Graphenium

Graphenium takes a repository and builds a persistent structural knowledge graph from it, capturing function callers, module dependencies, and component relationships in a format that MCP-compatible AI assistants can query directly. The core workflow is: run the install script, index your repo, and point Claude, Cursor, or any other MCP-aware assistant at the running graph service. Queries that would otherwise require the assistant to read and re-read source files return in approximately 20 ms, per the project’s documentation.

The differentiating architectural choice is persistence across sessions. Most AI assistant integrations treat codebase knowledge as ephemeral — the assistant reads files, builds context, and forgets when the session closes. Graphenium writes that structural map to a graph that outlasts any single conversation, so the assistant picks up where it left off without re-spending tokens on orientation.

The tool fits squarely into the navigator role: it answers ‘what calls this function’ and ‘what depends on this module’ faster and cheaper than file-by-file reading. Where it stops is at the boundary of reasoning — Graphenium supplies the graph; the assistant decides what to do with the answer. Teams who need the graph to reflect commits in near-real-time will need to trigger re-indexing themselves, as no automated sync mechanism is described in the available documentation.

Graphenium is written in Rust, ships as a single binary via install.sh, and exposes an API alongside its MCP interface. The repository is MIT-licensed with no paid or hosted tier. A .grapheniumignore file lets teams exclude paths from indexing, which is relevant for monorepos with generated or vendored code that would inflate the graph without adding navigation value.