TubeSum ← Transcribe a video

ChatGPT Operator is expensive... use this instead (FREE + Open Source)

0h 22m video Published Feb 21, 2025 Transcribed Aug 5, 2026 N NetworkChuck
Intermediate 10 min read For: Tech enthusiasts and developers interested in AI automation, with some familiarity with command-line tools and Python.
AI Trust Score 70/100
⚠️ Average / Some Fluff

"Delivers on the promise of a free alternative, but the demo is chaotic and includes a sponsor segment."

AI Summary

The video compares OpenAI's ChatGPT Operator with Browser Use, a free and open-source alternative for AI-driven browser automation. The host demonstrates setting up Browser Use locally, tests it with various LLMs, and runs a head-to-head challenge against ChatGPT Operator to see which can complete tasks like finding a Japanese VCR on eBay and purchasing a cloud VPS.

[00:00]
Introduction to Browser Use

The host introduces Browser Use, an open-source project that enables AI to control a browser, positioning it as a free alternative to ChatGPT Operator.

[01:12]
Browser Use Overview

Browser Use is impressive, with a paid version available. It claims to be 2% better than Operator at $30/month, cheaper than ChatGPT Operator's enterprise option.

[02:25]
Use Cases and Setup Requirements

Examples include adding items to cart, writing letters, and integrating with Salesforce. Setup requires a machine (Mac, Windows, Linux), Python 3.11+, and optionally Ollama for local models.

[03:31]
Local vs Cloud Models

Cloud models like OpenAI or Claude are better than local models due to more resources. The host uses a local setup with WSL on Windows.

[04:10]
Installation Steps

Install Python 3.11 via pyenv, clone the web-ui repo, create a virtual environment, install requirements, and install Playwright for headless browser control.

[05:49]
Configuration

Copy the example environment file to .env and edit it to add API keys for OpenAI, Anthropic, or specify an Ollama endpoint.

[06:44]
Running the Web UI

Launch the web UI by running the provided command, then access it at localhost. Choose an LLM provider and model.

[07:55]
Testing with Local LLMs

Testing with Qwen fails due to its limitations, but DeepSeek R1 14B works better. The host demonstrates adding a coffee to cart.

[09:37]
Using Claude with Browser Use

Switching to Anthropic's Claude 3.5, the host tests creating a VPS on Hostinger, showing that Claude performs well and uses the host's browser sessions.

[13:45]
Comparison with ChatGPT Operator

Browser Use works but required hand-holding and lied about completing tasks. ChatGPT Operator also made mistakes, like ordering 11 servers instead of 1.

[14:12]
Speed Test with Cloud Models

Using Claude, the task of finding and adding coffee to cart is completed quickly, demonstrating the advantage of cloud models.

[15:12]
Advantages of Browser Use

Unlike ChatGPT Operator, Browser Use can use your own browser with logged-in sessions, password managers, and settings, making it more powerful for personal automation.

[17:08]
Head-to-Head Challenge

A race between ChatGPT Operator and Browser Use (with Claude) to find a Japanese VCR on eBay. Operator wins, but Browser Use completes the task successfully.

[19:25]
CAPTCHA Handling

Browser Use attempts to solve a CAPTCHA but struggles, highlighting a limitation. The host notes that ChatGPT Operator cannot handle CAPTCHAs either.

[21:33]
Conclusion and Future Ideas

The host encourages viewers to suggest automation ideas and notes the potential for hacking, emphasizing the power of open-source AI agents.

Browser Use is a viable, free, and open-source alternative to ChatGPT Operator, offering flexibility with local or cloud LLMs and the ability to use your own browser. While it has limitations, such as CAPTCHA handling and occasional errors, it is a powerful tool for AI-driven browser automation.

Mentioned in this Video

Tutorial Checklist

1 03:57 Install Python 3.11 using pyenv: pyenv install 3.11
2 04:10 Set Python 3.11 as global: pyenv global 3.11
3 04:23 Clone the web-ui repository and navigate into it: git clone <repo> && cd web-ui
4 04:38 Create a virtual environment: python3 -m venv venv
5 04:53 Activate the virtual environment: source venv/bin/activate
6 05:05 Install dependencies: pip install -r requirements.txt
7 05:19 Install Playwright: pip install playwright && playwright install
8 05:49 Copy the example environment file: cp env.example .env
9 05:49 Edit .env to add API keys or Ollama endpoint: nano .env
10 06:44 Run the web UI: python web_ui.py
11 07:12 Open browser to localhost and select LLM provider and model

Study Flashcards (9)

What is Browser Use?

easy Click to reveal answer

An open-source project that enables AI to control a browser.

00:40

What is the cost of Browser Use's paid version?

easy Click to reveal answer

$30 per month.

01:28

What are the system requirements for running Browser Use locally?

medium Click to reveal answer

A machine with Mac, Windows, or Linux, Python 3.11 or higher, and optionally Ollama for local models.

02:52

How do you install Python 3.11 using pyenv?

medium Click to reveal answer

Run 'pyenv install 3.11' and then 'pyenv global 3.11'.

04:10

What is the purpose of the .env file in Browser Use?

medium Click to reveal answer

It stores API keys for cloud providers or specifies an Ollama endpoint.

05:49

Which local LLM failed in the demo?

easy Click to reveal answer

Qwen (Quinn) failed due to its limitations.

07:55

What advantage does Browser Use have over ChatGPT Operator?

medium Click to reveal answer

It can use your own browser with logged-in sessions, password managers, and settings.

15:12

What was the outcome of the head-to-head challenge?

medium Click to reveal answer

ChatGPT Operator found the VCR first, but Browser Use also completed the task.

18:24

What limitation did Browser Use show with CAPTCHAs?

medium Click to reveal answer

It struggled to solve CAPTCHAs, often clicking the wrong elements.

19:25

💡 Key Takeaways

📊

Browser Use claims 2% better than Operator

This is a bold claim from an open-source project, suggesting it can compete with a major commercial product.

01:28
💡

Local LLM limitations

Demonstrates that smaller local models struggle with complex tasks, highlighting the trade-off between cost and capability.

07:55
💡

Using your own browser is a game-changer

This feature allows for personalized automation with existing sessions, a significant advantage over Operator's isolated browser.

15:12
📊

CAPTCHA handling is a challenge

Shows a real-world limitation of AI agents, which is important for understanding their current capabilities.

19:25
💡

Potential for hacking

Raises awareness of the security implications of AI-driven browser automation.

21:33

[00:00] go find me a Japanese VCR that supports TBC on eBay and add it to my cart. Oh, open a web browser and do this thing.

[00:13] By the way, that's based on a true story. a few weeks back. Now I'm doing this with open AI's operator.

[00:25] so don't be surprised if it's kind of janky. which means you got to be paying OpenAI 200 bucks a month. But I found an open source alternative. It's like this. It'll open up a browser,

[00:40] It's free open source. I'll show you how to use it right now. network check from the future here coming up, and purchase a virtual machine in the cloud and they,

[00:57] Who can do it the fastest? Can they do it at all? I don't know. And by the way, We'll see what happens. The project is called Browser Use Enable AI to control your browser.

[01:12] and I have to say the project is very impressive. It does a lot. just know there is a paid version of this. If we go to their official website, meaning they've gotten some funding and here they're even touting their

[01:28] performance saying they are 2% better than operator at 30 bucks a month. They are still cheaper than chat. GPT operator, the enterprise option, yes, a month. Okay.

[01:42] I mean it made me stop and look at it and go, what? But anyways, use our own local stuff, even local ai. I mean our web browser be accessed via the web or rather,

[01:56] I said that backwards. I need some more coffee. research mode like Chad, GPT is, and it's very programmatic,

[02:10] meaning that if you are really into building AI agents, which I'm getting into, They have some examples here like add G items to your cart and go check out. And it's obviously all done with code on the side. Don't be scared.

[02:25] You can add my latest LinkedIn follower to my leads in Salesforce, Write a letter in Google Docs to my papa thanking him for everything and say, you don't have to know how to code anything and you just want to take it for a

[02:40] I can see one of the projects is called Web ui. This is very easy to set up. Alama. So a couple of things you're going to need to make this happen. First,

[02:52] you'll need some coffee. That's just the rules. I didn't make 'em. Maybe I did you'll need a machine to run this on if you want to run this locally,

[03:04] which is what we're doing right now. So Mac, windows, or Linux. I will be demoing right now this setup on Windows, which will be using WSL, So it's basically Linux and it will be very different from the bare bones of

[03:17] Linux or Mac setup. And if you're like, Chuck, I have no idea what WSL is, that's pretty much all you need. Oh, you know what? I lied. One tool you can use that's completely free, completely local,

[03:31] completely awesome is a llama. Go out to a llama.com, click on download. Have that going. If you want to use open AI or Clot or any of those cloud-based models,

[03:43] it's going to be better than a local model because they've got more resources. My favorite place to be and because I'm in Windows Land, I need to jump into Linux with WSL. I'll launch my Ubuntu, I think it's 2204.

[03:57] the first thing you want to do is make sure you do have Python three point 11 installed at least three point 11. The easiest way to do that is with PI ENV, All you have to do is type in PI ENV installed three point 11.

[04:10] and then you do PI NV global three point 11 to make it live and you can switch Now real quick, make sure you have Python three point 11 by typing in Python three dash dash

[04:23] version and you should see Python three point 11. Now we're going to clone this, and then we'll jump into that directory by typing in CD web dash ui. we're going to launch a Python virtual environment or create a virtual

[04:38] environment. We'll type in Python three dash M for module specify VENV, if you've never used a virtual environment, Pip install a virtual, ah,

[04:53] why is my curs bouncing around virtual ENV, just like that. We'll type in source vnv slash ben slash activate. Boom.

[05:05] going to be impacted by the things we install. Now we'll use the command pip install dash r and we'll type in requirements This is a file that's right here in our directory and it's going to describe the

[05:19] set, go. And we'll watch it happen while we're sipping some coffee And then one more thing we have to install is this tool called playwright, but I think it's essentially doing headless browser stuff. It's amazing.

[05:34] You just might take a moment. And finally, They do have an example environment file that we're going to copy to our own. So we'll type in CP Env example and we'll copy that file to

[05:49] env just like that. Now let's edit that env file nano. And here we can add any kind of API keys we want to have here. So open API anthropic. We can also specify an Alama endpoint,

[06:03] I do have an external alama server that's more powerful. Terry, I've built in this video here. He has dual 40 nineties. It's amazing.

[06:15] But I'll add his IP address and we'll use him for my stuff. going to show you what they feel like. And don't worry, I will end up revoking these keys. So it's okay that you see them right now.

[06:28] Then now all we have to do is run this command. Let's scroll down and find it. They do have a Docker option, but Docker can be kind of tricky. So this command right here, it's going to launch the web ui dot pi script,

[06:44] copy and paste that. Hit enter and we should be off to the races. Yeah, yeah, so now we're going to navigate out to our browser and go to local host port Now let's go full browser mode here. Actually, no,

[06:58] command line. Now, fair warning, They're super fun and you can go crazy with this, For now, we're going to do something just quick and easy.

[07:12] Here we have our LLM provider. I'm going to choose, let's see, nothing in the cloud, and then we'll choose our model name.

[07:24] Now we'll say if you're doing Quinn or Llama two, they're dumb. They have a really hard time doing this. I normally want to do it with deep eq, you do want to make sure you download the model just like so.

[07:39] If you want to get Quinn with alama, you'll open up your browser, I'm sorry, and you'll type in alama pull and that model name just like this. and here they have a little demo option, just a quick little thing to try out.

[07:55] Let's run it. Click on run agent and watch what happens. Oh, You can see on the right side, our terminal is thinking and things were failing. It's failing because Quinn is dumb. Yeah, it just couldn't do it.

[08:09] Let's try another LLM. Let's try deep seek R one 14 B. Pretty smart guy. so what's doing stuff? Check that out. what to look for. This is amazing. And notice how on the site,

[08:25] it's like Autocorrecting, it'll fail. Try again. It'll max out at five times. I'm going to try it with the local LLM just to see what else I can do with this. Let's do something simple like go out to network Chuck coffee,

[08:44] find the 4 0 4 error coffee and add it to it's called 4 0 4. Not found. so it made it to my site very quick. It's finding the search.

[08:59] It's like watching one of my kids try to use the computer. Oh wait, it's not called 4 0 4. Not found. What am I thinking? It can't find it. It's having such a hard time. Let stop him. Stop. It's okay buddy.

[09:11] I still can't believe we're able to do this and locally too, Now we'll add it to my cart. Why did it go to 200? Okay,

[09:24] It's a single origin. I forget where it's from now. I'm curious. Because that is something that Chad GT operator will not do. Okay, here we go.

[09:37] We'll be using Anthropic and Claude three five and we'll be using our own browser. This is cool because it'll keep my logged in sessions and here are the Essentially I wanted to log into hosting and create a VPS for me.

[09:53] And then here are the instructions for the operator. Same thing, One is going to use Ubuntu 24 0 4 as the OS one is going to use one of the It'll just be installing Docker and I'll launch them roughly at the same time.

[10:08] Ready? Set, go. We're off to the races again. we're already at hosting here. It's going to get logged in. Come on buddy. It's my browser. I know. I'll have to log it over here. I take control.

[10:26] It's going to set up A-K-V-M-V-P-S. Oh, it's going now. As you can see, It's searching for Ubuntu. It found it. It's so smart. I love it.

[10:39] Stupid thing. Okay, we'll set the root password. I have no idea if it'll actually do this, right? It did it. Oh my gosh, I want KVM two eight terabytes of bandwidth. That's a lot.

[10:54] Two virtual CPU cores. Okay, here's the thing, can it use the coupon code? did it add the coupon code network? Chuck 10? Is it doing it? Whoa, whoa. It's adding 20 servers. Stop. No, stop, cancel.

[11:10] Oh my gosh. I better stop that. I'm going to be very specific about the number of servers I want. but goodness, that was scary. I'm stuck on caps lock over here.

[11:27] I can't even do anything. I started over, gosh, it's stuck on caps lock. caps lock is currently not on for me. Try it freaking again. Eight gigs. A ram is really good for 6, 9 9 a month. That's crazy.

[11:44] That's a good server. Alright, here we are again. Don't do 20 servers please. It's doing one month but it put fricking 11 over there. Oh no. and I don't think it used my coupon code chat. BT is still screw me over.

[12:01] Just bought 11 servers. At least it wasn't 20 for a year. I have to restart chat GPT again. Alright, they're setting up my VPS. so the open source browser thought he was done, I think. Yeah,

[12:16] So he did make a server but he didn't stink and use my coupon code I don't think. Let's see if he actually made that many servers. Yeah, he did. they're amazing because they are a MD Epic CPUs.

[12:32] Chat g. PT is still trying to figure it out. Scroll down, dude. Chat. What it's hanging up on right now is the application options. Yeah, you can install regular Linux oss or you can do applications that are

[12:49] We'll choose Docker because that's what I told him to do. it's going but seriously, if you want to have a project in the cloud, Powerful servers coupon code.

[13:04] It's doing the coupon code and with never chucked 10, you'll get 10% off. Lies. It does exist. Maybe it's only a year to 12 months. There we go. I feel like we're missing something. I'm going to add one more thing.

[13:19] What's happening over here in Chachi PT Land? What's it doing? Oh, Chay PTs in the terminal instead of asking me, okay, it said it created it.

[13:31] Let's see. No it did not. It's not there. Chay Piz is a liar. No, no, stop. It's doing it again. No, no, stop. Okay, the verdict browser use works great if you want more servers than you want.

[13:45] but he had to have his hand held the entire time and he lied at the end. Anyways, thanks to hosting it for sponsoring the segment. If you want a VPS, you should get one right now. Use the code network. Chuck 10 for 10% off a year.

[13:57] you're stressing me out. I got to cut you off. So that's using local ai. Now we know that using any kind of cloud-based ai, like open AI or anthro, I want to test the speed.

[14:12] So we'll change from llama to anthro and we'll choose the Claude One of my favorites actually run this same task. Keeping in mind, many more cool things through programming. What have you do? That was fast.

[14:30] It's going, I found the coffee. It's on the right coffee page now. Okay, so smart and added it to cart. That's so cool. You did such a good job. Good job buddy.

[14:42] I do want to test Quinn one more time just to see if it was a fluke because I using. Let's see, let's do something simple. Go to YouTube and find a video from Network.

[14:58] Chuck. Let's see how it does. Dang, that was fast. Okay, go Quinn. That is so cool. Oh my gosh. And it started playing the video. That's so awesome. Now just so you know, I haven't played with this extensively,

[15:12] So the one big limitation with Chad GBT operator is that it's using this random So lemme show you what that means. If I were to ask it, like right now,

[15:25] I can take control of our eBay session and log into my own eBay But here I'm using the browser that can say finish up. You can have control. we can actually use our own browser with our own settings.

[15:40] Our password manager and our AI can handle it for us. That's so powerful, dude. It's still watching videos over here. That's so amazing. So I'm going to try and tell it to find a specific video and leave a comment.

[15:56] but I wanted to get to that point. And this is using Quinn. Yeah, Go on YouTube and find the video from Network Chuck covering Leave a comment saying what should we say here in

[16:13] 2025? Sorry, I couldn't think of anything better. Let's just try this. Go Quinn, you've got this. I could honestly do this for hours. Just imagine the automation things you can do. Are you kidding me?

[16:30] I'll notice you probably saw this deep research thing. I've not tried that, but you can also go to recordings and it will actually show you the result of what it was doing. So if you're like, what did this guy even do on my browser?

[16:44] Let's see if Deepsea can do it. Deepsea is moving. So I found some videos. Is it going to recognize that those are not the videos? Well scroll down.

[16:56] I'm like wanting to scroll stressing me out. Come on, you can do it buddy. Not the docker video or the network video, but it's on a video. Yikes. 666 comments. Please leave one please. Okay,

[17:08] I do want to see if Claude can do this very fast and then I want to jump into a test to head to head test between open AI operator and browser use. I want to do that same Japanese eBay situation. Alright, let's go. Claude,

[17:23] jump on it. You're there now leave a comment. But I love seeing the thinking on the right side here in the terminal.

[17:39] Or maybe try to leave the comment maybe. I didn't see that. Chuck, I need a Linux environment with a gui.

[17:52] I believe there is a headless version where you can just run in headless mode leave a note in the comments, encourage somebody that can work. Head to head versus chat GBT operator. To make it fair,

[18:09] I will use philanthropic Claude on my web UI here in my browser use and we'll I'm curious if they'll find the same exact VCR. Okay, ready? Set. Who should I start first? I'll start him first. Go,

[18:24] go and they're off. This is fun. My agent got to eBay first. Operators typing in first. Oh wait, hold on. GPT first browser use is still trying to figure out where the search button is.

[18:41] Yeah, and eBay. I'm sorry. Operator already found it. Yeah, add it to my cart. That was pretty quick. Come on browser. I'm going to try it with deep seek. Maybe that does better.

[18:56] I'll tell you one thing, even though mine might be a bit slower, Sam Altman's tracking what you're doing. Oh wow. It's probably going to go for the same VCR. There it is. Come on.

[19:11] Add it to cart. Add it to cart. Yes it did it. Will it proceed to cart? Yes it did. It finished. It did it. Oh my gosh, I know people comment that I do that. No, this is seriously amazing.

[19:25] Now I know for a fact the operator will not do this, Lemme show you what it looks like here. It's simply where you can test a capcha. Solve this capcha.

[19:46] Let's see if my local one can do this. I'm still on deep seek. This is all local I think, right? Yes. We're on alama deep seek R one 14 B,

[19:58] I really want you to win. Okay, we're here. It's got the caption up. Oh my gosh. Will this trip it up? What's the command line saying? I'm so curious.

[20:10] It's also probably having trouble recognizing stuff instead of click button with index zero. What's zero? It keeps clicking the cap button. Okay. Let's stop that. I want you to solve a capcha.

[20:29] click the I'm not a robot check box and then a capcha verification will pop up. Do what the instructions say. Okay,

[20:47] let's see if it does this. Let's get the terminal up here. We're on the site. It may have solved it because what happens when you finish it? Because it's just a demo, right? Let's do it side by side.

[21:05] Oh, so it should say I'm not a robot. Come on dude, I have so much faith in you. Oh, it's clicking squares now. It did something. It's learning. I don't know if it selected the right square that time though. Okay,

[21:17] we've got to end this. So this is an open source version of the chat. Anything we can run open source local is amazing. I wish I had the time to go crazy with this and program and do all kinds.

[21:33] Let me know if you want to see a video of just some sort of programming automation thing. Give me some ideas. I would love to hear that. Comment below. Also think about this, the hacking ramifications.

[21:47] and really there's no limit to what we can do. Think about hackers, he's never going to figure it out. That's all I got.

[21:59] I'll get you guys next time.

More from NetworkChuck

View all

⚡ Saved you 0h 22m reading this? Transcribe any YouTube video for free — no signup needed.