Agents: read the raw Markdown of this page, or start at llms.txt.
~/wiki/ideas
Tools: CLIs, linters, semantic grep and dev tools
TL;DR Command-line and CI tools that ask Jev questions: general CLIs, linters and scanners, semantic grep/sort/search. Unaudited community code, mostly days old (★ and push as captured 2026-09-23 to 25). Numbers are the authors'. Names collide heavily: check the owner before you install. Split from Tools and integrations hub: MCP servers, plugins and connectors on 2026-09-25 (nothing dropped); agent gates and screens are on Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks).
Pattern IDs: Decision patterns from the community (with fit verdicts). Licence MIT unless stated. "Our check" = our 2026-09-25 source and registry vetting (docs/sweep/2026-09-25-mrjev/partc.md), not a raw capture.
Name collisions. At least six tools install a command called jev (model-clis/jev, jevkit, shaharia-lab/jev-cli, jev-repl, Nasrallah-AL's jevctl, utk2103/jev-studio): run which jev before trusting it. jev-cli is three projects (crates.io: shaharia-lab; PyPI: tumf; npm jevctl: Nasrallah-AL); mattn/go-jev also ships a jev-cli (Community SDKs and libraries). jgrep is kyu1204 (npm jevgrep) or keltokhy (PyPI jev-grep); jevgrep is allebee (jevgrep-cli), nassim-arifette (@nassim-arifette/jevgrep) or kyu1204's npm package. uehaj's semgrep binary shadows the Semgrep tool. Alurith's jeff also collides with other tools (our check).
General CLIs and REPLs
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| jev CLI (model-clis) | one-question commands, --assert → exit 3, presets, map over JSONL |
brew, scoop, curl | 0 · 2026-09-18 | exit 0 ≠ "yes"; its "roughly 20 requests/second" = the documented 1,200 rpm (20 × 60) verified; --concurrency default 4, author advises ≤ 8. P10 |
jev (shaharia-lab; crates jev-cli) |
Rust CLI: yes/no, choice and rubric subcommands; exit codes and --abstain-band for CI; YAML batch files; jev mcp serve; the key is never a flag |
install script, brew install shaharia-lab/tap/jev; claude mcp add jev -- jev mcp serve |
27 · 2026-09-25 · Apache-2.0 | 255 options and 2–10 levels match Choice questions, Score questions (verified); "about a hundredth of a cent" per question unverified (its own 42-token example shows $0.000002). Installers check SHA-256 (our check). P10 |
| jev-axi (Nicholas Underwood) | pick/rate/check/rank/diff/triage/guard; setup safety hook |
npm i -g jev-axi |
19 · 2026-09-19 | file ranking did not cut agent cost. P03, P11, P16 |
| jevkit (Ariel Frischer) | Rust CLI (command jev); jev lint: 13 offline rules |
install.sh | 3 · 2026-09-19 | |
jevctl (Nasrallah-AL; command jev) |
verify, screen, classify, extract, find, rerank, match, route, compact, batch; exit 2 when a --fail-on condition matches |
npm install -g jevctl; plugin claude plugin install jev@jev-cli |
21 · 2026-09-25 | the plugin swaps Claude Code's compaction summary for keep/drop per tool call, adapted from tamaratran/fast-jev-compaction; it sends the transcript to Jev and needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 (undocumented, unverified). Its 25k state / 30k request caps sit under 32k/64k (verified, Models, aliases, pricing, rate limits, context). P07 |
| jev-repl (aoprisan) | terminal REPL and sketch page with offline mock answers; :turn/:trend re-ask as a conversation grows; jev eval scores a saved request on labelled cases |
npx jev-repl |
0 · 2026-09-25 | its cost example uses user-supplied "$0.20/$1.00 per Mtok"; as Jev's price that contradicts docs ($0.042/Mtok input, output free). P42 |
| wellposed (suraj-phanindra) | pre-send linter for requests: 45 offline structural checks, then Jev for what structure cannot decide; agent skill | npm wellposed |
2 · 2026-09-24 | a Choice without a "none of the above" option answered wrong at confidence 1.00 (author); the fix matches Choice questions (verified) → Failure reports: where Jev broke, lost, or was the wrong tool. P42 |
| jevcache (hyperspaceai) | cache on redacted (model, schema, state); CI replay |
curl | 71 · 2026-09-19 · no licence | answers not bit-exact (Measurements, access routes and open replicas) |
| Jeview (andududu) | local gateway on 127.0.0.1:4777 with a live map of every call; an events id links calls into chains |
clone, Node 24 | 53 · 2026-09-21 | not a transparent proxy: it adds events to Jev's JSON. Calls and your key sit in a local jeview.sqlite, readable by your user only, so any process running as you can spend it. "About ten cents an hour" demo cost unverified |
Names only: jev-studio (utk2103; PyPI jev-studio: a jev CLI with exit code 2 on a tripped check, MCP server, cookbook slash commands; it checks PyPI for updates, our check); jev-cli (tumf; PyPI jev-cli, dependency-free Python CLI plus stdio MCP server jev-mcp; its Vercel v4/ai/evaluation-model endpoint is unverified, our check).
Linters and code or prose scanners
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| perch (lakeday-org) | method-level bug scanner: ranked defects with severity and confidence; perch check <id> re-asks after a fix; custom rules in perch.yaml; skills for Claude Code, Codex, pi, Cursor |
npm install -g @lakeday/perch; key in PERCH_API_KEY |
179 · 2026-09-25 | no accuracy published. Unscoped npm perch is an unrelated 2016 package (our check). P10 |
| lintus (virolea) | lint rules in plain English (YAML), one Noul per rule per file, all of a file's rules in one request; whole tree, diff or staged; GitHub or JSON output | curl … install.sh | sh (checksum-verified, our check); crates lintus |
13 · 2026-09-20 | author: every file was flagged at ~0.15 until the criteria were reworded; fix wording, not the threshold. Default 100000-byte file cap ≈ 25k tokens at ~4 characters per token (our arithmetic); dense code tokenizes worse and the rules count too, so a file near the cap can exceed the 32k state + longest-question budget (inferred). No accuracy. P10 |
| patdown (tyler-dot-earth) | lints a tree against fuzzy rules in one markdown file; Pi write steering; Claude Code Write/Edit hook; GitHub Action | npx patdown; pi install npm:@patdown/pi |
15 · 2026-09-24 · 0BSD (GitHub: NOASSERTION) | author's direct-API probe (2026-09-18, jev-1.13.0, 14 probes): succeeded at 32,808 reported input tokens, failed from 97,536 bytes with HTTP 400 max_tokens_exceeded; 131,072 spaces cost 2,363 tokens → Request mechanics: billing, limits, latency, calibration and stability. The code is not in the docs we hold (unverified). P10 |
| Sniff Test (DanRWilloughby) | prose linter for AI writing tells: regex rules locally, one Noul per rule per paragraph; 0.4–0.6 = "no judgment" | npm install --global snifftest |
31 · 2026-09-18 | author, 166 paragraphs, 2026-09-17: Jev 63/80 faults, 1/54 false alarms, $0.0129 per 100 paragraphs, 182 ms; Haiku 4.5 66/80 with 37/54; Opus 5 77/80 with 0/54 at $3.08. Rule missing from Jev's reply on 13 paragraphs → Head-to-head benchmarks: Jev on public datasets and suites. P30 |
| Supercov (supercorp-ai) | coverage tool adding Jev quality and security scores per file; content-hash cache; Claude Code plugin | npx supercov quality |
120 · 2026-09-25 | vendor: about a cent per megabyte of source (1 MB ≈ 250k tokens at ~4 characters per token × $0.042/Mtok ≈ $0.0105, our arithmetic, consistent); your source goes to TypeSafe. P10 |
| oxlint-plugin-jev (wobsoriano) | Oxlint rules as English yes/no | npm i -D oxlint oxlint-plugin-jev |
58 · 2026-09-21 | P10 |
| jev-lint (mizchi) | ast-grep + one question; 99 rules | npx -y jev-lint check |
74 · 2026-09-22 | P10 |
| slop-grader (lukstei) | every rule × every line, feeds an agent fix plan | npx @lukstei/slop-grader |
20 · 2026-09-22 | P30 |
| Every (sufianetaouil) | one yes/no per function in a codebase, ranked | pip install every-cli |
8 · 2026-09-17 | 1,302 gin functions in 3.7 s for $0.018; self-test 10/10 recall on 20 labelled functions; its "waitlist" is the author's hosted tier. P17, P10 |
| commit-miner (devanshbatham) | labels commit diffs: bug fix, security/CWE, change type; HTML/CSV reports | cargo install --path . --locked |
36 · 2026-09-17 · no licence | describe, don't reuse code. P18, P35 |
| triagedy (m0rphtail) | security alerts as a Unix filter: JSONL in, typed decisions out; routing in code; local OpenAI-compatible fallback | clone, Rust | 2 · 2026-09-22 | calls Choice confidence "calibrated"; per Confidence vs probability confidence is a statistic of the distribution's shape, not a probability (contradicts docs). Weak benchmarks (author). P23 |
Names only: JevLint (iamtoomas, was huntedman; npm @jevlint/cli; one Noul per file per rule; JS plugins run with full permissions even in --dry-run, our check); jeff (Alurith; Go, read-only lint with a privacy filter; the author's first strict eval run failed its own quality gates → Failure reports: where Jev broke, lost, or was the wrong tool; Apache-2.0); typesafe-ai-playground (markjaquith; Rust experiments: option definitions once in state, catalogues over 253 codes split so each Choice stays ≤ 255, verified); jev-seo (AkashPriyadarshii; crates jev-seo, Rust SEO CLI + MCP server with a per-run Jev spend cap; its README hides a note aimed at AI answer engines and claims a release dated 2026-10-22 that crates.io does not have, so take nothing from that README as fact; our check).
Semantic grep, sort and search
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| jgrep (kyu1204) | semantic grep over code, diffs, CSV; 16 chunks per request; grep exit codes for CI | npm i -g jevgrep (command jgrep) |
34 · 2026-09-24 | init stores the key chmod 600. Not keltokhy/jgrep below. P17 |
jgrep (keltokhy; PyPI jev-grep) |
grep where the pattern is a description: lines, paragraphs, JSONL/CSV fields, diff hunks, whole functions | uv tool install jev-grep |
99 · 2026-09-24 | author via OpenRouter: SMS spam 5,574 messages F1 0.91 for $0.07 vs 17-term regex 0.72; AG News 7,600 articles 86.6% top-1; ~210 ms flat from 1 to 64 questions per call; repeats: 128 of 150 identical, rest moved ≤ 0.03. Warns input text can steer answers → Head-to-head benchmarks: Jev on public datasets and suites. Local-replica run (2026-09-22): Laya's 512-token window failed 300 decisions until records were cropped (Open replicas and Jev-compatible servers). P17 |
| jevgrep (allebee) | grep lines by meaning | uv tool install jevgrep-cli |
1 · 2026-09-21 | author's limits: a self-describing line scored 0.97; a line scored 0.86 in a batch, 0.01 alone (Failure reports: where Jev broke, lost, or was the wrong tool). Not @nassim-arifette/jevgrep. P17, P16 |
semgrep (uehaj; npm @uehaj/semgrep) |
grep by meaning across languages; AND/OR/NOT over meanings; regex prefilter; 30 lines per request | npm install -g @uehaj/semgrep |
144 · 2026-09-25 · MIT text (GitHub: NOASSERTION) | binary shadows Semgrep. Author: --chunk 1 flipped 18 of 200 log lines vs 1 when batched; judges solo verdicts less reliable (the opposite of allebee's result; Request mechanics: billing, limits, latency, calibration and stability); ±0.05 drift; 420 questions in one request worked; P 0.94 / R 0.98 on 51 lines (Claude-judged). P17 |
| jegrep (can1357) | semantic code search, no index: lexical candidates → Jev on short sketches → ≤ 40 full passages | cargo install jegrep |
91 · 2026-09-20 | fails over between OpenRouter and TypeSafe on 401/429/5xx and others; reads keys from ~/.env. Author, 30 queries: files 49/49, regions 93/94, ~$0.005 per query; labels positive-only, so precision is a lower bound. P17 |
| JevGrep (nassim-arifette) | semantic code search CLI + MCP returning exact excerpts with line numbers; asks approval per repo before sending code | npm install -g @nassim-arifette/jevgrep |
72 · 2026-09-23 | local token ceilings: 64,000 direct, 32,000 on Vercel/OpenRouter (its own conservative policy). P17 |
jsort (keltokhy; PyPI jev-sort) |
sorts texts along a plain-English dimension: one Noul per pair, Bradley-Terry fit with standard errors | uv tool install jev-sort |
23 · 2026-09-24 | author: 300 CommonLit excerpts r 0.824 vs ~0.88 ceiling for $0.046; a 2-option Choice pushed 93% of answers below 0.1 or above 0.9 where a Noul left 28% between → Head-to-head benchmarks: Jev on public datasets and suites. Local replicas (author, 2026-09-22): on the 289 excerpts all scored, DiffusionGemma r 0.789, Laya 0.032 (Open replicas and Jev-compatible servers). P17 |
| webctl (dorkitude) | agent web search: Jev scores results from several backends, MinHash + Jev dedupe, keeps relevant chunks | brew, apt, npm install -g webctl, go install |
143 · 2026-09-23 | author, 30 questions judged by Kimi: in Claude Code every webctl arm beat native search (9.17–9.30 vs 8.50) at 19–30% lower cost; under 0.3 points is noise. npm postinstall downloads the release binary with no checksum (our check). P17 |
| jev-reranker (shinpr) | JSON pipe: rerank, filter, extract | npm i -g jev-reranker |
4 · 2026-09-23 | P17 |
jevyoumean (syumai; command jym) |
semantic "did you mean?" for any CLI: one Choice over subcommands plus __none__, sharded above 254 options |
go install …/cmd/jym@latest |
16 · 2026-09-23 | auto-run needs p ≥ 0.95 and never a denylisted verb (rm, push, …); valid commands never touch the network; --all wraps every executable on PATH (author: dangerous). P12, P14 |
| doc-router (misbahsy) | which PDF pages need OCR | cargo | 26 · 2026-09-18 | 155 → 87 pages billed; 9 misses vs 28 for rules. P26 |
Browsers, harnesses and pipelines
| Tool | Does | Install | ★ · push | Note |
|---|---|---|---|---|
| fastbrowse (Agent Labs) | Jev picks actions, LLM reads, cited answers, --authorize gate |
uvx fastbrowse |
98 · 2026-09-22 | 41/42 vs Browser Use 39/42; median $0.0041 vs $0.3668/task. P12 |
| jev-browser (Tomohiko Hiraki) | Playwright SDK, CLI, MCP | release tarball | 7 · 2026-09-22 · Apache-2.0 | not jkudish/jev-browser (Tools and integrations hub: MCP servers, plugins and connectors). P12 |
| System One Harness | finite actions → questions, gated steps, traces | pip install -e . |
124 · 2026-09-21 · Apache-2.0 | P24 |
| evoke | sentence → program + args; runs, confirms or asks | curl; npm | 11 · 2026-09-23 · Apache-2.0 | P14 |
| firehose-judge (Leo Mata) | Cloudflare template: 8 questions per post, human lane | wrangler deploy |
0 · 2026-09-19 | ~260 ms, ~850 tokens/post. P16, P06 |
Elsewhere: semdecide, blink, jev-shell-history → Repos: coding agents, orchestration, memory, browser and computer use, integrations.
Related
- Tools and integrations hub: MCP servers, plugins and connectors — hub: MCP servers, plugins, connectors, agent routing
- Tools: guardrails for agents (tool-call gates, permission hooks, injection screens, rule checks) — tool-call gates, permission hooks, injection screens
- Community SDKs and libraries — SDKs by language and libraries; Community repos: what people built and how they use Jev — vetting
- Request mechanics: billing, limits, latency, calibration and stability — token limits, batching and stability evidence behind several rows
Sources
Links inline; raw captures (2026-09-23 to 25) in frontmatter. Names-only lines and "our check" notes rest on our 2026-09-25 vetting at source (docs/sweep/2026-09-25-mrjev/partc.md) where no capture exists.