SS
About Me
Frontier AI Paper BriefingsPokebowlClinical Trial EnrollerLittle Human Names
DisclaimersPrivacy PolicyTerms of Use
Privacy Policy·Terms of Use·Disclaimers

© 2026 Silvia Seceleanu

← Back to Explorer
Products·Anthropic·Nov 2024

★22. Model Context Protocol (MCP) Launch

Open JSON-RPC 2.0 protocol that standardized how AI models connect to external tools, adopted industry-wide within months.

Product Announcement
Summary

Open standard for connecting AI models to external tools and data sources via JSON-RPC 2.0. Solved the M×N integration problem. Released with Python and TypeScript SDKs plus pre-built servers for GitHub, Slack, Google Drive, Postgres. Became the de facto industry standard; adopted by OpenAI, Google, Microsoft within months.

Key Concepts

Model Context Protocol (JSON-RPC 2.0)

An open standard protocol based on the well-established JSON-RPC 2.0 specification for bidirectional communication between AI models and external tools. The protocol defines how models request tools, how tools respond with data or execute actions, and how errors are communicated. Using JSON-RPC (rather than inventing a new protocol) leveraged battle-tested infrastructure that developers already understand. The simplicity is intentional—MCP isn't meant to be sophisticated, just standardized.

Tool-Use Standardization

The creation of a common vocabulary and calling convention so that models from different providers can call the same tool without provider-specific integration work. Before MCP, Claude needed custom integration with GitHub, GPT needed different GitHub integration, and Gemini needed yet another version. Standardization collapses this M×N problem into M+N: each tool builds one MCP server, each model implements one MCP client, and everything interoperates. This is analogous to how USB standardized hardware interfaces.

Server-Client Architecture

The asymmetric relationship where models act as clients (requesting information, triggering tools) and external services act as servers (exposing capabilities, returning data). The server is responsible for security, authentication, and enforcing what the model can access. A GitHub MCP server exposes only certain repositories and actions, not the entire GitHub account. This architecture gives server implementers fine-grained control over what capabilities they expose, enabling trust-based integration.

Universal Connector Pattern

The abstraction that allows different classes of external systems (databases, APIs, web services, local applications) to be exposed to AI models through a common interface. A Postgres server looks like a GitHub server looks like a Google Drive server from the model's perspective—all are MCP servers. This universality enables rapid tool ecosystem growth: once developers understand how to build an MCP server, they can expose any external resource.

Open-Source Infrastructure Play

The strategic decision to release MCP as open-source, free, standardized infrastructure rather than proprietary technology. This reduces Anthropic's short-term competitive advantage but increases long-term ecosystem influence. By making the standard free and open, Anthropic makes it almost impossible for competitors to maintain a proprietary alternative. OpenAI, Google, and Microsoft all had to adopt MCP because rejecting a free open standard is politically untenable. The play mirrors Android (Google sacrificing iOS-level control for market dominance) and PCs (IBM opening the architecture to competitors but establishing the category).