Primitive Grouping

[chughtapan]

Large MCP servers e.g, Github, Azure, Workspaces, etc., organize primitives into semantic groups. Clients select the groups to use for a particular request / scenario, and servers only list primitives corresponding to those.

Currently, even though some servers and gateways implement grouping, they do so in an ad-hoc manner, which prevents clients from engaging with them meaningfully.

Examples:

• Servers: GitHub MCP Server, Azure DevOps, Workspace MCP
• Gateways: MCPJungle

References:

• SEP-2084
• SEP: Primitive Groups

[scottslewis]

Currently, even though some servers and gateways implement grouping, they do so in an ad-hoc manner, which prevents clients from engaging with them meaningfully.

IMO this does not capture server-side primitive grouping concerns. A couple of those concerns

1. The context exchange size that can result from having a large number of tools (and/or other mcp primitives)...e.g. a gateway server. Server-side grouping can/and does help deal with such context explosions by grouping tools (for example), thus reducing the runtime context exchange necessary for tool discovery, selection, and orchestration by LLM clients.

2. Server-side grouping provides a way to organize domain-specific mcp servers (e.g. supporting a specific team, process, or workflow).

3. Server-side grouping provides a means to dynamically restructure a domain model from 2.

There are other server-side use cases articulated from previous discussions use case discussion having to do with using grouping for server-side governance, and management.

Perhaps a 'server-side primitive grouping' discussion should be created, to focus specifically on server-side uses and concerns wrt primitive grouping.