zg is a CLI tool that indexes a local codebase once, then exposes four query modes via one interface: hybrid (semantic + lexical), BM25 keyword-only, vector-only, and ripgrep literal/regex. It installs from npm, runs embeddings on-device by default using small models (256-dim static lookup), and wires into Claude/Cursor/Codex via MCP with just two tools. The vendor reports ~40–50% reductions in tool calls and tokens on small benchmark sets (20 and 80 questions), but those are their own runs and haven't been independently replicated. Index updates are incremental; remote embedding requires explicit per-command auth. The restraint in the MCP surface—no silent index creation—is deliberate design. Deployable today under Apache 2.0. The source is a marketing post, so the benchmarks should be read as a proof-of-concept signal, not ground truth. Index size, latency for large repos, and performance on actual agent loop cold-starts aren't in the text. One small flag: the launch post cites eleven on-device models but the docs list ten—worth checking if that's a sync issue.