers ship main knowledge); client-specific (e.g., apps for web UIs only).

Join open community (Discord/issues/foundation) for feedback—MCP positions for full connectivity.