Coding agents burn an outsized share of their tool budget on search, and the failure mode is familiar to anyone who has watched one flail: the target is a symbol, so ripgrep nails it instantly, but the target is a behavior described in plain English, keyword matching whiffs, and the agent starts guessing terms, reading entire files, and stitching context together by hand. Every detour costs tool calls, tokens, and clock time.
Qwen Developer’s answer is zg, a local-first search layer that puts semantic search, BM25, and ripgrep behind one interface. It ships under the zvec-ai GitHub org with an Apache 2.0 license, installs from npm as @zvec/zvec-grep, and needs no GPU with the default model. One index, four retrieval routes: a hybrid default, —fts for BM25, —vector for pure semantic similarity, and —rg for literal or regex matching that needs no index at all. Results report fresh or possibly_stale, so an agent can skip the status preflight and just act.
The MCP surface is deliberately restrained. zg install detects Codex, Claude Code, Cursor, and OpenCode, then exposes exactly two tools by default: zvec_grep_search for known intent without exact strings, and zvec_grep_rg for known symbols or regexes. Index lifecycle stays with the CLI. A six-tool compatibility set exists but is opt-in, and the docs explicitly state an agent must never silently create, rebuild, or delete a persistent index. Embeddings run on device by default, with remote Qwen endpoints gated behind an explicit —allow-remote flag or a signed workspace grant rather than a mere credential.
The vendor’s own A/B tests claim a 43.52% drop in tool calls and a 38.58% cut in agent time on an 80-question BrowseComp-Plus sample, alongside a 1.5-point Judge score bump on 20 SWE-QA-Bench questions. Those sample sizes are small enough that independent replication is the obvious next step. Still, the architecture is the interesting part: a loopback-only MCP server, anonymous workspace indexes, and context-economy output shaping suggest Qwen is designing for agent ergonomics first and benchmark theater second. That instinct has been rare in the agent-tooling wave so far.
