Optimizing token usage when using Affine MCP server for AI access

[slfields]

When using Affine MCP for using AI to peruse my documents, I note it's using large number of tokens per document.

Are there any configuration options to optimize the token usage?

Can the MCP be configured to use TOON or other MCP native encodings that reduce the token usage?

[DAWNCR0W]

A few practical things help here, even before changing the encoding format.

I maintain a community AFFiNE MCP server, and the biggest token wins have usually come from reducing the exposed tool surface and avoiding full-document reads until the agent has narrowed the target set: https://github.com/DAWNCR0W/affine-mcp-server

The patterns I would recommend are:

1. Start with a smaller tool profile when possible. For example, expose read/search tools for browsing workflows and enable authoring/destructive tools only when the session needs them.
2. Use discovery tools first: list/search docs, list tags, filter by tag/title/updated time, then call read_doc only on the few documents that are actually relevant.
3. Prefer explicit workflows such as “find docs tagged X, then read the top 3” over broad “read my workspace” prompts.
4. If the client supports separate MCP configs per task, keep a read-only AFFiNE config for research and a fuller config for editing sessions.

TOON or another compact encoding could help for large structured responses, but in my experience the first-order issue is usually tool and document selection rather than the JSON syntax alone. A compact response format is most useful after the server already supports narrow, predictable reads.