Seven free, mostly-unknown GitHub repos that plug TypeSafe's Jev decision model into Claude Code and Codex to make routine agent decisions dozens of times cheaper and faster.
Jev is a decision-only AI model, not a text generator, and several new open-source repos now let you plug it into Claude Code or Codex to replace expensive LLM inference for yes/no and scoring decisions.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You're building Claude Code or Codex agent workflows that make lots of small yes/no or scoring decisions, like fact-checks, ad detection, click targets, or candidate screening.
You want to cut agent latency and per-step API cost without touching the reasoning model you already use for planning.
You're comfortable installing an MCP server or cloning a GitHub repo and wiring in an API key.
SKIP IF…
You're looking for a general-purpose LLM replacement. Jev doesn't write text, plan, or hold a conversation.
You don't currently run any agentic workflows that make repeated automated decisions.
TL;DR
The full version, fast.
Jev is TypeSafe AI's decision-only model: instead of generating text, it assigns a probability to a fixed set of options and returns a typed judgment in well under a second, for a fraction of a cent. This video walks through seven free, low-star GitHub repos that plug Jev into Claude Code or Codex: a UI-morphing text box, a voice-controlled browser, a computer-use agent that replaces screenshot-based clicking, an ad blocker, an MCP server exposing eleven judgment tools, a live 'BS meter' for video, and a directory of community-built use cases. The through-line: keep Claude or GPT for planning and reasoning, and route every fast yes/no, score, or selection to Jev to cut both latency and cost.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Cold open framing Jev as a non-text-generating decision model, 40-200x faster than a typical LLM, with a promise to show installable, low-star GitHub repos.
01:37 – 03:27
02 · Repo 1: Shapeshift — a text box that becomes any UI
A single input field morphs into an event card, timer, calculator, or color picker as you type, using Jev to classify intent offline in real time.
03:27 – 04:44
03 · Repo 2: Jev Voice Browser — commanding a browser by voice
Jev decides intent and target roughly every 300ms per spoken word, so a real Chromium browser starts acting on a voice command before the sentence finishes.
04:44 – 07:22
04 · Repo 3: TypeSafe Computer Use — clicking without screenshots
Aaron's repo replaces the screenshot-then-predict loop of typical computer-use agents with OCR plus classification, cutting cost and latency by more than an order of magnitude versus Opus 5.
07:22 – 08:04
05 · Repo 4: an undetectable ad blocker
A browser extension that classifies every DOM element live as ad or non-ad and removes it; presented as the least practical repo of the seven, mainly as a demonstration of Jev's speed.
08:04 – 11:06
06 · Repo 5: jev-mcp — eleven judgment tools for Claude Code
An MCP server (jkudish/jev-mcp) that gives Claude Code or Codex eleven typed judgment tools (verify, screen, find, gate, and more), demoed catching a prompt-injection attempt and fact-checking pricing claims for fractions of a cent.
11:06 – 12:49
07 · Repo 6: Jevmeter — a live BS meter for any video
ChetasLua's repo scores every sentence of a video with Jev and renders it as a 16:9 overlay, demoed against a presidential debate clip for under five cents total.
12:49 – 14:25
08 · Repo 7: shipwithjev.com — a directory of use cases to mine
Not a GitHub repo but a categorized directory of 551-plus community-built Jev projects, meant to be pasted into Claude Code alongside your own context to surface applicable ideas.
14:25 – 15:21
09 · Where to start
Of the seven, the presenter says jev-mcp is the most immediately practical because it plugs straight into an existing Claude Code or Codex session; suggests auditing past sessions for places Jev could have been used.
15:21 – 15:34
10 · Outro
Sign-off, pointing back to the free Skool community and the one-prompt setup guide in the description.
Atomic Insights
Lines worth screenshotting.
Jev isn't a language model. It doesn't write text, it assigns a probability to a fixed set of options and returns a decision in milliseconds.
Jev can be 40 to 200 times faster than a typical LLM like GPT or Claude at making a decision, because it skips text generation entirely.
TypeSafe's computer-use repo replaces screenshot-based clicking with direct OCR and classification, cutting an agent's per-step cost from around $0.05 with Opus 5 to about $0.0002 with Jev.
A Jev-driven computer-use step runs in 0.13 to 0.38 seconds versus roughly 5.2 seconds for a screenshot-based Opus 5 step, a 14x to 40x latency drop.
The Jev Voice Browser controls a real Chromium browser by voice and starts acting on a spoken command in about 300 milliseconds per word, often before the sentence finishes.
The jev-mcp server exposes eleven typed judgment tools to Claude Code and Codex, including jev_verify, jev_screen, jev_find, and jev_gate, each returning a confidence score in 150 to 500 milliseconds for a fraction of a cent.
In a live jev-mcp demo, four fact-check verdicts run against a page with an injected hidden instruction cost a combined $0.00062, and the injection itself was caught and blocked automatically.
A creator built a live BS meter that scored 1,191 Jev calls across a presidential debate for a total cost of $0.0497, producing 5,955 yes/no verdicts.
Jev is positioned as an addition to Claude and GPT, not a replacement: use frontier models for planning and strategic reasoning, and Jev for any decision that can be reduced to a yes/no, a score, or a single choice.
Business-development teams can use Jev to score cold-outreach candidates on person fit, company fit, and criteria match, then only hand the highest-probability matches to the LLM for outreach copy.
A high-frequency trading bot assigned a 91% buy probability to a trade decision in 400 to 500 milliseconds, a speed only possible because it skips LLM inference.
Most of the featured repos have under 1,000 GitHub stars because Jev-based tooling is new enough that the ecosystem hasn't caught up yet.
shipwithjev.com catalogs 551-plus use cases built with Jev, organized by category, so builders can paste the URL plus their own context into Claude Code to find where Jev fits their existing product.
An 'undetectable' ad blocker built on Jev classifies every DOM element live as ad or non-ad and removes it, but even its own creator says purpose-built ad blockers still outperform it.
Takeaway
Route Decisions to Jev, Not Claude
JUDGMENT VS REASONING
Jev's real innovation is turning agent decisions into priced, timed events instead of blank checks, so cheap deterministic judgment calls stop consuming the same expensive inference budget as genuine reasoning.
02Repo 1: Shapeshift — a text box that becomes any UI
A single text input can replace a whole menu of small apps (calendar event, timer, unit converter, bill splitter) if something classifies intent fast enough to render the right UI as you type.
Fast intent classification can run fully offline with a built-in keyword classifier, so a tool like this needs no account and no API key to work at all.
03Repo 2: Jev Voice Browser — commanding a browser by voice
Voice control of software only feels instant when the model decides what you meant before you finish speaking, not after; a ~300ms-per-word decision loop beats typical multi-second voice-to-action lag.
Pairing fast intent detection with an existing automation layer (Playwright here) is a cheaper way to add speed than trying to make the automation layer itself faster.
04Repo 3: TypeSafe Computer Use — clicking without screenshots
Screenshot-then-predict is the bottleneck in most computer-use agents: every click requires processing a full image before deciding where to go next.
Reading the page's actual content (OCR or DOM) instead of a screenshot cuts both the cost and the latency of a single agent action by more than an order of magnitude.
A cost-per-decision comparison is a better way to evaluate an agent architecture than a per-token price, because it captures the full round trip, not just model pricing.
05Repo 4: an undetectable ad blocker
Not every fast-classification demo is the most practical use of the underlying technology; the creator himself says purpose-built tools still beat a novelty implementation.
Live per-element classification of a webpage (ad or not-ad) is a workable pattern for any tool that needs to make a running decision about DOM content as a page loads.
06Repo 5: jev-mcp — eleven judgment tools for Claude Code
Judgment work (verify a claim, screen content before it enters context, pick the best candidate, gate a task as done) is a distinct category from reasoning work, and routing only the judgment calls to a cheaper model is where the savings come from.
A prompt-injection attempt embedded in a webpage was caught and blocked automatically because the judgment layer scored the page's hidden instructions before any content reached the agent's context.
Four verification calls against a page, three claims plus a safety screen, cost about six one-hundredths of a cent combined, the kind of cost that makes it reasonable to fact-check everything instead of nothing.
Installing a new capability into an existing coding agent can be as simple as a single npx command that the agent runs for you, rather than a manual setup process.
07Repo 6: Jevmeter — a live BS meter for any video
Scoring every sentence of a live video feed in near real time, a 0.4-second median per call, is now cheap enough to build as a side project, not an infrastructure investment.
A live scoring tool run across a full presidential debate cost about five cents total for nearly 6,000 individual judgments, a cost structure that was not realistic with a frontier LLM.
The creator's own disclaimer is worth taking seriously: the tool shows a model's probability estimate, not a verified fact-check, and both speakers get graded on the same fixed set of questions.
08Repo 7: shipwithjev.com — a directory of use cases to mine
A directory of hundreds of real, shipped use cases is more useful for idea generation than a blank prompt, because you can filter by category before asking an LLM to map ideas onto your own business.
Feeding an LLM your own context (skills, goals, existing product) alongside a curated idea list produces more relevant suggestions than asking it to brainstorm from nothing.
09Where to start
Of the seven repos shown, the presenter rates the MCP judgment-tools server as the most immediately practical, because it plugs directly into the coding agent you're already using.
You can audit your own past agent sessions and ask the LLM to flag which steps could have been offloaded to a cheaper judgment call, turning this from a one-time demo into an ongoing optimization habit.
Glossary
Terms worth knowing.
Jev
A decision-only AI model from TypeSafe AI that assigns a probability to a fixed set of options instead of generating text, built for fast yes/no and scoring calls rather than open-ended reasoning.
MCP (Model Context Protocol)
A standard way for a coding agent like Claude Code or Codex to call out to an external tool as a typed function, used here to expose Jev's judgment calls inside an existing agent session.
Computer use
An AI agent capability where the model looks at a screen and decides where to click or type next; the traditional approach takes a screenshot and runs it through the model before every action.
Typed judgment
A response format where a model returns a structured probability or verdict, like a confidence score, instead of free text, making the output easy for code to check without further parsing.
TypeSafe AI
The company that built Jev, positioning it as a fast, cheap decision layer meant to sit alongside a reasoning model rather than replace it.
“It's just a way to speed AI up and save money whilst you do it.”
plain-English summary of the entire video in one sentence→ IG reel cold open↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
I feel like everyone in the AI space right now is talking about Jev, a new kind of AI model that doesn't speak. And because of that, it can handle decisions 40 to 200 times faster than a typical LLM like GPT or Claude.
But what people haven't noticed is that there are some amazing free githubs that are now coming out because Jev is still very new that are now enabling us to unlock even more capabilities with the model. And unlike the top Claude githubs, these literally only have 600 stars, 286 stars, 955 stars, and I'm finding some absolute gems.
So in today's video, I'm gonna make Jev useful for you. In the last video, I showed you how I was now using Jev as a filter to make decisions based on probabilities for my business development team. I showed you how you can use it for SEO, how I was using it to build a low latency trading bot, because instead of spending time and energy on inference, now it could just assign a probability to an outcome and execute a trade.
But if you watch that video and you still didn't know how to get started, this video is going to be super practical because these are githubs that you can literally download right now, plug into your cloud code or your codex and Start combining these LLMs with Jev. And as I spoke about in yesterday's video, I truly feel like if you combine the top models like Opus 5 .5, which is an extremely smart model with Jev for decision -making, you can unlock crazy power and you can also save a lot of money.
And by the way, I'm going to make it easy for you. If you want to set up these GitHubs in my free school community down below, I'm going to leave a prompt, which you can just copy and paste into your cloud or your codex. And it's going to go ahead and install all of these GitHubs for you.
So in brand new sessions, you're going to be able to immediately start. Start using this stuff. It's under the free assets library of my school community, alongside all the assets from all of these videos.
Okay, so the first GitHub, which is very cool, is called Shapeshift, which is very cool because it's a text box that actually morphs into a UI as you start typing. So it could be an event card, a checklist, a timer, a color picker, a bill splitter, a poll, a converter, or more. And the reason why I can do it so quickly is because of how Jev works.
You essentially enter in your command instead of... the LLM having to do inference. It simply assigns a probability to determine what you are trying to ask the model for, and then it can create an intent.
So an event extremely quickly, which can then develop the UI of your choice. This is a very great visual demonstration of how quick the model is. So let's actually follow one of its prompts here.
Let's search up Minecraft diamond, and you could see how it finds the color straight away. You could go dinner with Alex on Friday at 7 PM, and it's already going to start adding a calendar event. This is all happening in real time.
It's not like, you know, Claude needs to process something. You could do a timer, so timer 25 minutes, and it's going to automatically pull up a timer. You can even do math, 500 divided by 5, and it happens automatically.
So on its own, this is a cool demo. But imagine if you incorporated this into your workflows or into your apps. Like imagine if you were trading and you went buy $500 of Bitcoin every time price drops 5 % and it uses that to trigger the order automatically.
Like that kind of latency is really powerful. And I think this is the type of UI that a lot of the major developers like Google, for example, with calendar invites or Meta with DMing contacts, this instant UI generation, I think is really powerful. And I think it's a great example of how Jeff works.
under the hood. So this is one of them. I'm going to get into even more powerful ones now.
The last one is something that can actually make you money literally right now. And I'm going to show you a hack, which you can use if you're interested in a startup around Jev, because it's opening up a whole new raft of possibilities. That one's going to be, I think, really good for the people that are interested in that.
This next one's awesome. It's called Jev Voice Browser, and it allows you to control a real browser by voice, which obviously you can already do with Claude and GPT. But this is just so much faster because as you...
saying it it's already acting it doesn't have to do that initial inference part before it loads up a web page so you can literally be speaking into jev you can see an example here he's saying go to wikipedia .com bang it's already on wikipedia now he's saying click on the first link bang you see it already clicks on the first link like this is right as the text is coming up so it's happening immediately eternal blue already popped up go to joshua tree joshua tree pops up so you can actually control it in real time now if you've ever used chat gpt or claude and you've tried to do voice command to your browser you'll know it is very slow and there's like a few second lag if you've ever tried to use a Hermes agent to do voice control there's a couple second lag so this is going to significantly shrink the workflow time on browser use so this is definitely one worth installing because it's super easy once you install it then you can just start commanding your browser by voice there's you know literally no downside to do so so this is a really practical one you can start using now and I was actually filming a demo for you but because I am recording audio for this video my computer just couldn't handle the internal audio recording
for the web browsing and the video recording at once. So I just showed you the demo of someone else, but I literally tested it out and it's lightning quick. Okay, so this next one, which is called computer use, makes computer use so much quicker.
So computer use is essentially when an agent just has to do something on your computer. The problem with ChatGPT and Chord is that every single time they need to click on something on your computer, they essentially need to process it via an image. So they take a screenshot and then they predict where they need to click in order to do something.
But because... of the way Jev thinks, instead of taking five seconds, it only takes 0 .3 seconds because it predicts where to click next based on the command that you've given it. So once again, it's a fundamental architecture shift in terms of how it actually works.
And the stats are quite crazy. If you look at Typesafe's Jev, the price is around 0 .042 per million tokens in output free. Opus 5 is around $5 and $25 per output, which makes Jev 119x cheaper per input token.
And on a cost per decision basis, it's 155 times cheaper than Opus. And on a latency basis, this is what actually matters. I think it's actually less about the money and more about the speed.
It's 0 .13 to 0 .38 seconds per command versus Opus, which, as I said, needs to take screenshots, which is around 5 .2 seconds. So 14 to 40 times faster. And this repo is created by Aaron, who basically just used Jev's technology to create a computer use tool.
So as soon as you... download this, you can start using Jev for computer use. It's crazy.
He's only got 955 stars, but that's because these models are super, super new. And you can see he used it to break down the TechCrunch page and the pricing and amend the website. So instead of having to, you know, screenshot every time like a traditional model and click around, it can just read the back end of the website and, you know, speed up browser use significantly.
Just to be clear, I don't think Jev is a replacement for Claude. I think it's an aid. I think there are some things it does better, some things it does worse.
And as I said in my last video, You want to use Claude for inference. So the planning, the strategic thinking, and you want to use Jev for whenever there is, you know, a probabilistic thing that needs determining because it will basically just determine like a yes or no score based on a percentage.
Jev thrives better with that kind of decision making, which anything that's able to be distilled into a binary outcome, Jev is probably going to do well. But the GitHub's that I'm showing you today, you can literally just chuck into Claude code. So if you're using Claude and then you call upon one of these features, Claude can just call upon.
one thing when it thinks Jev will do really well in that particular scenario. So this is an addition to your existing LLMs, not a replacement in my opinion. Now, this next one isn't something that's that usable in my opinion, and there are other great solutions for this out there, but I thought it was a really good example of how Jev operates.
So I put it in this video and then we'll get into one of the really heavy hitters that I think everyone will get a lot of value from. And this was an ad blocker. So it basically reads the page live, scans if there is an ad, and if there is an ad, you can see it automatically gets rid of all...
of the ads because it determines whether there is ad space on the page. So you can see here, entering another site, detects an ad, deletes it straight away. But it isn't that efficient.
I think there are better ad blockers out there that you can find, but it's another example of the shift in how AI is actually going to be used going forward and the speed benefits that you get. Okay. This one is very cool.
If you want to use Jev alongside Claude code, and this is essentially an MCP tool, which plugs into Claude as a set of judgment tools. So if you're So using Claude, it can pull on Jev to do a variety of things.
For example, verify, checking claims against evidence, screen. judging content before it enters context. Find, picks the best candidate by meaning.
Decide, settles bounded alternatives. So each judgment comes back typed as a probability. So it will give a confidence score to the LLM to save time and money on the inference side.
Because it could be between 150 to 500 milliseconds for a fraction of a cent. And these are cheap mechanical checks that agents will otherwise skip or take too long and will be too expensive to do. Like frontier models aren't good at this stuff.
So this MCP, once you install it, will abstract the experience away. So you can start calling on Jev to do this stuff whilst using your existing agent harness that you're probably already wanting to use.
And once again, that's why Jev doesn't replace Cloud Code or Codex. You want to stay in those harnesses. You want to stay using those models.
As I said, you want to use Jev when it's good at a particular thing. And making these judgment calls is something that it is fundamentally very good at and designed to do. I'll give you a live example.
So I just asked Claude to use the Jev MCP tool and show me the cost for each call. I'm using Jev Verify. to verify a pricing structure for a product.
So you can see Jev here comes back with an outcome based on a probability that it assigns. So it gives me a confidence score, a verdict. based on the preset criteria and that cost me you know 0 .000023 so if you needed to make a decision or if you needed to verify a piece of information is true instead of using Claude to determine that you could use Jev and save money and save time doing so and I think the real value in this isn't just for questions because you know I understand people already have and myself as well their normal way of conversing with AI this is better when stacked on top of an automated end -to -end workflow so if you have a task which requires a variety of steps and decisions, Jeff can plug into that workflow at certain key touch points.
For example, with business development, we scan candidates for our agency and we will essentially scan whether someone is a good fit. And that is based on a few things. Are they the right person at the company?
Is the company someone we're actually interested in working with? And does this candidate match our criteria? And it could give us a confidence score and then develop a curated list of probabilities as to who fits our criteria, feed that information back to the LLM and then...
automate outreach from there, as opposed to having the LLM do all of the inference around the probabilities. Because if you're scanning thousands of contacts, that is going to take a lot of time and cost a lot of money. And that was why in yesterday's video, I used the example of a trading bot, which I built, which is actually a great example of Jev, because high frequency trading relies on a lot of decisions happening very quickly.
You can see here, this was around 400 to 500 milliseconds. This decision was being made in when it comes to news trading, when it comes to high frequency algo trading, you need to make decisions really quickly. Jev can assign a probability in this case, it made a buy call and 91 % probability trade based on the set of rules that you actually give to it.
So that is where it's going to come really in handy over anything else. So it's these workflows where you need fast decision -making and you need it done in a cost -effective manner. And I think anything that involves, you know, mass data is going to fit this criteria.
Okay. The last GitHub I want to show you is I think another very fun example of how Jev works. And then I'm going to get into the best way to make money.
with Jev. I did find a resource which can help you actually do that. So this is quite interesting.
There is essentially a GitHub here called Jevmeter. And Jevmeter can basically come up with like a BS meter live. And this has some interesting applications.
So you can see here, this creator gave the presidential debate access to Jev. And Jev was essentially looking at a set of criteria. Whilst the debate was happening live, it was coming up with a BS score based on every single line that was being said live.
This does have some interesting applications. Like the immediate one that springs to mind is like polymarket and betting markets, because the inference is much faster than Claude. It could probably assign a probability and then place a bet quite easily.
So I think there'll be use cases across those markets. They'll probably be arbitrage, but that's very interesting. But I think there are also many other interesting use cases for live stuff, for analyzing live streams around, you know, consumer data and behavior, all sorts of indexing around when something is said.
For example, if you have a webinar and you're selling a product, when you say a certain thing, does that uptick? to sales, other sales dropping off? What is the language that you used?
What's the probability of that language actually landing a client? There's lots of interesting things that I could see Jeff being used for in a live forum, taking the data in from other live creators. If you're doing like TikTok shop, for example, I could potentially see use cases there.
It's a little bit abstract for now. And this is more of a fun one than maybe a practical one that you'll be jumping into and using right now. But this is a very interesting one.
I think the capability to analyze live. video is very very interesting and i think the use cases for that over time will start to become apparent now earlier i said i would show you a site that i think is good if you are interested in a startup or making money and it's called shipwithjev .com this has a variety of ideas use cases and posts that have been curated because i you know gave you six githubs i'm kind of counting this as the seventh even though it isn't a github it is a repository of information that can help you a lot of people developing their own use cases, their own ideas.
And this website filters them by build type. So you can go agents and browsers. You can see a variety of things people are building in terms of browsers.
If you click on games and real time, you can see games that people are developing with Jev. If you look at trading and markets, you can see trading bots. People are building with Jev.
Content and growth. You can see how people are using Jev to enhance content. And you can see a bunch of...
ideas for tools and apps. So what you could actually do is you could take this URL, you could go into Cloud Code, paste it into Cloud Code and ask it, hey, this is, so give it your context file. This is the context about me, who I am, my strengths, my weaknesses.
These are my goals, which ideas could be interesting to either integrate into my existing business or startup or be used to create a new startup based on my skill sets and interests. And there could potentially be some ideas to actually ship stuff with Jeff, either if you're shipping within. an existing idea that you have.
So this could help you implement Jev into your existing workflows, or this could help you potentially start something new. Remember, Jev isn't an LLM. So I love the Opus 5 .5 model.
I think it's the best model in the world. That is going to help you determine what to do and what you can implement. And then Jev will just be a part of.
the workflows behind everything to speed things up and save money. And that's pretty much, you know, how I would describe Jev. If you're still a little bit confused and you just want to come away with this video with a really simple explanation, it's just a way to speed AI up and save money whilst you do it.
And you'll find a lot of your use cases could probably be sped up and be made lower cost if you implemented Jev. And hopefully some of the GitHubs that I showed you today, some are more fun, some are, you know, actually quite practical, especially the browser use stuff, the computer use, the voice. voice use, that stuff I think is super, super practical.
You can start implementing this today. Probably the most practical one actually is the, and the one I'm getting the most use out of already is the MCP tools, because that is how you can actually just stack it into Claude to get instant outcomes and really start using it right away. And if you don't know how to use this, you could just ask Claude, hey, based on my sessions, could you analyze them and tell me where I could have used Jev based on these commands?
And it will actually show you some ways that you can use it and help guide you as to how you should be using it in the future. I think that's cool as well. Hopefully you enjoyed this video.
I'll keep you up to date with the latest tools and the best workflows and everything. The one prompt setup guide from today's video will be available in the description down below in the free school community. I'll see you in the next one.
Have a lovely rest of your day. Peace out.
The Hook
The bait, then the rug-pull.
Jev isn't another chatbot: it's a model that never writes a sentence, only picks an answer. This video is a guided tour of seven free, barely-discovered GitHub repos that hand Jev off to Claude Code or Codex to make routine agent decisions dramatically cheaper and faster.
Frameworks
Named ideas worth stealing.
00:00model
Jev vs Typical LLM (ELI5)
Typical LLM (GPT, Claude, Fable): you ask a question, it writes the answer word by word, a paragraph of text. Flexible, does anything, slower and pricier, can hallucinate.
Jev: your code hands it info plus fixed options, it picks one answer instantly, a yes/no, a score, or one choice. 40-200x faster and cheaper, can't hallucinate, decides, doesn't write.
The recurring on-screen diagram that frames the whole video: one model writes, the other decides, and Jev handles the millions of tiny decisions behind the scenes.
Steal forexplaining any 'small model routes, big model reasons' architecture to a non-technical audience
08:04list
Jev's 11 MCP judgment tools
jev_verify - checks claims against evidence
jev_screen - judges content before it enters context
jev_noul - returns a calibrated probability for a stated proposition
jev_find - picks the best candidate by meaning
jev_rerank - scores and sorts every candidate
jev_classify - batch-assigns items to classes
jev_decide - settles bounded alternatives
jev_compare - judges how two passages relate
jev_extract - pulls field values with regex plus judgment
jev_review - scores a proposed diff before the task is called done
jev_gate - reviews a patch and verifies completion claims in one call
The full judgment-tool surface exposed by the jev-mcp server (by Joey Kudish), each one callable from Claude Code or Codex once installed.
Steal fordeciding which of your own agent's steps are 'judgment' work worth offloading to a cheaper model
CTA Breakdown
How they asked for the click.
VERBAL ASK
01:30product
“Join My NEW Free Skool Community and Claim the One Prompt Setup Guide”
Mentioned around 90s as where a copy-paste prompt to install all seven repos lives, then repeated at the very end and in the description. Framed as where the free downloadable asset lives rather than a hard sell.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A 13-minute breakdown of the Chinese open-source model that nearly matches Opus 4.8 intelligence at one-fifth the price, and the four-step setup to wire it into Claude Code.
AI Edge's host installs seven free, high-star GitHub repos live on screen, turning stock Claude Code into a social-research engine, a browser operator, an ads department, and a 279-agent company.
Elon's new agent platform runs a team of AI employees on their own always-on computers — tested here as a CFO, an EA, a Dubai real estate scout, and a YouTube research analyst.
A tutorial arguing prompt engineering is dying now that models are smart enough to self-correct — and a walkthrough of Claude Code's /goal, /loop, and /schedule commands with a live website-audit and YouTube-monitoring demo.
Five free GitHub repos, one install each, aimed at five specific failure modes in Claude Code: forgotten context, buried answers, AI-sounding prose, token bloat, and unchecked assumptions.
00:17
00:32
00:41
00:52
01:04
01:10
01:24
01:33
01:51
02:07
02:11
02:25
02:49
03:00
03:12
03:27
03:35
03:45
03:59
04:11
04:19
04:34
05:09
05:18
05:32
05:44
05:56
06:07
06:19
06:31
06:42
06:54
07:06
07:29
07:35
07:49
08:04
08:16
08:27
08:39
08:51
08:59
09:14
09:26
09:37
09:49
10:01
10:12
10:31
10:53
10:53
11:06
11:34
11:46
11:58
12:09
12:21
12:33
12:56
13:08
13:19
13:34
13:41
13:54
14:06
14:18
14:41
14:53
15:04
15:16
13:11
24:49
22:56
21:37
20:35
17:51