jevgrep

Semantic code search for coding agents and humans. Command: jg.

Built by Remotehost. Repository: remotehostai/jg.

Describe the behavior you need; get source excerpts with exact file/line locations, compact excerpts and a report of how much code was searched. Jev evaluates relevance through Vercel AI Gateway. jg exact delegates literal/regex search to ripgrep without changing its arguments or semantics.

Hosted search, email login and CLI connection management live at https://jevgrep.com. Install the CLI and run jg login. Dry-run and exact searches work without an account.

This release does not establish exhaustive semantic search or measured coding-agent task speedup.

Install

Install the CLI:

npm install -g @remotehost/jg
jg --help

The CLI is open source under the MIT license. Hosted inference requires a Jevgrep account.

Requirements

Requires Node.js 22+ and ripgrep. Python syntax parsing additionally uses Python 3. This package contains only the CLI/MCP client. The website and hosted backend are separate projects. Hosted semantic search requires a Jevgrep login; dry-run and exact modes work without a login. Package: @remotehost/jg (public npm package).

Selected source and bounded context are sent to the configured backend for real inference. Filesystem paths remain local. Use --dry-run to inspect the search plan without network calls; add --dump-candidates --json to inspect bounded candidate source.

Search the whole repository

From the repository root:

# Inspect everything that would be evaluated; no model calls or login required.
jg --all --dry-run "where do we reject expired sessions?" .

# Sign in with jg login, then search:
jg --all "where do we reject expired sessions?" .
jg --all --json --limit 30 "retry a failed network operation" .

--all evaluates every eligible snippet in the supplied paths, without a lexical shortlist. With no path it searches the current directory, including when invoked by an agent with non-interactive stdin. Run from the repo root or supply that root explicitly. To search stdin in this mode, supply - explicitly.

The default budget is 20,000 snippet evaluations. Use --max-evaluations N to set a smaller budget; exceeding it fails before any model requests rather than silently omitting code. Discovery still enforces the file, byte and chunk limits below. Ignored files, hidden files, binary files and other reported skips are not evaluated. --all describes search coverage, not guaranteed semantic recall or one shared context window containing the entire repo. Jev evaluates separate snippet batches.

Scans use up to three concurrent requests, with no artificial per-minute quota or fixed delay between batches. Gateway rate-limit responses fail the scan explicitly rather than returning partial results. Ctrl-C cancels work; cached evaluations can be reused on a subsequent run. Latency depends on scope and Gateway response time. Dry-run plans include the estimated request count. minimumRequestSpanMs remains zero for schema compatibility; it is not an estimate of actual completion time.

All-mode returns the top five matches by default. Use --limit (up to 100) to return more; omittedMatches reports matches excluded by result or output limits.

Search behavior

• Discovery uses rg --files, preserving ignore rules and omitting hidden files. -g accepts repeatable ripgrep globs. Explicit files bypass directory ignore rules.
• Every search reads current working files, including uncommitted changes. Duplicate real paths are removed. Binary, non-UTF-8, oversized files and giant/minified lines are reported as skipped; read errors fail the search.
• JavaScript/TypeScript use TypeScript's syntax tree; Python uses the standard AST. Typical functions/methods stay whole. Large declarations and unsupported or malformed languages use overlapping windows. --chunk-lines N selects fixed windows; stdin defaults to individual lines.
• Syntax chunks include up to 2,000 characters of file header/import/constant context to resolve references. That context is not itself the target being matched.
• Local BM25 ranks identifiers, symbols, paths and text, with deterministic vocabulary expansion. Up to 48 candidate snippets go to Jev by default (--candidates, max 256). This lexical shortlist can miss semantically relevant code with unrelated wording.
• --all judges every eligible snippet, with a default budget of 20,000. --broad retains its smaller default budget of 256. --max-evaluations can set either budget up to 20,000. Larger scopes fail before model calls; narrow paths or raise the budget. Discovery itself is bounded to 20,000 files/chunks and 32 MiB of source; individual files are capped at 1 MiB.
• Requests contain at most 16 snippets and 24,000 text/context characters, with up to three API requests in flight. The backend evaluates each snippet in a separate Jev state (up to four concurrent model calls per batch), so neighboring snippets do not supply evidence for one another. Dry-run request counts are API batches; an uncached scan makes one model call per snippet. Transient network errors and HTTP 502/504 responses get one bounded retry; authentication and rate-limit failures do not. Any batch failure fails the whole result instead of reporting a partial search as success. Ctrl-C cancels in-flight requests.
• Probabilities are cached for ten minutes by query, exact text/context, endpoint, account-token hash and client prompt version. Cache files contain only hashes, probabilities and expiration, never source text, queries or tokens. Content edits invalidate entries. --no-cache bypasses it. Model aliases may change within a TTL.
• Matches are ranked by score; overlapping windows are deduplicated. Defaults: threshold 0.5 and limit 5. A score is the product of two independent judgments, that the snippet acts on the requested entity and that it performs the requested operation. It orders results; it is not a calibrated probability, and the default threshold is chosen from a small labelled corpus, not a representative one.
• JS/TS and Python declarations longer than 24 lines are split into the blocks inside them, so a router or a long loop is judged and quoted one behaviour at a time. The split covers every line of the declaration, so coverage is unchanged.
• In a terminal, source lines longer than the window are clipped to one row and marked with …, keeping line numbers aligned. Redirected and --json output is never clipped, so byte budgets do not depend on the terminal.

Output and exact search

jg --json "reject expired sessions" src/
jg --files "retry failed requests" src/
jg --broad --full "release resources on cancellation" src/network/
jg --dry-run -g '*.ts' "billing rules" src/
jg exact -n --glob '*.ts' 'refreshToken' src/

Plain queries search the current directory recursively, even when an agent runs with non-interactive stdin. Supply paths to scope the search. Reading stdin always requires an explicit - (for example, cat file.log | jg "expired" -).

Default output contains up to five matches, each with its path, exact excerpt line range, optional symbol and source text. Excerpts contain at most 12 lines and 1,200 characters. For longer snippets, the excerpt is taken from the narrower block Jev judged relevant when there is one, and a local query-based window picks the lines inside it otherwise; both only locate the excerpt and neither changes Jev’s relevance decision. A match whose path is a test carries kind: "test". Tests often are the clearest evidence of a behaviour, so they are ranked on merit and labelled rather than demoted. One coverage summary follows; scores, request timings and parser details are hidden unless --debug is supplied.

--json and MCP share a versioned result schema:

{
  "schemaVersion": 1,
  "matches": [{
    "path": "src/auth.ts",
    "startLine": 42,
    "endLine": 44,
    "symbol": "validateSession",
    "text": "if (session.expiresAt <= Date.now()) {\n  throw new Error(\"expired\");\n}"
  }, {
    "path": "test/auth.test.ts",
    "startLine": 12,
    "endLine": 13,
    "kind": "test",
    "text": "const expired = token({ expiresAt: past });\nassert.equal(await call(expired), 401);"
  }],
  "coverage": {
    "mode": "shortlist",
    "files": 30,
    "evaluated": 48,
    "eligible": 260,
    "selected": 48,
    "skipped": 0,
    "selectionComplete": false,
    "evaluationComplete": false
  },
  "truncated": false,
  "omittedMatches": 0
}

Terminal output uses aligned line-number gutters and exact source text. Interactive progress updates one stderr line; redirected output contains no terminal escapes. JSON and MCP retain their versioned structured format.

--max-output bounds each result representation to 8,000 UTF-8 bytes by default, including metadata and the CLI trailing newline (range: 1,024–64,000). MCP returns the same bounded payload as structured content and a compatible JSON text block. The protocol envelope and stderr progress are outside this budget. Output is never cut into invalid JSON: excerpts are shortened or lower-ranked matches omitted. truncated signals any shortened excerpt, omitted match, candidate, or diagnostic detail. omittedMatches counts excluded matching snippets, not excluded files. An excerpt's endLine describes its displayed text; excerptTruncated and sourceEndLine identify a shortened source range. --full requests full snippets but still respects the total budget. Increase --limit or --max-output, or read the source range, for more context.

selectionComplete means every discovered eligible snippet was selected. evaluationComplete additionally requires evaluation to finish with no skipped files; it is always false in dry-run mode. Ignored and hidden files are excluded from discovery. No coverage flag guarantees model recall, and a no-match response is not proof that a behavior is absent.

--dry-run prints a compact plan with file/snippet counts and request estimates, without source code or model calls. With --json, it adds dryRun and plan fields to the same schema. --dump-candidates explicitly includes candidate text and context within the output budget, with omittedCandidates reporting the rest. --debug adds match probabilities and bounded diagnostics; neither option is needed for normal agent use.

--files prints only unique matching paths, with truncation notices on stderr. Progress, when attached to a terminal or explicitly debugging, goes to stderr; JSON stdout stays machine-readable. Exit codes: 0 matches/preview, 1 no matches, 2 error, 130 interrupted. jg exact preserves ripgrep output and exit status. To search a query named login, status, mcp or another subcommand, put -- before it. 0.4.0 changes the earlier preview JSON contract; consumers should use schemaVersion: 1, startLine, and the coverage fields above. kind is optional and only ever "test"; consumers that ignore unknown fields are unaffected.

Coding agents

An MCP server is included: jg mcp /absolute/workspace. It exposes search_code with query, scoped paths, result limit, probability threshold, broad search, all-snippet search (all: true), evaluation budget (max_evaluations), dry-run, cache-bypass, max_output, full, debug, and dump_candidates options. It validates workspace boundaries and returns compact source excerpts plus coverage. After installing the client and signing in, register it with Codex:

codex mcp add jevgrep -- jg mcp /absolute/path/to/workspace

Restart the agent session after registration. No global agent configuration is silently modified.

Use semantic search for unknown symbols and behavior-based discovery. Keep ripgrep for exact symbols, regex and exhaustive references; do not alias rg to jg.

Authentication and hosted backend

jg login
jg status

Login opens Jevgrep's browser device approval page. Sign in by email, verify the code matches your terminal, then approve. Credentials are stored at ~/.jevgrep/config.json with mode 0600. Each CLI connection has a separate token that expires after 90 days. jg logout revokes the current token and removes the local login; you can also revoke connections at https://jevgrep.com/account.

Version 1.0.0 promotes the 0.6.0 preview CLI to the default npm release. Search behavior and the versioned result schema are unchanged from 0.6.0. Version 1.0.1 licenses the CLI under MIT.

Version 0.6.0 changes result scoring and chunking, so results differ from 0.5.x. A score is now the product of the two judgments rather than the lower of them, declarations longer than 24 lines are judged as the blocks inside them, and matches under a test path carry kind: "test". The default threshold stays 0.5, which a three-run evaluation on the bundled corpus put between the highest graded hard negative (0.46) and the lowest correct answer (0.55). Scores are on a different scale from 0.5.x, so a pinned custom --threshold should be re-checked; cached scores from earlier versions are not reused.

Version 0.5.0 moves authentication and search to Jevgrep's independent backend. Old Jevgate logins are not migrated; run jg login again. Source and query text are not stored in Jevgrep's database; account, hashed token and usage metadata are. The server calls Jev through Vercel AI Gateway using server-side credentials. This distribution does not include the website, database or infrastructure.

Configuration:

| Variable | Purpose | | --- | --- | | JEVGREP_ENDPOINT | Backend origin; default https://jevgrep.com | | JEVGREP_TOKEN | Explicit token for automation | | JEVGREP_CONFIG_DIR | Login storage directory | | JEVGREP_CACHE_DIR | Probability cache; default ~/.cache/jevgrep |

Saved credentials are bound to the login origin and are not forwarded on endpoint changes. Only HTTPS origins and local HTTP are accepted; requests reject redirects. .env is not automatically loaded by the CLI; use Node's --env-file if needed.

Distribution repository

This repository is a generated CLI distribution. Development, tests and application services are maintained separately. Make changes in the internal source repository. This public repository contains only the client distribution. The CLI is publicly distributed on npm under the MIT license.

Contributions and support

This is a generated distribution repository. Development and tests are maintained internally, and external pull requests are not currently accepted. You can report CLI bugs or request features through GitHub issues. Include the package version, operating system and a minimal reproduction using non-sensitive sample code. Do not include credentials or private source code.

License

The CLI is licensed under the MIT License. Third-party dependencies retain their own licenses. The separately maintained hosted service, backend, and database code are outside this CLI distribution.

Not affiliated with TypeSafe, Vercel or OpenAI.