User repository on GitHub
Smart web search CLI for agents, backed by Jev. Saves a lot of tokens.
- Forks
- 16
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Webctl
A CLI search wrapper that fans out to up to three providers, dedupes the merged results, has Jev score them for relevance against a stated goal, and optionally scrapes and chunk-filters pages so an agent reads less. Saves context, at the cost of a Jev call per batch.
- Stars
- 156
- Language
- Go
- License
- MIT
- Published by
- Community
- Author
- dorkitude
- Last verified
- Oct 2026
About this project
Smart web search CLI for agents, backed by Jev. Saves a lot of tokens. I originally built this to give my Pi/Kimi K3 chat stack something akin to what Claude/Codex already have out of the box, but now I have Claude/Codex using it too... and it's pretty freaking dope. It really saves a lot of the context window. So now it's free for all!
Quoted from the project's README. Read the full README.
- Languages
- Go
Highlights
- Pipeline: providers, dedupe, Jev scoring, threshold, optional scrape, print. Scores run 0–10 with a default cut of 6; --noul and --rubric change the question.
- --scrape --filter-chunks cuts pages into chunks scored for relevance, typically reducing 25 chunks to three or six and reassembling them as the page text.
- Dedupe runs in two passes: normalized URL or title collapse, then MinHash candidate pairs sent to Jev in one batch to decide same content.
- Rate limits use a shared cooldown ladder on disk, from 15 minutes up to 72 hours and then one probe a day, so a throttled provider is skipped.
Watch out
Token savings are conditional and the author says so: snippets are usually enough, chunk filtering pays off on long PDFs, comment threads, transcripts and whole articles, and for a short fact it adds nothing. The cost is a Jev scoring call per result set and per scraped page, so a keyless setup degrades rather than breaks — without a key, webctl falls back to keyless endpoints and DuckDuckGo, which throttle by IP after a few dozen searches a day and trigger the cooldown ladder. Two configuration catches are documented and easy to miss: once any search key is configured only keyed providers are used, so an unset key silently removes a backend, and cooldown state lives in one file under your home directory shared by every process, which matters when you run webctl concurrently. The benchmarks directory is referenced for token numbers, but the figures themselves are not in the README. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It fans out. By default it tries three backends and fuses their rankings by reciprocal rank; you can pin a single provider, take the first that answers, or request every available one. SearXNG or Degoog go first if configured, then whichever providers you have keys for, with Brave as the author's setup default because of its free monthly allowance. The catch: once any key is configured, keyless endpoints and DuckDuckGo drop out of the rotation entirely.
Relevance to the goal you pass, not truth. Results are scored 0–10 and everything under the threshold is dropped, with the default cut at 6. The dedupe step is a separate judgment about whether two excerpts are the same content; if so, the better-scored copy survives and engine tags merge. You can swap the score for a yes/no noul question, define a custom rubric, or skip filtering with --no-filter, which works with no key at all.
A lot on long documents and very little on short lookups, and the README is upfront about the asymmetry. The chunk path fetches the top kept results, strips edge menu runs, chunks text at 2,000 characters with 20% overlap for judging, then trims output to a default 20,000 characters so a result fits an agent tool window. Per page you can request 50,000 characters by default. The author points to a benchmarks directory for the token comparisons, but those numbers are not in the README.
A 429 or a 402 is recorded in a cooldown ladder kept in a file under your home directory and shared across processes. Strikes lengthen the skip window from 15 minutes through 1, 4, 12, 24 and 72 hours, after which the provider is parked at one probe a day. A spent quota starts one step further along than a throttled one, and any success resets the count. Because the file is shared, parallel webctl calls in different terminals will see each other's cooldowns.
Homebrew (brew install dorkitude/webctl/webctl), an apt repository for Debian and Ubuntu, npm install -g webctl, or go install the cmd/webctl module; prebuilt macOS and Linux binaries are attached to every release.
Yes. Jev scores each result 0-10 with a default cut of 6; --min-score 8.5 is stricter, --min-results 5 keeps at least five marked backfilled, --noul asks a yes/no question instead, --rubric defines a custom scale, and --no-filter skips Jev.
By default it scrapes the top three results (--scrape-top), requests up to 50,000 characters per page (--max-chars), chunks at 2,000 characters for judging (--chunk-chars), and trims output to 20,000 characters (--max-output) so a result fits an agent tool window.
More Jev resources like this
1.8kGitHub stars
Rust CLI that gives agents computer use through OS accessibility trees instead of screenshots: snapshot a live UI, act on stable element refs, then re-verify. Ships a jev-desktop skill where Jev picks the operation and target from accessibility-tree candidates.
Jev Demos & Experiments#community#jevbest
Community
lahfir/agent-desktopRust
lahfir/agent-desktopRust1kGitHub stars
Sandboxed virtual machines for agents: each sandbox is a microVM that starts in about 500 ms, keeps files and state between sessions, and runs locally or in Celesto Cloud behind a CLI and Python SDK.
Jev Demos & Experiments#community#jevbest
Community
CelestoAI/celestoPython
CelestoAI/celestoPython75GitHub stars
A Node package that distils Jev-style typed answers into small local classifiers, trained either from saved labels or from a Jev teacher, then served in Node or a browser worker.
Jev Demos & Experiments#community#jevbest
Community
AndrewPrifer/jimothyTypeScript
AndrewPrifer/jimothyTypeScript147GitHub stars
Apache-2.0 open-weight stand-in for a hosted decision model: Qwen3.5-4B with a distilled adapter scores noul, choice and score questions through a systemone-shaped server, with no generated tokens.
Jev Demos & Experiments#community#jevbest
Community
allebee/jevk5Python
allebee/jevk5Python56GitHub stars
macOS agent loop that moves a real USB HID mouse and keyboard under a vision model, with an experimental decision-model lane picking the next operation and screen target from an element table.
Jev Demos & Experiments#community#jevbest
Community
affirmitv/ghosthandsPython
affirmitv/ghosthandsPython38GitHub stars
Standalone CLI skill that drives isolated Playwright Chromium with Jev choosing scoped in-page actions, defaulting to Vercel AI Gateway with TypeSafe as an option.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
hqman/jev-browser-skillTypeScript
hqman/jev-browser-skillTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.