[00:03] lighter. What's quietly eating your memory every time your AI agent opens a browser? Most people never check. OpenCode just did, and they didn't like what they found. So they tore it down and rebuilt [00:15] it from scratch. Today I'm breaking down OpenCode Browse 0.1.0. What it is, what changed, and why it matters. And later, I'll show you what happened when I pointed it at a real website and handed it a big research [00:29] job. So what is OpenCode Browse? It lets your AI coding agent use your real browser, your actual browser with your tabs, your logins, and your extensions. It has two parts, an MCP server and a browser [00:45] extension. MCP stands for model context protocol. It's a standard way for AI agents to find and use tools. Together, these two parts let your agent open tabs, click around, read pages, and take [00:58] screenshots. It works with Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera, and Chromium. And it's not just for OpenCode. It works with any MCP client, including Claude [01:11] Code, Cursor, Codex, and Zed. Now here's the big news. OpenCode just released Browse 0.1.0, and they rewrote the whole thing in and they rewrote the whole thing in Rust. Three headline changes, 25 times [01:25] less memory, 23 times smaller, and no more playwright with browser actions exposed directly to your agent. People mix up those first two numbers, so let's be clear. 25x is memory. That's how much RAM it uses while it runs. 23x is size. [01:42] That's how big the program is on your computer. The new stack put real numbers on it. The new Rust program is about 4.3 megabytes. The old Node version was around 53 to 64 megabytes. And memory [01:55] for each Browse connection dropped from roughly 190 megabytes to 5.6 megabytes after a few calls. That's a drop of around 97%. Your agent might keep Browse running in the background all day. At almost 200 [02:11] megabytes, you feel it. At under six, you barely notice it. So you can just So how did they do it? The old version ran on Node and leaned on Playwright. Playwright is a popular tool for controlling browsers with code. [02:26] It's powerful, but it's big and it brings a lot your agent never uses. So Dax Rod, a developer on Open Code, rebuilt it in Rust and replaced only the parts they actually needed. Rust is a programming language known for being [02:40] programming language known for being fast and light on memory. The result is Okay, quick pause before I show you the next part. If you want to learn tools like this step by step, check out the AI Profit Boardroom. We've got a 30-day [02:54] roadmap that walks you through getting set up with AI tools like this one, one step at a time. There are four coaching calls every week. So if you get stuck setting up something like Browse, you can bring it to a call and work through [03:06] it. You also get the videos, the prompts, and the files to copy so you're never starting from a blank page. Link in the description or go to ai profit boardroom.com. Right, the part I find most interesting [03:19] is that browser actions are now exposed directly to your agent as tools. Dax explained what that unlocks. All the browser tools are now available in code mode. So your agent can mix browser actions with your other MCP [03:33] tools in one go. His example was this. Open a page, grab a list of things, loop over them, and call another MCP tool for each one. One flow instead of a long chain of back and forth steps. So how is this different from normal [03:48] browser automation? A typical Playwright setup often launches a separate browser with a fresh profile. No logins. So when it hits a page that needs you signed in, it's stuck. Browse uses the browser you already have. If you're signed into a [04:03] extra setup. It's also built to stay out of your way. Agent tabs open in the background inside collapsed tab groups so you keep working. When the agent hits a login or a two-factor code, it hands the tab back [04:17] to you. You do that step, click continue, and the agent picks up where it left off. And if you already have a tab open, you can click a button in your toolbar and hand it to your agent. It can also take screenshots, record the [04:30] screen, and capture network traffic as a HAR file. That's a log of everything a page loaded, which is really handy when something breaks. And you can set an allow list so a session can only touch the sites you choose. Now let me show [04:43] you what happened when I tested it. I gave my agent a prompt to act as a community growth assistant for the AI Profit boardroom. Browse our public pages and build a growth plan without inventing anything it couldn't verify. [04:56] growth blueprint with our value proposition, the six audience groups from our 42 niche playbook page, a seven-day onboarding path, five starter tutorials, and seven daily community themes. It even flagged older blog pages [05:11] that no longer matched the homepage and noted that it never opened the private member area. So what else can you use this for? For coding, your agent can build a page, open it in your real browser, and take a [05:23] screenshot to check it. If something fails, it can check the network log to see why. For testing, it can click through a flow on a site you're already logged into, like your own dashboard. So how does it stack up against other AI [05:36] browser tools? Some of them run in their own separate browser. That's safer for some jobs, but you lose your logins. Browse works inside your own browser, and because it's an MCP server, it isn't locked to one app. Now some honest [05:50] limitations. First, this is version 0.1.0. It's early, so expect changes. Second, it uses your real browser, which means your real accounts. A web page can hide text that tries to trick your [06:05] agent. That's called prompt injection. So use the allow list and think about a separate browser profile with only the accounts that task needs. Third, some websites are good at spotting automated browsing, so some sites may still block [06:19] it. So how do you get started? Go to opencodellc.ai and find the browse page. The easiest route is to install the browser extension first. It walks you through the rest. If you like the command line, [06:31] the page also has one line installs for Mac and Linux, Windows PowerShell, npm, and Bun. And the code is open source on GitHub under anomalico/browse. Once it's set up, start small. Give your agent one simple job on a page you're [06:47] already logged into. Then build from there. Here's the takeaway. Browser tools for AI agents used to be heavy. Now this one is a few megabytes and works inside your real browser. Here's the thing. There's a window right now [07:01] where most business owners haven't figured this out yet, and that window won't stay open forever. Learning these tools early gives you a real head start. If you want to get ahead, join the AI Profit Boardroom. You get daily [07:14] tutorials so you keep learning new AI tools as they come out. You get four coaching calls every week where you can bring your questions and get unstuck. There's a prompt library so you're not writing every prompt from scratch, and [07:26] find other members in the community. Link in the description or go to aiprofitboardroom.com. If you want the full process, the SOPs, and over 100 AI use cases like this one, join the AI Success Lab. It's our free [07:41] community. Links in the comments and description. You'll get all the video notes from this episode plus access to 95,300 members. The notes cover everything from today, from what changed in browse 0.1.0 [07:55] to how to set it up, so you don't need to pause and write anything down. Just come in, grab the notes, and learn alongside everyone else. If this video helped, go join the AI Success Lab now. I'll see you inside.