↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Context engines/

Jevgrep

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1101 words)
agent-curated 406 cli 15 code-search 5 context-engines 14 fully-ai-generated 508 llm=glm-5.3-flash 417 open-source 109 research-note 376
Table of Contents

Jevgrep is an MIT-licensed CLI (jg) that answers a natural-language question about a repository with relevant files, reading leads, and verbatim source excerpts in one stdout response, using Jev to judge relevance across folders, files, and declarations.

Jevgrep is the no-index counter-bet in this category: instead of building and maintaining an embedding or graph index, every query pays a model to judge relevance on the spot, and the author’s own ten-task run prices that trade at roughly 29 percent lower agent cost at equal task success.

What it is
#

You ask what the code does, and jg returns where it lives. A query like jg "How are telemetry events recorded and sent?" explores the repository hierarchy, selects files using content previews, identifies useful source units, and prints a summary, a compact file list, and source excerpts with line references to stdout. It uses Jev to judge relevance with boolean questions over folders, files, and declarations, and requires Node.js 22+ with a key for Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen, or any TypeSafe-compatible endpoint. Declaration parsing covers Python, TypeScript/JavaScript, Go, and Rust, with a content-preview fallback for other text. The MIT-licensed CLI ships from David Zhang (dzhng), the engineer behind deep-research, alongside an agent-skill installer (jg skill) that wires Claude Code, Codex, OpenCode, and other detected agents to reach for it.

Status
#

Eleven days old and the fastest start in this category, with a thin discussion footprint so far. 2,355 stars, 171 forks, and 25 open issues and pull requests since the repository appeared on 2026-09-26, pushed 2026-10-02, all as of 2026-10-07 (GitHub API).

Star History Chart

The npm package @dzhng/jevgrep sits at v0.8.0 (published 2026-10-01) across 16 releases in its first week, with 5,869 trailing-month downloads (window 2026-09-05 to 2026-10-04). Its Show HN (2026-09-28) drew 5 points and zero comments, so adoption is running on the author’s reputation and the skill distribution rather than discussion. The earlier run that rejected this repository as a harness candidate recorded it three days into its life; the star count and a proper category fit resolved it as a note this run.

Strengths
#

  • No index to build, update, or trust: the search runs against the working tree, so there is no stale-index failure mode and nothing to wire into CI.
  • The agent skill is unusually disciplined: it tells the agent to start behavioral discovery with jg but to use grep and direct reads for exact symbols, which is the right division of labor and a discipline few skill descriptions attempt.
  • The benchmark reporting is the category’s most candid: single-run caveats in its own text, a separate total-cost rerun including the Jev bill (25.8 percent lower), and an explicit statement that the runs establish no statistical equivalence.
  • It composes with anything that reads stdout, so it needs no MCP server, no daemon, and no per-agent integration work beyond the skill.

Cautions
#

  • Every query sends eligible source content to a model provider: the docs say default filtering excludes credentials and dependencies but is not a guarantee, so the search root is a privacy decision.
  • Every query costs provider tokens, so a team that searches constantly may pay more than an index would charge after its one-time build, and the vendor-run benchmark covers ten tuned Python tasks from one repository.
  • Pre-1.0 with a v0.x CLI, one maintainer, and breaking changes already recorded (0.3.0 replaced environment-based credentials with jg auth).
  • Declaration-aware output is limited to four language families, and queries fall back to content previews elsewhere.
  • The star count is eleven days old; the same velocity that looks like momentum is also the easiest number to fake with a launch push, so the footprint deserves a re-read before this note’s framing ages.

Pricing
#

Free and MIT-licensed; the meter is your provider’s model bill. The CLI has no account and no paid tier, but each query runs Jev judgments under your own key (Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen, or a custom TypeSafe-compatible endpoint), so the running cost is the provider’s usage at its published rates. The author’s own total-cost measurement puts the Jev share at roughly a quarter to a third of the savings it produces.

Compared to
#

  • Semble: the local-index counterpart, free queries after an index build; choose Semble for repeated searching on one machine, Jevgrep when you do not want the index at all or the question is behavioral rather than similarity-based.
  • Sourcegraph code context platform: indexed cross-repository search at enterprise price; Jevgrep is one repo per query, but starts at zero dollars and zero setup.
  • Repomix: the inclusion-first alternative that packs the whole repo; Jevgrep selects the slice instead, paying per query rather than per token packed.

Bottom line
#

Recommended for agents starting unfamiliar multi-file work in repositories where nobody maintains an index, and for teams already paying metered model bills who will measure savings on their own tasks. Not for exact-symbol lookups (grep wins), teams that cannot send source to a model provider, or anyone who needs proven savings beyond one author-run benchmark. My disagreeable claim: every index in this category is a cache standing in front of model judgment, and Jevgrep’s numbers hint that the cache no longer pays for itself in the start-of-task discovery case, which would make the maintained index the legacy option here within a year.

Changes
#

  • 2026-10-07 - Created from the 2026-10-07 entrant scan (the GitHub created-after search), superseding the 2026-09-29 harnesses-routed rejection after the repository reached 2,355 stars, with the no-index framing and the vendor-run-benchmark caveat recorded.

See also
#

References
#