Flexible tool annotations

[laurentsimon]

Is your feature request related to a problem? Please describe.

The tool annotations are static and cannot be customized. It has two limitations:

1. Different use cases may need to label their tools differently to enforce policies (an example is Proposal: Leveraging Action Manifests for Better Agent Security #440). It's likely that certain annotations will differ based on ecosystem (Android vs Cloud agents vs ...). A way to customize these annotations would be useful so we don't have to create extensions or different SDKs for different ecosystems.
2. It's currently impractical to experiment with new annotations without breaking the ecosystem of MCP SDKs. As such, it's hard to validate what works and what does not to inform the specs.

Describe the solution you'd like

A potential solution would be to turn the annotations into a list of key / value pairs. Something like:

toolAnnotations = [
  {
     "mcp.org/openWorldHint": {
        "value": true,
      }
      ...
     "mycompany.com/something": {
        "value": { object },
        "critical": true,
        "experimental": true
      }
  }
]

All existing annotations would be expressed this way and new ones may be added overtime. In the example above, an agent implementation would ignore fields it does not understand, and fail only if the annotation is marked "critical", for example. (like x509 certificate extensions)