An in-progress MCP server that gives AI coding agents a structural map of a codebase, so they can retrieve the right code instead of paging through whole files and burning tokens.
The problem
Coding agents waste enormous context reading entire files to find one function.
Full-text search helps, but it doesn't understand structure — it can't ask for
"the definition of getUser and its callers" as a first-class query.
What I'm building
- A tree-sitter-based indexer that parses source into symbols, definitions, and references.
- An MCP server exposing structural queries (definition, references, neighbourhood) as tools an agent can call.
- A ranking layer that returns the smallest relevant slice, not the whole file.
Status
Work in progress. Early results show meaningful token savings on agent runs; the accuracy evaluation harness is the current focus.