AI coding agents like Claude Code, Cursor, and Copilot re-read entire files every time they need to understand your code. 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 bill.

We built Code Context Engine, an open-source tool that indexes your codebase locally using tree-sitter AST parsing, stores vector embeddings in sqlite-vec, and serves relevant code chunks through the Model Context Protocol (MCP). The result: 94% token savings, benchmarked on FastAPI with 20 real coding queries.

For embedded and systems teams, AI coding costs compound fast: large C++ codebases, monorepos with hardware abstraction layers, and strict toolchain constraints. Code Context Engine runs entirely local (no cloud calls for indexing), works with C, C++, Rust, Python, and Zig via tree-sitter grammars, and integrates with your existing editor through the Model Context Protocol.

In this talk, we walk through: tree-sitter AST parsing for semantic code chunks, hybrid retrieval combining vector similarity with BM25 via Reciprocal Rank Fusion, confidence scoring, file diversity filtering, content-hash embedding cache (96% hit rate), secret redaction and PII scrubbing, and cross-session memory using a SQLite-backed knowledge graph with CALLS/IMPORTS edges.

We show live benchmarks, demonstrate MCP integration across Claude Code, Cursor, VS Code, Gemini CLI, and Codex, and share engineering decisions behind choosing sqlite-vec over LanceDB (99% smaller install), truncation compression over LLM summarization, and RRF over learned reranking.

Attendees learn how to build a local RAG for code, measure token savings rigorously, and reduce AI tooling costs. MIT-licensed: github.com/elara-labs/code-context-engine