Model Context Protocol: Launch, SDKs & How MCP Works

What shipped in November 2024

Anthropic released the Model Context Protocol on November 25, 2024, alongside a specification, software development kits, and example servers for systems including GitHub, Slack, and Postgres. Claude Desktop could connect to local MCP servers immediately. Toolkits for remote production servers were still described as forthcoming. [1]

Atlas interpretation: The practical aim was to stop rebuilding the same connection for every AI application. A developer could put a database integration behind an MCP server, then make it available to compatible assistants and editors through the same interface. The database-specific work still existed; the reusable part was how an AI application discovered and requested that work. Support depended on the features each side implemented. [1][3]

How a request reaches a tool

MCP separates the application from its connectors. The host is the AI application: it manages the conversation, model interaction, and permission decisions. Inside it, a client maintains a connection to one server. That server exposes particular data or operations. A host can run several clients and decides which context to share with each server. [3]

Atlas interpretation: Imagine asking an assistant for the number of open support tickets. An MCP server could advertise a ticket-counting tool with a name, description, and expected inputs. The application's client discovers it with tools/list. If the model selects it and the application permits the action, the client sends tools/call with the chosen inputs. The server queries the support system and returns the result for the assistant's answer. MCP defines those messages; someone still has to implement the query and supply the right access to the support system. [4][3]

Tools were only one part of the original design. Servers could also expose resources, such as a file or database schema to use as context, and prompts, reusable message templates. These serve different purposes: supplying information, offering a prepared interaction, and performing an operation. An application did not have to implement every optional feature to speak MCP. [5][6]

Another route into software

Atlas interpretation: A month earlier, Anthropic had released Claude's computer-use beta, which let developers build systems that worked through screenshots, mouse movements, and typing. MCP offered a different interface. A ticket-counting tool can return a number without the model locating a search box and reading a results screen. Computer use can work through an interface built for people; MCP requires an integration that exposes the needed capability. [7][4]

A common interface still needs access controls

The original specification left authentication and authorization outside its core, allowing clients and servers to arrange their own methods. It assigned consent and security responsibilities to applications and required servers to enforce access controls. A standard way to request an action does not determine whether a particular user is allowed to take it. [6][5][4]

Atlas interpretation: For the support-ticket example, counting tickets and deleting them need different permissions even if both operations use tools/call. A connector that exposes too much authority remains dangerous behind a standard interface. MCP makes capabilities easier to connect; the application and the system behind the server still have to decide what should be allowed. [4][3]

The later move to a foundation

On December 9, 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. It joined Block's goose agent framework and OpenAI's AGENTS.md convention as founding projects. That gave the protocol an institutional home alongside contributions from other companies, more than a year after the original release. [8]