What is jevgrep?
jevgrep is a command-line tool, run as jg, that answers questions such as "Where is authentication checked before a request reaches a handler?" with a list of relevant files, reading leads and verbatim source excerpts with line references, all printed to stdout. It walks the repository from folders to files to declarations and asks Jev, TypeSafe AI's decision model, whether each branch is relevant. It returns evidence for a coding agent to read, not a generated answer.
It is written in TypeScript, licensed MIT, and installs from npm as @dzhng/jevgrep on Node.js 22 or later. A companion skill, installed with jg skill, teaches Claude Code, Codex, OpenCode and other agents when to call it. Declaration parsing covers Python, TypeScript and JavaScript, Go and Rust; other text falls back to plain previews.
Searches call Jev through a provider you choose at jg auth: Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen or a compatible custom endpoint. There is no local or offline mode.
Is jevgrep worth it?
jevgrep is one of the few agent search tools that publishes its cost evidence in full, and the evidence supports a modest claim. In the maintainer's rerun of ten Python SWE-bench tasks with OpenAI's gpt-5.6-sol at medium effort, both the baseline and jevgrep 0.4.3 solved 8 of 10. Total cost fell from $7.62 to $5.66, a 25.8% saving, with $1.48 of that spent on Jev itself at list price. The README's headline of about 30% is the older figure that leaves Jev out, and the maintainer says so.
The method is different from embedding-based search. Instead of building a vector index, jg asks Jev yes-or-no relevance questions about folders, files and declarations, follows the branches that pass, and prints excerpts. Nothing is indexed ahead of time, so there is no watcher to keep current, but every search costs provider calls; validated answers are cached locally to soften repeats. Concurrency defaults to 32 requests across all stages.
The data trade-off is plain. Eligible source goes to Jev through whichever provider you configured, on every search. The default filter skips ignored, hidden, dependency, binary and .env files, and drops whole files that look as if they contain credentials, but the README calls this best-effort detection rather than redaction. jg files previews what a search would read without a key or network call, which is worth running once on any private repository.
The evidence is thin in one respect: each task was run once, the baseline was reused from an earlier run, and the ten tasks were tuned rather than held out, as the results page states. The 0.5.0 run cut Jev spend by about 59% but left combined cost 2 to 3% higher than 0.4.3. The project is six weeks old, with 3,176 stars, 242 forks, releases up to v0.8.0 on 1 October and commits as recent as 4 November.
Against grepai, which embeds code locally with Ollama and can stay entirely offline, jevgrep needs no index and no local model but cannot keep code on the machine. For a team already paying a gateway for its agents, that is a small extra hop; for one with a no-third-party rule, it rules jevgrep out.
Where it earns its place
A backend engineer dropped into a large Python service they have never seen
Find where request retries and timeouts are handled before changing them
Ask jg "Which tests cover retry behaviour when a request times out?" and hand the printed files and excerpts to the agent. In the maintainer's study Jev cost about $0.15 a task at list price.
An engineering manager paying for a 20-seat coding-agent rollout
Bring down the token bill without lowering the solve rate
Install the skill in each repository with jg skill --yes and compare gateway spend before and after; the published study saw 25.8% lower total cost on ten tasks, which is a starting estimate, not a guarantee.
A security-conscious team lead at a fintech
Check what an agent tool would send off the machine before approving it
Run jg files on the repository to list, by directory, every file a search could read, then add --exclude patterns for generated or sensitive paths. It needs no key and makes no network request.
Before you commit
- Setup
- npm install -g @dzhng/jevgrep on Node.js 22+, then jg auth to pick a provider and save its key, and jg skill (Node 22.20+) to install the agent skill. No Python, Bun or ripgrep needed.
- Integrations
- Skill installers for Claude Code, Codex, OpenCode and other agents; Jev through Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen or a custom TypeSafe-compatible endpoint. Output is plain stdout.
- Data and privacy
- Searches send eligible source to the chosen Jev provider. Keys sit in an owner-only config file; validated answers are cached locally; no report files are written. The README describes no telemetry. MIT.
- Cost at scale
- Jev is $0.042 per million input tokens on Vercel AI Gateway with no output charge. In the ten-task study Jev added $1.48 while cutting the coding model's spend from $7.62 to $4.18.
- Limits
- No offline mode or provider fallback; environment-variable credentials are ignored; search times and timeouts are not published; 11 open issues and pull requests.
What’s good
- Published SWE-bench cost study with per-task figures and stated limits
- No index to build or keep current
- Returns verbatim excerpts with line references, not a summary
- jg files previews exactly what a search would send, offline
- MIT licence; one npm install on macOS, Linux or Windows
What’s not
- Every search sends source to a remote Jev provider
- Benchmark is ten tuned tasks, one run each
- Declaration parsing for four language families only
- No jg upgrade command; skill and CLI update separately
Best for
Teams running Claude Code, Codex or OpenCode on unfamiliar repositories who already route model calls through a gateway and want to cut the cost of the agent's file search.
jevgrep alternatives
jevgrep: common questions
- Is jevgrep worth it?
- jevgrep is a well-measured way to cut a coding agent's file-hunting bill by about a quarter, provided you are content to send your source to a model provider for every search. We rate it 3.9 out of 5.
- How much does jevgrep cost?
- Free under MIT. Searches are billed by your Jev provider; Vercel AI Gateway lists Jev at $0.042 per million input tokens.
- What are the downsides of jevgrep?
- Every search sends source to a remote Jev provider; benchmark is ten tuned tasks, one run each; declaration parsing for four language families only; no jg upgrade command; skill and CLI update separately.
- Who is jevgrep best for?
- Teams running Claude Code, Codex or OpenCode on unfamiliar repositories who already route model calls through a gateway and want to cut the cost of the agent's file search.
- How much does jevgrep cost for a team or in production?
- Jev is $0.042 per million input tokens on Vercel AI Gateway with no output charge. In the ten-task study Jev added $1.48 while cutting the coding model's spend from $7.62 to $4.18.
- Is jevgrep safe for company data?
- Searches send eligible source to the chosen Jev provider. Keys sit in an owner-only config file; validated answers are cached locally; no report files are written. The README describes no telemetry. MIT.
- What are the best jevgrep alternatives?
- The closest alternatives we have reviewed are grepai and LangGraph Bigtool.
Answers checked Nov 5, 2026.