MCP: Connectivity Protocol for 2026 Production Agents
MCP hit 110M monthly downloads in 18 months—faster than React. For 2026 agents tackling knowledge work, combine skills, CLIs, and MCP with progressive discovery and programmatic tool calling to enable efficient, scalable connectivity across SaaS apps.
MCP Delivers Standardized Agent Connectivity with UIs and Tools
MCP (Model Context Protocol) lets agents ship full interfaces—served via MCP servers deployable to cloud, ChatGPT, VS Code, or Cursor—without plugins or client-side rendering. Servers provide rich semantics for UI rendering, long-running tasks, resources, authorization, and governance, enabling platform-independent decoupling. Agents interact human-like via UIs while models use tools, supporting experiments like MCP applications. In 18 months, MCP grew from a local-only spec (mostly Claude-written) to 110M monthly downloads—half React's time—powering 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 enterprise systems to agents privately.
2026 Agents Need a Multi-Layer Connectivity Stack
Shift from 2024 demos and 2025 coding agents (local, verifiable via compiler/UI) to general knowledge-worker agents for finance, marketing—requiring SaaS/shared drive access. No single solution (computer use, CLIs, MCP) fits; use all:
- Skills: Domain knowledge in simple, reusable files (minor platform differences).
- CLIs: Auto-discoverable for local/sandboxed coding (GitHub/Git, pre-trained); ideal for bash discoverability.
- MCP: For rich semantics, UIs, tasks, elicitation, enterprise features (auth/policies); excels sans sandbox.
Production agents seamlessly compose them. Current agents lag, needing better harnesses.
Client-Side: Progressive Discovery and Programmatic Tool Calling
Avoid dumping all tools into context (causes bloat). Implement progressive discovery: Use tool search (Anthropic API or custom) to load MCP tools on-demand via a 'tool loading' tool. Claude Code saw massive context reduction post-implementation.
Replace serial tool calls (latency-heavy inference orchestration) with programmatic tool calling (code mode): Give models an execution env (V8 isolate, Monty, Lua) to script compositions. MCP's structured outputs provide return types for typing/filtering. Example: One call filters JSON vs. two sequential. Fallback: Prompt cheap model for structured extraction. Compose with CLIs/APIs/executables too—mimics hardcoded bash scripting but generalized.
Server-Side: Design for Agents, Leverage MCP Semantics
Ditch 1:1 REST-to-MCP wrappers (produces poor tools). Design like human/agent interaction: Provide execution envs (e.g., Cloudflare MCP server) for server-side scripting. Ship MCP apps, skills-over-MCP (updated guidance w/o registries), tasks, elicitations. Roadmap:
- Core: Stateless transport (Google proposal, June) for hyperscaler scaling (Cloud Run/K8s); async tasks (agent-to-agent comms).
- SDKs: TypeScript/Python v2 (lessons learned; fastMCP outperforms current Python).
- Enterprise: Cross-app access (single IdP login, Okta/Google); server discovery (well-known URLs for crawlers/agents).
- Extensions: Skills-over-MCP, web-only (e.g., apps for HTML).
Join open community (Discord/issues) for feedback. 2026: Full connectivity ships agent UIs dynamically.