This is just the beginning.
More local models, more ways to run AI on your own machine.
Starting today with WebAI, bringing models like Gemma and Apple's Foundation Models into the browser 👇
What if you could use AI without sending your data to the cloud?
Introducing WebAI, a set of browser APIs that lets developers integrate locally running AI models into websites and extensions they build.
It’s opening up new possibilities - now in early preview in Opera One and Opera Neon, with Opera GX to follow.
🧵 Here's how it works.
So, what exactly is WebAI?
It lets websites and browser extensions use AI models running directly on your computer.
No separate AI app. No external AI server. No API keys.
Just AI capabilities built into the browser, ready for developers to use.
Here's where it gets interesting.
WebAI gives developers access to three core APIs:
• Prompt API: Generate, summarize and work with text.
• Decision API: Classify information and make decisions.
• Semantic Embeddings API: Find connections between pieces of content.
All powered by compatible local AI models.
AI can build you an app in an hour and a website in ten minutes. That part is real, and honestly, it's impressive.
The problem is what most people do next. They look at what came back, see that it works, and call it done. Then they put it out there and can't figure out why nobody cares.
Usually it's because you can tell nobody cared while making it either.
AI is great at getting something to work. But getting something to work and making something people actually want to use are two different things.
Take a website. You ask AI to build a landing page for a fitness app, and it gives you something that looks perfectly fine.
Big headline: "Transform Your Fitness Journey."
A few generic promises about reaching your goals. Some stock-looking visuals. A big "Get Started" button.
Nothing is technically wrong with it. It's just a website that could belong to a thousand different apps.
Now imagine changing that headline to:
"Made it to the gym twice this week? That's two more times than last week."
Suddenly, you're speaking to a specific person with a specific problem. Someone who struggles with consistency, not someone looking for another app promising to transform their life.
That's the difference. AI built the page, but someone had to decide who it was for and what would actually resonate with them.
And it's not just websites. You see it in AI-written posts, presentations, videos, and almost anything else people are generating.
None of this means you should stop using AI. It means the first output probably shouldn't be the final one.
When anyone can make something in minutes, the interesting part becomes what you choose to change.
So what gives it away for you? What's the first thing that makes you think "yeah, nobody touched this after AI made it"?
New models are easy to notice. The harder part is figuring out what should actually change in your work.
A model can be faster, cheaper, and score better on benchmarks. That still doesn't tell you which tasks can be run on their own, or when you need to stay involved.
If you already run your own agents, chances are the thing slowing them down isn't the model. It's what happens when a task needs to touch a real, logged-in webpage. Here's what changes when you connect an agent to a real browser session instead.
That's what Opera Neon's browser connectors are built for: agents you already run, given access to a real browser session instead of a workaround. Worth testing on whatever task currently needs you to manually bridge the gap.
Start using Opera Neon's browser connectors today: opr.as/2bpb
None of this requires switching your whole stack. Your existing agents plug into it the same way they'd plug into any other MCP server.
There's also a token-efficiency side worth knowing: reading through a real browser session this way runs meaningfully leaner than pulling the same context through a raw API call.
The connection itself is lightweight. An agent that speaks MCP can read tabs, page content, and screenshots from the browser session directly, without you rebuilding your existing setup around it.
In practice this means tasks that used to require a person to manually bridge the gap, copying something from a logged-in dashboard into your agent's context, can now happen directly.
Connecting to a real browser session gives your agent the same access you have. Whatever you're signed into stays signed in. The agent sees what you'd see, not a stripped-down version of it.
Most agent setups work with an API, a scrape, or a headless browser with no session behind it. Fine for public pages. It breaks the moment a task needs something behind a login.
Automation can handle a lot, but a login can stop the whole workflow.
A script can fill out a public form. Without an authenticated session, it can't see your inbox, dashboard, or account data.
What's the last task an "automated" workflow handed back to you?
A lot of products call themselves "AI powered" when the AI is really just sitting on top of a process designed years ago.
A search box becomes a chat. A form starts suggesting answers. Useful? Sure. But once that part is done, you're back to the same approvals, handoffs and copy-pasting between tools.
One step got faster. The workflow didn't.
"This AI can read the web" sounds straightforward. But what does it actually have access to?
Can it fetch a public page? Access an open browser tab? Work within an authenticated session? Those are very different capabilities.
Authentication uses a secure MCP server URL, so only authorized clients can access the session. You can also control which read and write tools are available.
Worth checking which kind of "reads the web" claim you're relying on before you trust the answer.
Download Opera Neon today: opr.as/2bpb
Opera Neon’s MCP Connector lets external AI clients connect to your active browser session and use live context from open tabs, page content, and authenticated sessions.
With the relevant tools enabled, they can navigate, extract information, fill forms, and more.
So "can browse the web" and "can work inside your browser" are not the same claim.
Before you trust an answer or an action, check what source and context the AI actually had access to.
That matters when the useful context is tied to your session: an authenticated dashboard, a logged-in web app, or the tabs you’ve gathered for a research task.
Same URL. Different context.