NanoNets/Graft
Graft test: turn 222 files into a queryable code graph
A real project produced 1,794 symbols and 4,464 relationships, then exposed the code structure to Codex through MCP.
01 · PROBLEM
What I wanted to solve
In a large repository, coding agents can spend substantial context rediscovering entry points, functions, and call paths. Keyword search shows fragments but does not reveal every file affected by a change.
02 · PROCESS
What I did
We ran Graft structural analysis on a real project, used tree-sitter to map files, symbols, calls, and imports, then loaded its MCP connection so Codex could query the repository by symbol and relationship.
03 · RESULT
What actually happened
The graph covered 222 files, identified 1,794 symbols, and created 4,464 relationships. MCP loaded successfully, giving the agent a structural entry point instead of relying only on file-by-file search.
04 · PITFALLS
What others should watch for
These counts prove that the graph was generated; they do not prove a fixed token-saving percentage. The base graph can run locally, while deep summaries use a model selected by the user and must be measured separately.
05 · VERDICT
My conclusion
Graft has demonstrated that it can create a queryable structural map for a real repository. The next test should compare token use, tool calls, speed, and repair accuracy on the same task with and without the graph.