get-usage-dependency-links walks the graph rather than the text, so the answer to what depends on this covers indirect callers a search across files would miss. Keeping documentation search on its own tool is the other quiet win: a question about how the project is configured returns prose, not the first function whose name happened to match. Everything is answered from the pre-built graph, so what you get back reflects the indexed repository rather than uncommitted work on your disk.
A client over a pre-built CodeGPT graph of a repository, where each node holds an implementation, its documentation and its relationships to other nodes.
- The complete implementation of a named class, function or method, retrieved from the graph rather than by reading files
- The first-level relationships of a node: what references it, what it calls or uses, and where it is declared
- An adjacency list of every functionality that a change to one entity would affect
- Natural-language search over code entities that matches on meaning rather than on keywords
- A separate natural-language search over the repository's documentation files, so a setup question does not return source code
- The folder tree under any path in the graph
A CodeGPT graph already built for the repository, addressed by CODEGPT_API_KEY, CODEGPT_ORG_ID and CODEGPT_GRAPH_ID.
One command plus a key — npx -y mcp-code-graph@latest CODEGPT_API_KEY CODEGPT_ORG_ID CODEGPT_GRAPH_ID, then supply credentials