Local Persistent Code Indexing for AI Agents

_febin_p_ · reddit · 2026-07-19

The author built a local persistent code indexing daemon (mcp-injector) for AI coding agents, aiming to prevent agents from repeatedly reading entire repositories and wasting context.

Core Approach

Scans repos using language-specific AST parsers, mapping symbols, files, and line numbers into a local SQLite database (WAL mode)
getprojectmap returns an AST-folded project graph: preserving function signatures, imports, types, and interfaces, while replacing function bodies with compression markers
This allows the model to know "what exists" first, fetching raw source code on demand

Benchmarks & Engineering Details

In one repo, the initial project map shrank from 892k tokens to 143k tokens, an 84.9% reduction
To ensure stable hits on Anthropic's prompt cache, the author normalized volatile factors like file order, timestamps, mtimes, and newlines, guaranteeing byte-level consistent maps for the same repo across multiple generations
Incremental updates are handled via inotify/FSEvents and a git post-checkout hook, rebuilding only changed files
Provides higher-level MCP tools like BM25 symbol search, raw file retrieval, dependency/impact traversal, Mermaid graphs, git context, regex search, and DB schema checks
Payloads are validated before writing back to prevent agents from accidentally writing compressed representations into source code
Uses entropy heuristics to detect and redact suspected secrets before indexing

Original post →

Local Persistent Code Indexing for AI Agents

More from coding & agent