BREAKING
Code Graphs Cut AI Token Use 95%
How the knowledge graph works
1Parse AST with tree-sitter
2Store graph in SQLite
3Expose tools via MCP
4Query only what's needed
0%
Max token cut
0%
VS Code repo cut
0%
Fewer tool calls
0%
Shorter time
Tokens per query: before vs after
Before12000
After2000
Strengths and limits
Strengths
Traces blast radius of changes
Clearer PR impact scope
Runs 100% locally, often open source
Limits
One-time index build cost
MCP-only agents
Still needs design intent, ADRs
Smarter agents, fewer tokens
AI NEWS BLITZ
A new technique is slashing how many tokens AI coding agents burn through.