Aside AI Browser Launches with SOTA Browser Agent and Local Workflows

Foreword

Today I came across a newly released AI browser: Aside.

It does not just cram a chat panel into the browser sidebar — it makes the browser agent the core selling point. The official page puts it plainly: Aside is "The SOTA browser agent," capable of working across logged-in websites, tabs, messages, documents, and local files.

My first impression after a quick spin: the UI is well-crafted. This does not feel like an engineering-demo-grade agent browser — it looks closer to a browser product you could use day-to-day. One thing that stood out: after logging in with ChatGPT, usage seemed to pull from Codex credits rather than regular ChatGPT conversation credits. That is something I will need to watch more closely, but if browser agent usage is indeed tied to Codex credits, Aside's positioning would be quite different from typical AI browsers.

• Aside Official Site
• Aside Browser Agent Page
• Aside Benchmark Results Repo

Official Pitch: SOTA Browser Agent

Aside's most visible claim is its browser agent benchmark results.

The official page states that Aside ranks first on three browser agent benchmarks:

• Online-Mind2Web
• BU-Bench-V1
• Odysseys

The page also lists comparisons including OpenAI, Anthropic, and Browser Use. On Online-Mind2Web, Aside's page shows:

The public benchmark repo also shows Aside's record on Online-Mind2Web: 297 out of 300 tasks passed, for a reported pass rate of 99.0%. One task is marked as impossible; excluding that one gives 297/299, or 99.3%.

For now, I am treating this as "Aside's official and public repo benchmark results," not as proof that it will outrun every browser agent in everyday use. Browser agent benchmarks are heavily affected by website state, login flows, anti-bot measures, and whether tasks are still completable — all of which affect the numbers. Still, 99.0% is an eye-catching figure.

How It Differs from Typical AI Browsers

There are plenty of AI browsers out there, and the problem is usually not "can it answer questions" but "can it actually get things done."

Aside's direction is closer to rebuilding the browser as an agent workbench. The official page says it can handle:

• Logged-in websites
• Email, dashboards, internal tools
• Comments, replies, follow-ups
• Docs, spreadsheets
• Local files

These are exactly the places where browser agents tend to get stuck. Many agent demos look impressive, but when they hit login screens, tab switching, forms, files, payment flows, or sending messages, they stop and hand off to the user. Aside's pitch is to bring those real workflows into scope.

The official page also mentions that sensitive actions — payments, posting, sending messages — wait for user confirmation. That matters, because once a browser agent can operate on logged-in sites, its权限 is much broader than a regular chatbot. Without a clear confirmation mechanism, the risk would be high.

UI Feel: More Polished Than Expected

What stood out to me most was the UI.

Aside's interface is not a rough automation tool. It clearly integrates the browser, agent tasks, context, and operational flow into a cohesive design. The visuals are polished and the status presentation feels like a complete product rather than a thin wrapper.

That matters for an AI browser. The core of a browser agent is not just running benchmarks once — it is whether you are willing to keep it in your daily workflow. If the UI feels clunky, cluttered, or confusing, users will go back to Chrome or Arc. At least on first impression, Aside looks more like something you could use long-term than most agent browsers.

ChatGPT Login and Codex Credits

Another thing I tested was how it behaves after logging in with ChatGPT.

Aside's official page mentions you can bring your own subscription — ChatGPT, Claude subscription, or your own API key. My impression is that after logging into ChatGPT, it seems to use Codex credits rather than general ChatGPT conversation credits.

I will keep this as "what I observed so far" because credit usage ultimately depends on the official account page and usage history. But if this direction holds, Aside's positioning leans more toward an agent work tool than a chat browser.

In other words, it pulls together Codex, browser-use, and AI browser into one thread: not asking AI to read a webpage for you, but having AI directly operate the browser to complete tasks.

My Assessment So Far

Aside is worth paying attention to, but I would separate it into two layers.

The first layer is product experience. The UI is polished, the agent interface has a solid level of completion, and the official page explicitly accounts for real-world issues like logged-in websites, local files, and sensitive action confirmation. That is more practical than building yet another sidebar chatbot.

The second layer is the benchmark claims. Aside presents strong browser agent scores on its official page and public repo — especially the 99.0% on Online-Mind2Web, which is impressive. But browser agent benchmarks are naturally affected by website state and evaluation setup, so I see these as a strong signal rather than a guarantee that every daily task will be handled reliably.

What I want to test next are a few real-world use cases:

• Help me organize data within logged-in websites
• Cross-tab research across multiple pages
• Read local files and then perform web actions
• Handle forms and follow-ups
• Whether sensitive actions actually pause for confirmation

If these hold up, Aside would not just be a good-looking AI browser — it could genuinely become a new benchmark for the browser agent space.

Sources

• Aside Official Site
• Aside Browser Agent Page
• Aside Benchmark Results Repo