MCP Is About to Eat the Agent Framework Market. LangChain Is the First Course.

Hey guys, Mr. Technology here.

Hot take: Model Context Protocol is going to eat the agent framework market in 18 months, and LangChain, CrewAI, and AutoGen are not going to survive it as standalone companies. They are about to become jQuery — a layer that briefly mattered, then got absorbed by the platform, then became a thing you find in legacy codebases and migration guides. The cull is going to be ugly.

1. LlamaIndex's founder already admitted it

In March 2026, Jerry Liu, founder of LlamaIndex — the second-largest agent framework by adoption — published "Why LLM Frameworks Are Being Replaced" and conceded the thesis of his own company. The abstraction the frameworks were charging for is now a commodity protocol. When the founder of the framework publicly tells you the framework business is over, the framework business is over. That is not a competitive take. That is an exit signal.

2. MCP did in 12 months what HTTP took 8 years to do

MCP shipped as an open standard from Anthropic in November 2024. By mid-2026, every major model provider has implemented it on both sides of the wire. OpenAI, Google, Microsoft, Mistral, Cohere, and xAI all ship native MCP clients and servers. The MCP server directory has grown from roughly 1,000 servers in Q1 2025 to over 15,000 by June 2026, with Cloudflare, Stripe, Shopify, Linear, Notion, and Supabase shipping official servers. The protocol is what USB-C is to cables — nobody argues about it, you just use it.

3. The frameworks are pivoting and admitting the play

LangChain's entire 2026 strategy is now LangGraph + MCP adapters. The wrapper around raw model calls is gone — the new value prop is state management and long-running workflows. CrewAI rebranded in Q1 2026 to position itself as an orchestration layer over MCP. AutoGen pivoted to Microsoft 365 integration and is now an enterprise SKU, not a developer framework. SmolAgents has been folded into Hugging Face's MCP tooling. Every framework has two paths: become a state layer over MCP, or become a feature in someone else's platform. None of them are a standalone developer framework in 2027.

4. The economics make it inevitable

A framework makes money by owning the abstraction between you and the model. MCP kills the abstraction for the highest-value parts of that stack. Tool calling, resource access, prompt templates, server discovery — the things frameworks charged the most for — are now a free, open protocol. What is left? State, retries, observability, deployment. That is infrastructure, not framework. Six-figure price points, nine-month procurement cycles, platform engineering as the buyer. The developer-led, self-serve, fifty-grand-to-land motion is dead.

5. The coding agents already won the future

The most important MCP clients in 2026 are not LangChain or CrewAI. They are Claude Code, Cursor, Codex, Gemini CLI, Cline, and Continue. Coding agents were the first wave to treat MCP as a first-class protocol, and they are the wave with the most users and revenue. Every meaningful MCP server built in 2026 is built first for a coding agent, second for a chatbot, third for a LangChain wrapper. The framework is now a third-party integration target, not a primary surface. The center of gravity has moved.

The Take

By Q4 2027, 80% of "agent code" in production is a model call plus a list of MCP servers. The framework layer is reduced to: pick a state engine, pick a deploy target, write some glue. The frameworks that survive pivoted in 2026. The rest are going to spend 2027 doing the math on a down round.

I have been on this beat for 18 months. I am calling this one now: MCP eats the agent framework market in 18 months, and the frameworks will thank you for it. The alternative is selling infrastructure to platform engineering for a fraction of what they were getting from developers. The protocol won. The frameworks lost. The only question is which of them admit it first.

— Mr. Technology