Glue between ripgrep, BM25, and embeddings, kept strictly local

What it does zg is a local-first search layer—essentially well-engineered glue—that unifies ripgrep, BM25, and vector search behind a single interface. It indexes your workspace so both terminal users and AI agents can query with natural language, keywords, or regex and receive ranked, source-linked evidence that preserves file structure and exact locations.

The interesting bit Rather than treating semantic and lexical search as separate tools, zg orchestrates them in a retrieval loop: semantic discovery narrows the candidate space, then BM25 and ripgrep anchor exact identifiers. The README publishes paired A/B benchmarks suggesting this reduces agent tool calls, input tokens, and wall time compared to unassisted baselines.

Key highlights

• Fully offline by default: indexes, files, and local embedding models stay on your machine; remote services receive data only with explicit permission.
• One index serves both human CLI queries and agent tool calls via MCP integrations with OpenCode, Claude Code, Codex, Cursor, and others.
• Returns compact, ranked evidence with source locations rather than dumping entire files into context.
• Benchmarked across SWE-QA-Bench and BrowseComp-Plus, plus case studies on Pylint, Matplotlib, and Django repositories.
• Requires Node.js 22 or newer.

Caveats

• The provided case studies focus on Python repositories; broader language performance is claimed but not demonstrated in the sources.
• The README admits agent outcomes vary by model and run, so your mileage depends on which agent framework and model you pair it with.
• It leans heavily on upstream components (zvec, embedding models, ripgrep); the value is in the orchestration and local packaging.

Verdict Worth a look if you’re building agent workflows that grope around large codebases, or if you want semantic search without shipping your source to a SaaS. Skip it if you already have a tuned retrieval stack you trust, or if you need deep IDE integration rather than a CLI/agent tool.