正在加载视频...

视频加载失败

🎇 I made a small Chrome extension: a side panel that drives any site's WebMCP tools with Jev! When you type, on every keystroke it picks the relevant page's tool, fills in the arguments, and tells you how sure it is. Here it is grocery shopping:

76,866 次观看 • 1 个月前 •via X (Twitter)

54 条评论

Sarah Drasner 的头像
Sarah Drasner1 个月前

This demo shows not everything needs a chatbot. Jev can't write words, it's answers typed questions: pick one, yes or no, how much. Paired with WebMCP, though, turns out that's plenty. "got anything gluten free for under two dollars?" doesn't need a conversation, it just needed search_products.

Sarah Drasner 的头像
Sarah Drasner1 个月前

A thing I didn't expect: it's actually a great way to evaluate your WebMCP tools. You can see how every tool gets a live probability. If two keep fighting over a sentence, your descriptions overlap. If an argument comes back shaky, the schema is vague. This could be useful for better tool descriptions.

Sarah Drasner 的头像
Sarah Drasner1 个月前

This is also fast and cheap. I hammered at this thing this weekend, and I spent *checks notes* 32 cents. It's fast because nothing is screenshotting or guessing at the DOM. WebMCP is handing over a menu of named tools with described params, enums, etc. Then a small decision model answers "which tool, which args?" in a couple hundred ms. This is one request instead of larger conversation and scraping loops, pretty powerful!

Sarah Drasner 的头像
Sarah Drasner1 个月前

Extension is undergoing review but you can play with it here, I've open sourced it: And here's the demo site if you want to play around:

Sarah Drasner 的头像
Sarah Drasner1 个月前

I make demos like this so people will be inspired to try things out themselves, especially on new pathways that didn't exist before. If you do make stuff, I would love to see it! 🔥

Sarah Drasner 的头像
Sarah Drasner1 个月前

This is extension is publised now! You can find it at: The demo shopping site is at if you want to try it out. (Don't forget to go to chrome://flags/#enable-webmcp-testing and flip the flag to enabled)

Miano 的头像
Miano1 个月前

Showing the confidence per tool is the clever bit. It turns the agent from a black box into something you can actually debug, and it tells you when your tool descriptions are stepping on each other.

Sarah Drasner 的头像
Sarah Drasner1 个月前

Exactly

David Boyne 的头像
David Boyne1 个月前

Very cool, I'm loving WebMCP at the moment, and combined with Jev, my mind is racing!

Sarah Drasner 的头像
Sarah Drasner1 个月前

It’s a powerful combination!

Luke Haines 的头像
Luke Haines1 个月前

Genuinely great job

James Q Quick 的头像
James Q Quick1 个月前

Ohhhh that's super useful!!! Love that!

Sarah Drasner 的头像
Sarah Drasner1 个月前

Thanks James!

Sasha Sheng (Hiring) 🫶🏼 的头像
Sasha Sheng (Hiring) 🫶🏼1 个月前

Oh wow! So cool!

Sarah Drasner 的头像
Sarah Drasner1 个月前

Thanks! I'm really excited about this pairing! Thanks for all your hard work at TypeSafe AI Sasha

Abdul Wasey 的头像
Abdul Wasey1 个月前

Live probabilities turn tool descriptions into an eval surface. Overlap becomes visible before it turns into a routing bug.

Yashwin Ghosh 的头像
Yashwin Ghosh1 个月前

The bigger idea to pair Jev with tool calling is quite unique! Great work :)

Sarah Drasner 的头像
Sarah Drasner1 个月前

Thanks Yashwin!

juan pineda 的头像
juan pineda1 个月前

Wow that's really impressive!

Seb ⚛️ ThisWeekInReact.com 的头像
Seb ⚛️ ThisWeekInReact.com1 个月前

Nice! Have you tried using a local Jev-like model? Eventually, one that can run in a browser? (Laya)

Sarah Drasner 的头像
Sarah Drasner1 个月前

I started to this weekend! A lot of potential there. cc @etiennenoel

John Rood 的头像
John Rood1 个月前

Adding a tool can break routing to tools you didn't touch. The regression test is whether the old queries still pick the right tools in the new catalog.

Bruno Lourenço 的头像
Bruno Lourenço1 个月前

Doesn't JEV require a finite set of inputs? How can it fill the tool's params? 🤔

Sarah Drasner 的头像
Sarah Drasner1 个月前

The WebMCP tools become the finite set of inputs

Sarah Satonov 的头像
Sarah Satonov1 个月前

on every keystroke tho? thats gonna hammer the model and your battery like crazy 💀 curious if you debounce it or its raw fire-on-type

Sarah Drasner 的头像
Sarah Drasner1 个月前

Not at all, it’s extremely efficient and cheap. But also, I intentionally built it this way for educational purposes- so that you can see what’s happening. You don’t have to build it this way. Exposing the model thinking is the point of the extension.

Sarah Satonov 的头像
Sarah Satonov1 个月前

okay thanks for information

Jatin Garg 的头像
Jatin Garg1 个月前

When tool selection happens on every keystroke instead of at call time, the model becomes the autocomplete engine, not the executor.

Sarah Drasner 的头像
Sarah Drasner1 个月前

It’s still executing. Very quickly too ☺️

Miguel de Icaza ᯅ🍉 的头像
Miguel de Icaza ᯅ🍉1 个月前

Genius

Kizuno18 的头像
Kizuno181 个月前

exposing confidence scores right at the interface level is huge for agentic loops. in headless pipelines we gate tool calls with strict thresholds—if confidence drops below 0.85, it aborts to a deterministic fallback instead of hallucinating arguments into the DOM.

Sarah Drasner 的头像
Sarah Drasner1 个月前

💯

Jonathan Moore 的头像
Jonathan Moore1 个月前

This is amazing. I'm certain WebMCP will only accelerate from here, and at @stripe / @link we're quickly building on top of it.

Joel Nishanth 的头像
Joel Nishanth1 个月前

No way I was thinking of doing this too!!!

Juan F. 的头像
Juan F.1 个月前

Amazing. Love your projects

NotaDEV 的头像
NotaDEV1 个月前

grocery shopping by keystroke, that's a fun one to watch actually work

Maceo (YC P26) 的头像
Maceo (YC P26)1 个月前

I'd want this in the shop's search box, without having to install an extension

Sarah Drasner 的头像
Sarah Drasner1 个月前

If you check out the site, it has this as well:

Sarah Drasner 的头像
Sarah Drasner1 个月前

There’s nothing keeping it from being in page, I made the extension this way to teach *devs* For the in page agent I used Gemini, but I could be using Jev there too (started building this before Jev came out)

LottieFiles 的头像
LottieFiles1 个月前

The confidence meter is such a nice touch. Was that the plan from the start, or something you added once you saw it flip-flopping?

Sarah Drasner 的头像
Sarah Drasner1 个月前

That was the plan! The purpose was to show devs what’s happening

Asaf Mazuz 的头像
Asaf Mazuz1 个月前

Great project! Added it to my Jev-based project directory Hope more people will be able to try it and see how Jev can help them also. Thank you so much for building and sharing this 🙏

Srinivasan K K 的头像
Srinivasan K K1 个月前

I'll try this for sure.

Steve Kaliski 的头像
Steve Kaliski1 个月前

very cool

Nick 的头像
Nick1 个月前

Yessssss! So cool

Mav1972 的头像
Mav19721 个月前

I got hooked on WebMCP, I even did the DevPost WebMCP competition 😆. That motivated me to create this tool

AI Mastery Guide 的头像
AI Mastery Guide1 个月前

grocery shopping demo, cool

Aark 的头像
Aark1 个月前

This is a powerful combo

Shez Malik 的头像
Shez Malik1 个月前

any site's the real flex

Jhon Dennis 的头像
Jhon Dennis1 个月前

每次敲键盘都在选工具 这个交互挺变态的 但准确率一上来 购物这类杂活真会直接被接管

Itay 的头像
Itay1 个月前

time for @Instacart to open up a MCP.

AXIALIS | Desktop App Tools 的头像
AXIALIS | Desktop App Tools1 个月前

Does confidence stay useful as the tool catalog grows, or do semantically adjacent actions start collapsing into the same probability band?

Perpetual Asogwa 的头像
Perpetual Asogwa1 个月前

Did this with webmcp a while back:

Sarah Drasner 的头像
Sarah Drasner1 个月前

Sure. This isn’t about just WebMCP though. It’s about the intersection with System One models like Jev

相关视频

Jev is now multimodal. It can see. Here’s how it works and what you can build with it: Jev is a decision model. You give it information, ask questions, and get answers your code can use directly. Previously, that information was text or JSON. Now it can include images. Start with how a Jev request works. You send a “state,” which is simply the information you want Jev to examine. Then you ask questions about it. Each question specifies the kind of answer you need. → Choice picks from options you define, such as “thumbs up” or “fist.” → Noul returns the probability that something is true, such as whether a hand is visible. → Score rates something on a scale you define, such as how open the hand is. With multimodal Jev, images can be part of both the state and the questions. Put a camera frame in the state, and your questions can ask about what’s visible in it. Put a reference photo in a question’s instructions, and you can ask whether an object in the camera frame matches the one in that photo. Jev can now also identify an object and locate it for you. The new Box type brings a YOLO-like capability to Jev. Describe what you want to find, such as “the hand in this image,” and it returns a bounding box around it. Your app can use those coordinates to highlight the object or crop it. One detail matters here. Box returns coordinates even when the target is absent, so pair it with a Noul presence check before using the result. My demo asks about presence, gesture, and location together. Jev answers these independent questions in parallel, and the app combines the results. These capabilities open up prototypes for gesture controls, locating buttons in screenshots, and checking product photos against visual references. Jev’s focus on fast decisions also makes it promising for interactive applications where response time matters, with image latency worth testing in your own workflow. Images and Box are still experimental and require the images “danger zone” header. I cover how it works, the API, and the demo app in the detailed video below. Chapters 0:00 Intro 0:36 Live demo 1:41 What's covered 2:22 How Jev works 3:06 Question types 3:57 Playground demo 6:59 Demo app walkthrough 10:30 Using Multiple images 11:25 Parallel questions 12:40 Danger zone headers 13:07 Outro If you’re new to Jev and want to understand how decision models work, I’ve written a detailed article about it. It’s quoted below.

Akshay 🚀

26,832 次观看 • 1 个月前