Jevgrep cuts coding-agent orientation costs by roughly 30% on a small benchmark

Coding agents burn a surprising share of their token budget just orienting themselves in an unfamiliar codebase — reading files speculatively, following wrong leads, backtracking. Jevgrep attacks that specific waste. You give it a behavioral question and a directory; it returns relevant files, source excerpts, and declaration locations in a single stdout response, letting the downstream agent skip the orientation phase and go straight to implementation.

The mechanism is worth understanding. Jevgrep walks the repository hierarchy, uses content previews to select candidate files, then identifies source units and surrounding context. It does not force results into a fixed-size list — if a file qualifies but an excerpt cannot be confidently extracted, the file location is still returned. Declaration parsing is available for Python, TypeScript/JavaScript, Go, and Rust; everything else falls back to text search. The output is explicitly framed as evidence for the agent, not a generated answer.

The cost claim is the headline, and the README is unusually precise about it. A ten-task SWE-bench comparison found that both the Jevgrep-assisted agent and the baseline solved 8 of the 10 tasks. Sol-only cost dropped from $7.62 to $5.44, a measured 28.6% reduction rounded to the advertised ~30%. A later rerun including Jev's own inference cost measured 25.8% lower total cost at the same solve rate. A subsequent 0.5.0 evaluation held the 8/10 solve rate while cutting native Jev cost by roughly 59% versus the prior run, at a 2–3% combined cost increase — a tradeoff the authors accepted. The README is careful to note these are single-run observations on ten tuned Python tasks, not statistical equivalence claims.

The setup has two distinct parts that are easy to conflate. Installing the CLI gives you the jg binary. Installing the skill — a separate step via jg skill — is what actually teaches your coding agent (Claude Code, Codex, OpenCode, and others) to invoke it. The skill installer detects which agents are present and asks where to install; it does not configure credentials. Authentication is handled separately through jg auth, which saves a provider key in an owner-only config file. Supported providers include Vercel AI Gateway, TypeSafe, OpenRouter, and OpenCode Zen.

One practical note on data handling: searches send eligible source content to Jev through the selected provider. Default filtering excludes hidden files, dependencies, binaries, and obvious credential files, but the README explicitly says these filters are not a guarantee. The jg files command lets you audit what a search would read before sending anything.

The tool is narrow by design. When you already know an exact symbol or path, a direct read or ripgrep call is faster. Jevgrep is for the questions that span unfamiliar files — the kind that currently cost agents the most.