Anthropic's MCP standard hits 10,000 servers, now under Linux Foundation

The protocol's ownership is the part that changed most. MCP began as an Anthropic-published specification in November 2024, but the December 9, 2025 donation to the Agentic AI Foundation — co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft and AWS — means competing AI vendors now adopt the same standard. For a company weighing whether to build integrations, that lowers the risk that one vendor's policy shift makes the work unusable.

The architecture explains where that work goes. An MCP host is the AI app a user operates, MCP clients handle communication with each server, and MCP servers expose external tools' functions to the AI. Servers offer three things: tools that let the AI act, resources that let it read data, and prompts that supply templated instructions. Tools carry the most risk because they can modify data, which is why the guidance is to start with read-only resources and expand permissions gradually.

Adoption on the tool side is spreading among services used in Japan, including freee's freee-mcp, released as open source in March 2026, plus Backlog, Notion and Atlassian servers. That breadth is also the exposure: unofficial servers, local servers that can run arbitrary commands on a PC, and prompt injection — where hidden instructions in a web page or email trick the AI — are the risks flagged. Admin-controlled plans from ChatGPT Business and Claude Team are presented as the way to keep the list of approved servers visible.