MCP

MCP, short for Model Context Protocol, is an open protocol connecting AI applications to external tools and data. A coding assistant can use it to retrieve design context; a desktop assistant can connect to a local service. The practical attraction is a common integration interface: applications and capability providers can implement the same protocol instead of building a separate connector for every pairing. Anthropic launched MCP on November 25, 2024, crediting David Soria Parra and Justin Spahr-Summers as its creators.[1][3]

In their April 2025 creator interview, Soria Parra and Spahr-Summers recalled starting around July 2024, frustrated by moving material between Claude Desktop and an IDE. They pursued complementary experiments: a Desktop integration and an IDE proof of concept. That everyday friction supplied a concrete starting point for a broader problem—how to make useful context available across applications without repeatedly rebuilding the connection.[2]

An important precedent was the Language Server Protocol, which lets editors share language tooling through a common interface. The creators described MCP in similar terms: many applications on one side, many integrations on the other. The launch package made that idea usable through an open specification, SDKs, reference servers and local-server support in Claude Desktop.[2][1][3]

MCP separates three participants. The host is the AI application. It creates clients, the connectors through which it communicates with servers that provide capabilities. The architecture documentation separates JSON-RPC message semantics from transport: standard input/output connects local processes, while Streamable HTTP supports remote communication. MCP defines the exchange; it leaves the application's use of models and management of context to the application builder.[4][3]

• Resources supply data or context that an application can retrieve.[3]
• Prompts supply reusable message templates or workflows.[3]
• Tools expose callable functions, allowing an application to request an operation.[3]

Figma's Code Connect integration makes this exchange concrete. An AI coding application communicates through its MCP client with Figma's server to request design context. For mapped components, the returned context can include properties, imports, usage snippets and instructions connecting the design to existing code. The application then decides how to supply that context to its model when generating code. MCP carries the request and response; Code Connect supplies knowledge of the components, and the coding application manages the model interaction.[10][11][4]ⁿ¹

The connection can also carry actions. Figma's Write to canvas feature lets an assistant create or modify native canvas content by running JavaScript through Figma's Plugin API via the remote server. This turns a tool call into a change in a design file. Figma describes canvas writing as beta quality that may require manual review and cleanup.[12][11]ⁿ²

MCP's design has continued to change. The March 26, 2025 revision introduced an OAuth 2.1 authorization framework and replaced HTTP+SSE with Streamable HTTP. The July 28, 2026 release made a further consequential change: it removed the initialization handshake and protocol session identifier in favor of self-contained requests carrying their own metadata. These are versioned changes; existing clients and servers may still implement earlier revisions.[5][8]

Stateless protocol requests do not require stateless applications. The 2026 release explains that application state can persist through explicit handles, allowing independent requests to reach different server instances. It also introduces Multi Round-Trip Requests: when an operation needs more input, the server returns input_required, and the client retries with inputResponses. That supports mid-call input without requiring a permanently open reverse stream. The distinction is useful for builders: state becomes an explicit part of the application's interaction rather than something hidden in a transport session.[8]

A common interface can expose consequential capabilities, including private data access and code execution. The specification calls for consent and access controls while acknowledging that the protocol cannot itself enforce those principles. Its security guidance explicitly forbids token passthrough: a server must not accept tokens that were not properly issued to it and simply forward them to downstream APIs. Correct authorization requires preserving the boundaries between services.[3][13]

There is also a model-level problem. In an April 2025 analysis, Simon Willison explained how untrusted incoming content can instruct a tool-using model to disclose private information. He emphasized that this risk extends beyond MCP. The engineering implication is that protocol compatibility does not establish the trustworthiness of retrieved instructions or the safety of an action. Builders still need to decide what data an assistant can reach and what operations it may perform.[14][3]

Anthropic announced MCP's donation to the Agentic AI Foundation, a Linux Foundation directed fund, on December 9, 2025, while retaining the project's governance model. MCP remained listed in the foundation's project directory as of September 15, 2026. The August 2026 roadmap identifies David Soria Parra and Den Delimarsky as Lead Maintainers and sets priorities including agentic messaging, HTTP transport, identity and security, improved primitives and SDK experience. Those priorities describe intended work.[6][7][9][3][10]