Safety & review
Agent action filter
Pre-check on every agent action. Best setup: Jev filters everything and only suspicious actions go to Gemini. That caught all tested attacks at lower cost.
The independent JEV directory
Small decisions.
Big possibilities.
Real ways builders use JEV to turn context into fast, structured decisions. Find an idea, read the original post, and vote for what inspires you.
THE DIRECTORY
Explore what's possible.
Every idea points to the builder who shared it.
Showing 69 ideas
Computer use
Ad stripping
Chrome extension walks the DOM and removes ads in real time.
Agents
Agent dispatch
Replaced the model that routes each Firstmate task to an agent. On 25 tasks it matched Fable's routing in about 200 ms, about 71% cheaper and 90% faster.
Agents
Agent handoff gate
Lead agent verifies worker evidence before accepting a result, instead of re-reading every job. After about 10 hours, lead-agent token use was about 60% lower.
Agents
Agent memory retrieval
Used Jev on his custom memory system. He reports 94% fewer tokens and 2-3x faster memory retrieval. He calls it a quick test.
Agents
Agent progress watch
JevScope judges whether a coding agent is aligned, progressing, or stuck. Traces alone do not say that.
Search & data
AI assistant reranker
Shipped Jev to production as their AI assistant's reranker, still behind a feature flag for selected customers. Early numbers: 3.3x faster (390 ms vs 1.29 s p50), slowest call 0.5 s vs 3.1 s, 55% cheaper, 0 errors.
Safety & review
AI review speedup
Ported an AI review feature from GPT models (about 2 minutes) to Jev. Same results in testing, in seconds instead of minutes.
Safety & review
API abuse classifier
Added Jev to DeepAPI. On this benchmark it ignored all legitimate API calls and flagged all abusive ones, for less than $0.001.
Creative & real-time
Atari lab
Open-sourced lab where game state plus typed questions drive the agent, with replayable evidence.
Safety & review
Auto-approve hook
Open-source jev-guard hook for cheap auto-approvals, because built-in auto-approve modes are missing or expensive.
Search & data
Auto-part pair matching
Decides whether two supplier listings are the same part. On 35k items, Gemini was about $23 and about 4 pairs/s; Jev was about $1.50, about 150 pairs/s, and finished in 4 minutes. They agreed on 89%. He frames it as a test, not a stated production deploy.
Search & data
Book-writing triage
Replayed 3,300 real decisions: source ranking, message routing, and already-sent checks. Jev caught 13 support preferences their classifier had dumped in other. Correct-source-first went from 60% to 90%.
Computer use
Browser control
Sends Stagehand's accessibility tree and candidate actions to Jev, which picks the next action. One remote-browser task cost $0.001.
Computer use
Browser-agent action picker
Jev is live on rtrvr.ai, a free browser agent. Recorded tasks took about 40% less time. The LLM plans and writes code; Jev chooses the next browser action. One find-creators task took 40 seconds.
Search & data
Bulk email classifier
Jev classified 500 emails in seconds for 3.5 cents.
Agents
Claude Code rule picker
jev-rules asks one yes/no per project rule in a single Jev call and injects only the rules that score yes into Claude Code. In his recordings Jev judged 12 rules in 367 to 773 ms, about $0.00002 per prompt. Over 8 checkout prompts, delivered rules were about 330 tokens versus about 1,390 for loading all 12.
Search & data
Clinic intake
Dental triage as 25 typed facts plus a code rule table, shared by the phone agent and the mailbox. On 1,240 emails, time per email went from 2.0 seconds to 0.29 seconds. Won 19 of 23 blind disagreements. Confidence bands choose act, review, or human. Silence is the decision for 23% of the inbox.
Safety & review
Code review
Script.it scores the raw git diff with Jev and only calls an LLM when Jev flags something. Zero false positives on their review data, about 50x faster and 100x cheaper than a GLM, Grok, and Gemini ensemble. The ensemble caught more bugs. Jev caught about 3 in 4.
Agents
Codex context handoff
A context-handoff plugin uses Jev to filter old read-only tool records while keeping instructions, file paths, and unfinished work. Ships as a Codex skill, MCP tool, and CLI, open source under MIT.
Agents
Coding-agent auto mode
She had turned auto mode off in her coding agent because classification was slow, and turned it back on now that they are using Jev.
Agents
Coding-agent decisions
28 decisions across sandbox security, search ranking, and accessibility. Tokens dropped from 49,000 to 8,739, about 3x faster, and cost went from $0.24 to $0.0026.
Agents
Coding-agent file filter
Put Jev in front of Claude to filter candidate files in Aition. 33% fewer files to the model, 25% fewer tokens, 23% less cost, $0.037 to score 543 files at 0.38s per file, and 0 loss in recall.
Safety & review
Command safety review
Benchmarked Jev as the shell-command reviewer in Vercel's fx. Up to 18x faster at p95 and more accurate than GPT Luna, and it may become the default.
Agents
Context compaction
Score every tool call and drop the junk instead of a summarization prompt. Matt Van Horn's roundup says a Claude plugin took one session from 1 million tokens to 86,000 in one second. Link is the roundup, not the original build posts.
Linked post is a roundup, not the original build post.
Safety & review
Custom X feed filter
Chrome extension hides engagement bait, empty hooks, and ads before they are seen. Rules are plain English; Jev checks every post in half a second. A follow-up says it cost about $0.002 for 100+ posts.
Agents
Deal and workflow gates
Tested Jev against gpt-5.4-mini, gpt-5.6-sol, and gpt-6-astra on deal-signal detection, request routing, tool selection, memory decisions, and workflow discovery. Jev was 16x cheaper than mini and about 230x cheaper than astra. Stronger on deal-signal detection; larger models were better on complex memory judgments.
Computer use
Desktop computer use
Turns what the driver sees into bounded action IDs and lets Jev pick one. One run: 44.9 seconds and about $0.001, versus Astra at 294.9 seconds and about $1.45. They say that is one run only.
Safety & review
Etherscan scam scanner
jevscan is a Chrome extension that detects malicious addresses and transactions on Etherscan in real time using Jev.
Agents
Failed-test routing test
Tested Jev against GPT-5.5 and a few lines of if/else on one job: deciding what a coding agent does after a failed test. Jev's routing cost was 125x lower than GPT's.
Computer use
Flight search
Wired Jev to Browser Use, with a small LLM only for typing. Finding flights took 7 seconds and cost $0.0039. Open-sourced.
Search & data
Flight Wi-Fi checker
FlightWifi is a Chrome extension that checks whether a flight's Wi-Fi supports video calls, email, or nothing, detects Starlink versus traditional Wi-Fi, and shows the verdict on Google Flights.
Agents
Herd agent router
Built an agent router using Jev plus Herd, and posted a video of it. Light on measured results.
Agents
Inbox sweep triage
Jev is inside Agently's inbox sweep. On the same 47 messages: about 14s to under 1s per sweep, 7x cheaper per decision at the same context, 30% cheaper while reading 5x more of each message. Sweeps every 5 minutes instead of every 55. 38 ignored, 8 sorted to him, 5 with a reply drafted, 1 flagged as unsure.
Agents
Instant tool-call compaction
Uses Jev for instant compaction: score every tool call and drop what is irrelevant, instead of a summarization prompt. This is her own post, not the roundup.
Computer use
iOS simulator control
Added an agent command to AXe, a non-agentic iOS simulator tool. Jev picks the next action from the prompt and the current state.
Search & data
Job-page crawler
Rebuilt a job crawler with Jev: start at a company homepage, find Careers, and score roles against his profile. An LLM version took about 5 minutes; Jev took just over 20 seconds in his test. Packaged as the skill jev-job-hunter.
Search & data
Job-post classifier
Classified work model, seniority, benefits, and location on 40,000 jobs for $4, then deleted the old regex classifier.
Creative & real-time
Kalshi BTC gate bot
Kalshi 15-minute BTC yes/no bot. Jev calls yes, no, or lean; the bot fires at most one order per window with stake locked at $2. Confidence gate is at least 0.55. The judge is paper; the trader uses live Kalshi keys, with a kill switch.
Computer use
Keystroke file launcher
Predictive launcher: Jev reads intent so typing a phrase like the pdf I just downloaded already ranks the newest PDF, with a confidence on every keystroke, in about 100 ms.
Search & data
Legal database router
A classifier test uses Jev via OpenRouter to take legal prompts and choose which database in his legal chatbot to search.
Safety & review
LinkedIn authenticity filter
Chrome extension scores each LinkedIn post on six authenticity checks in one call.
Creative & real-time
Live trading
Jev reads a price feed, chooses buy or sell, and places real orders on Kuru's on-chain book every 300 ms block. Open-sourced.
Safety & review
Multi-site junk filter
Chrome extension classifies X notes and accounts as LLM-written or engagement farming, skips YouTube sponsorships, blocks ads, labels dark patterns, flags fake Amazon reviews, and removes cookie banners.
Search & data
NL filter harness
A production layer around Jev: policy, confidence gate, shadow mode, recipes, and action-level evals. On one 24-row natural-language filter, Claude CLI took 48.9 seconds and Jev plus the harness took 1.3 seconds.
Search & data
Norwegian hearing scorer
First look, not a full benchmark: 24 Norwegian hearing responses. He says Jev reads Norwegian, costs $0.22 per 1,000 docs, and at 0.9+ confidence was right 14 of 15 times.
Creative & real-time
Polymarket suggestion extension
Chrome extension for Polymarket suggestions using Jev. He says Claude did most of the building. Light on measured results.
Search & data
Postgres natural-language filter
jev() is a PostgreSQL extension that filters rows in natural language with no index and no embeddings. 129 rows judged in about 1 second for $0.0009; the second run was 6 ms from cache.
Agents
Prompt routing
Jev picks which LLM handles each prompt. About 20x cheaper, 6x faster, around 140 ms.
Search & data
RAG chunk filter
Retrieve as usual, then run Jev on every chunk and delete the irrelevant ones. Also from Van Horn's roundup. The original post was not opened separately.
Linked post is a roundup, not the original build post.
Creative & real-time
Realtime Minecraft agent
Jev and GPT-6 Astra play Minecraft in real time through his agentic system. They can fight multiple zombies at once. Jev makes the quick decisions; Astra plans ahead.
Creative & real-time
Realtime Pong
Jev at about 200 ms through Vercel AI Gateway plays against GPT-5.6 Sol and Claude Haiku. The ball only moves when the model decides.
Search & data
Resume-job relevance scoring
Benchmarked Jev on HiringCafe (2.5 million monthly active users) scoring resume against job-description relevance. Human-labeled Spearman correlation: Jev 0.79 at $0.02, versus Gemini 3.1-Flash-lite 0.72 at $0.29.
Agents
Routing-threshold tester
Janus tests real traffic to find where Jev should handle easy decisions and a stronger model the hard ones. On his tests: up to 53% lower cost, 80.2% accuracy vs 78.8% with the stronger model alone, 302 ms median latency vs 2.27 s.
Search & data
SEC filings and trial criteria
Two production jobs. On 120 hand-built biotech 8-K cases Jev matched production. On 30,582 real filings it was worse (3.9% wrong vs 0.9%). On trial-eligibility facts it caught 75% of planted errors, with many false positives. Kept as a first-pass filter.
Safety & review
Security decision benchmark
Ran 34 security cases through Jev and four chat models. Jev was 47x cheaper than Claude Haiku 4.5 for the full run and 438x cheaper than Claude Opus 5, at 0.29 seconds per decision versus Opus 5's 10.81. Accuracy: Jev 14 of 20 incident decisions and 9 of 14 trace decisions. Under four adversarial probes Jev produced zero downgrades.
Safety & review
Security workflow triage
Working Jev into security workflows for incident response and threat hunting: ranking hunt results, scoring hosts, classifying noisy activity, labeling false positives. He does not claim these are shipped yet.
Agents
Software-factory supervisor
Open-sourced Foreman, a software-factory foreman built with Jev. Coding agents work the floor; Foreman assesses progress, completeness, tests, drift, and verification, and intervenes when needed.
Agents
Task router
Alan's task router now runs on Jev for faster routing. Lighter on measured detail than the others.
Creative & real-time
Town policy simulator
Type a law, Claude invents the reactions, and Jev decides how each of 40 citizens responds. The town sim plays those choices out.
Creative & real-time
Trading-signal judge
Demo where Jev reads dozens of structured trading signals and shows a decision. He says it is still experimental and there are no real trades yet.
Agents
Two-call agent router
An agent router on Jev. Call 1 scores department, urgency, and frustration, then code picks the agent. Call 2 scores complexity and stakes, then code picks the LLM and reasoning depth.
Search & data
Viral post classifier
Jev plus SuperX scores posts for virality. Every post gets 61 questions in about 1 second for $0.0004, fitted on 9,481 posts from 207 creators. It picks the viral post 2 in 3 times. Only the text goes into the classifier.
Computer use
Voice browser control
Controls the browser with his voice. The transcript goes to Jev, which returns probabilities in about 300 ms, and the browser clicks. Cost is $0.0002 per decision.
Computer use
Voice computer assistant
Local Whisper plus Jev classifies what he says. A small Swift app passes the screen's accessibility tree so Jev can decide what to click.
Computer use
WebMCP browser benchmark
Jev plus Mercury 2.5 using WebMCP solved 100% of their tasks at roughly 112x lower model cost than GPT-6 Astra with computer use and code execution. Without WebMCP their Browser Use setup solved 25/49; with WebMCP, 49/49. Jev picks the tool; Mercury generates arguments.
Agents
Workflow question classifier
Jev is now available in Dify's Question Classifier node, used to route inputs across workflow branches.
Agents
Xal context compaction
Shipped optional Jev compaction to Xal. He says it now compacts super fast.
Safety & review
YouTube slop filter
Bring-your-own-key Chrome extension that scores and filters AI slop while you scroll, and caches the results.
GROW THE DIRECTORY
Built something with JEV?
Share your X post and help someone find their next idea. Suggestions are reviewed before they appear.