A look at the new, open-source Model Context Protocol for connecting LLMs to data sources and why its success depends on overcoming middleware challenges
There's a long history of “middleware” in our industry. Everyone wants it. There's always a hot one, but it rarely makes to the finish line and often disappoints. Bluesky: @folletto . X: @stevesi and @stevesi Bluesky: Erin Casali / @folletto : Will the AI hype give enough boost to these kinds of middleware integration to then create better flow of data across systems? — Because honestly the “Small Pieces Loosely Joined” approach was nice, but got stuck by massive tech companies protecting their shareholder value. — Next round? [embedded post] X: Steven Sinofsky / @stevesi : Permalink with expanded background and history of “middleware”. https://hardcoresoftware.learningbyship ping.com/ ... Steven Sinofsky / @stevesi : Quick MCP thought... MCP, if history is any guide, will go down one of two paths: 1. Everyone will use it. 2. Everyone but one key platform will use it. When (1) happens two things will be true. First, no one will effectively monetize it. Second, every vendor will also add unique aspects to how they consume (client) or produce (server) the exchange...
Context & Ripple Effects
MCP began as Anthropic's open-source protocol for connecting AI assistants to external data, placing it within a broader contest over how AI systems integrate with tools and information. This analysis shifts attention from the protocol’s specification to the operational layer that has historically made middleware difficult to sustain.
The concern is timely because major AI platforms have pursued differing approaches to integration and modularization. MCP’s value proposition is interoperability, but that also makes it vulnerable to the incentives that fragment shared interfaces.
First-order effects
- MCP adopters must solve practical middleware problems—reliable connections, compatible implementations, and incentives for both clients and servers—rather than treating an open specification as sufficient.
- Vendors using MCP can differentiate through custom extensions on either side of an exchange, potentially reducing out-of-the-box interoperability even while retaining the protocol’s name.
Second-order effects
- Tool and data-source providers may face pressure to support MCP alongside proprietary integration paths, increasing implementation and maintenance work.
- If major platforms extend MCP differently or withhold full participation, customers could encounter the same cross-platform friction that a common protocol is meant to remove.
Third-order effects
- MCP is a test of whether AI-agent integration can become shared infrastructure instead of a collection of platform-controlled connectors; that outcome depends on governance and implementation discipline, not adoption claims alone.
- If interoperability remains incomplete, value is likely to concentrate in managed layers that handle authentication, reliability, and vendor-specific differences rather than in the base protocol itself.
The trend: AI-agent ecosystems are moving toward common connection standards, while competition shifts to the managed and proprietary layers built around them.