AI coding agents like Claude Code and Cursor re-read entire files every session. On a medium project, that is 45,000 tokens per query when the agent only needs 4,900. Input tokens are 85-95% of your AI tooling bill.

We built Code Context Engine (open source, MIT), a local code indexing tool that serves only the relevant code chunks through MCP (Model Context Protocol). The result: 94% token savings, benchmarked on FastAPI with 20 real queries, 0.4ms latency.

In this session with live demos, we walk through:

- Tree-sitter AST parsing for semantic code chunks (functions, classes, modules) across Python, JavaScript, TypeScript, Go, Rust, Java, PHP
- sqlite-vec for vector search, FTS5 for keyword matching, stored in three SQLite files (2 MB vs 217 MB for LanceDB)
- Hybrid retrieval: vector similarity + BM25 via Reciprocal Rank Fusion with confidence scoring
- Code knowledge graph with CALLS/IMPORTS edges for automatic graph expansion
- Content-hash embedding cache achieving 96% hit rate on re-index
- Secret redaction (AWS keys, GitHub tokens, JWTs) and PII scrubbing before indexing
- Cross-session memory: decisions and code areas persist across sessions

Live demo: index a project, ask questions, show before/after token counts, demonstrate savings in real dollars.

Works with Claude Code, VS Code, Cursor, Gemini CLI, and OpenAI Codex. One index, all editors.

github.com/elara-labs/code-context-engine