What is MCP?
MCP (Model Context Protocol) is an open standard that specifies how AI applications read files, call functions and obtain context. Anthropic introduced it on November 25, 2024; per Wikipedia it was designed by David Soria Parra and Justin Spahr-Summers.
The architecture has three roles. The host is the AI application that talks to an LLM, such as Claude Code. The client is created by the host, one per server. The server provides tools (such as database access) or resources (such as documents). They communicate over JSON-RPC 2.0. For designers, the most direct example is Figma's MCP Server: through it Claude Code reads your design file rather than a screenshot.
Why does MCP exist?
Before MCP, every pairing of an AI application and an external tool needed its own connector, so N applications and M tools meant N×M integrations, which Wikipedia calls the N×M problem. MCP lets the tool side write a server once, and any MCP-capable application can use it.
Other vendors adopted it quickly: per Wikipedia, OpenAI followed in March 2025 and Google DeepMind also adopted it in 2025. In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI.
How does MCP work?
When the host starts, it connects to each MCP Server and asks for its list of tools and resources, then hands those descriptions to the LLM. When the model decides it needs a tool, the host sends the call through the matching client and the result returns to the conversation.
Take Figma as the example: its server offers tools such as get_design_context, get_variable_defs and get_screenshot. After you paste a Figma link, Claude Code calls those tools to get structure, variables and a screenshot before it starts writing code. Servers can run locally or remotely; Figma's is remote, with sign-in through OAuth.
What does MCP mean for me?
The benefit is connecting once and sharing across several AI tools, so your design files, document stores and project trackers can be read by the same workflow. The cost is security: per Wikipedia, Invariant Labs described "Tool Poisoning" in April 2025, where malicious instructions hidden in tool descriptions are visible to the model but not to users, and a critical flaw was found in mcp-remote in July 2025 (CVE-2025-6514, CVSS 9.6). So connect only servers you trust, and watch what data each one is authorized to read and write.
The protocol is also not settled. Figma co-founder Dylan Field has said publicly that MCP "is not a perfect protocol," with plenty of room for improvement.
At Config India in October 2026, Figma announced 20 new MCP partners (TechCrunch named Amp, ClickUp, JetBrains Junie, Linear, Notion, Mistral and others). A designer using Claude Code only needs claude mcp add to connect Figma's remote server so it can read layout and variables, instead of anyone writing a custom integration.
Pros: write a server once and share it across AI applications, with a fast-growing ecosystem. Cons: the security model is still catching up (Wikipedia cites a May 2026 NSA statement that rapid proliferation has outpaced it), and the protocol itself is still being revised.