Google joins 'Agent Plugins,' a platform that allows AI agents to carry their skills and MCP servers with them, in collaboration with OpenAI, Microsoft, Amazon, and others.
Google has joined the core maintainers of 'Agent Plugins 1.0.0,' an open specification that allows AI agent skills and MCP servers to be carried together. Agent Plugins was a specification released by the Technical Steering Committee (TSC), which consists of representatives from Amazon, Cursor, Microsoft, OpenAI, and Vercel, and Google has announced that it will also be working on supporting its own products.
Agent Plugins package your skills, tools, and more - Google Developers Blog
AI agents utilize 'Agent Skills,' which are sets of specific work procedures and knowledge, and the 'Model Context Protocol (MCP),' which connects the AI agent to external tools and services. While Agent Skills and the MCP server can be used individually by different AI clients, distributing multiple functions as a single package requires developers to recreate packages for each client because the directory structure and configuration file format differ for each AI client.
Agent Plugins are not intended to replace Agent Skills or MCPs themselves, but rather to standardize the format of 'boxes' that hold multiple components. By allowing different AI clients to load the same package, it is said that the need to rewrite code for each client will be reduced.
Agent Plugins are structured as a single directory, with a 'plugin.json' file containing the plugin name and other information placed in the root directory. Agent Skills are stored in the 'skills' directory, and MCP server settings are described in 'mcp.json'. Because the locations are fixed, corresponding AI clients do not need to guess the file locations using their own methods.
The MCP server allows you to explicitly specify connection methods such as 'stdio' which uses standard input/output, or 'Streamable HTTP' which uses HTTP. Furthermore, even if one MCP server fails to start, the design ensures that the entire plugin remains usable, and available skills and other components continue to load.
On the other hand, they won't be standardizing AI client-specific features. If a client wants to add client-specific features, they can create a dedicated area like 'com.example.client,' and clients that don't support it will ignore that part. The idea is to standardize only the parts that can be standardized, while leaving room for each AI client to add its own unique features.
Furthermore, Agent Plugins is a useful mechanism for bundling and distributing multiple skills or MCP servers together, and it is not necessary to make individual skills or MCP servers into Agent Plugins. Installation methods, distribution methods, permission management, sandboxing, verification of the reliability of the distributor, and the user interface are also not covered by Agent Plugins 1.0.0.
From Google, Kevin Hou, Senior Staff Engineer at Google DeepMind, will participate as a core maintainer. In addition, Google's 'Agents CLI' and 'Data Agent Kit' already support the Agent Plugins format, and the Data Agent Kit will allow compatible clients to use skills to connect to BigQuery, Spanner, Cloud SQL, etc., as well as MCP servers.
Google has indicated that it plans to expand Agent Plugins support to even more Google products that handle Agent Skills and MCP servers in the future.
Related Posts:
in AI, Posted by log1d_ts