MCP server cuts markdown vault token use by about 60%
A local keeps ongoing project knowledge in a markdown vault, so the same does not need to be entered again for every new chat. The model can see only 10 tools rather than more than 20. Tool descriptions consume tokens on every turn even when the tools are not used, so a smaller tool set lowers cost and reduces the chance of choosing the wrong tool.
Search initially returns ranked excerpts instead of complete files. If more detail is needed, the model can follow wikilinks for and titles, request an outline, fetch one section, and use the full note only as a final step. On a real vault, this approach used about 40% as many tokens as a simple RAG setup that loaded full files.
The cost per answer also stayed roughly level as the vault grew. Writing is restricted to a fixed folder structure and an approved list of tags; unknown tags become instead of being saved automatically. Similar folder names are matched to existing ones, and every change enters an append-only identified by a hash of the access token rather than the token itself.
Key points
- Only 10 tools are shown to the model, reducing token use on every turn.
- Search provides ranked excerpts before any complete document is loaded.
- More information is fetched gradually through wikilinks, outlines, sections, and finally full notes.
- The tested vault used about 40% of the tokens required by basic full-file RAG.
- Controlled tags, standardized folders, and an govern every write.