01
Problem
Claude Code answered “what calls this?” and “what breaks if I change that?” by reading source into context — expensive on big repos, incomplete when files don’t fit, and impossible past the context window.
← Selected work
Open source
Claude Graph
Structural code answers without dumping the repo
7×–236× fewer tokensOfflineOn PyPI
Stack
PythonTree-sitterSQLiteMCPClaude CodePyPI
02
Solution
Build a persistent local call graph (Tree-sitter → SQLite), expose focused MCP tools and an offline HTML viz, and ship a one-command install that wires .mcp.json + Claude skills. Claude queries the graph instead of dumping the repo.
03
Where AI fits
Doesn’t replace the model — it gives Claude Code structural memory of the repo. Fewer tokens, faster answers, and no cloud round-trip for the graph itself.
04
Outcome
On Django (3,040 files, ~45k nodes, ~939k edges): structural queries use 7×–236× fewer tokens than sending matching files. Full Django dump is ~1.5M tokens — past Claude’s window without a tool like this. Published on PyPI as claude-graph (v0.2.3+).
Want something like this for your team?