Improve LLMs with Resources
Opened this issue · 1 comments
msanatan commented
MCP spec: https://modelcontextprotocol.io/specification/2025-06-18/server/resources
FastMCP documentation: https://gofastmcp.com/servers/resources
There are some major benefits for converting read-only tool calls into resources:
- Better performance as there are less tools in the context
- Better UX as users can select resources relevant to the context
- LLMs can discover resources on their own w/o having to make a tool call, so potentially faster
- Good MCP hosts cache the resources, so it's faster
- Actually follows the protocol
This seems like the ideal way to encode data for models. Here are some things we should add:
- Scene hierarchy, we should make a compact representation of the game objects in a scene for the models
- Serialized game objects. Right now we do searching and the like, but we should get a game object with instance ID or path as a resource
- Logs, instead of a tool call it could be a resource as well
- Packages
- Menu items (I recently made this a tool call, it should be a resource)
- EditorPrefs
- Tests
- Editor state (read only items in manage_editor)
- Tags
- Layers
What else could be added?