I Spent a Day Inside an AI Browser — Here's What Actually Works (and What Doesn't)

Ezra Pham — Wed, 15 Jul 2026 06:01:04 +0000

Last week, I finally got my hands on Aside — the AI browser that launched June 24 (YC F25) and hit 1.5M views on the founder's announcement post in its first week.

I've been building agent systems for 2 years. I know how hard browser automation is. So when a 3-person team claims their browser outperforms Claude Fable on agentic benchmarks, I had to test it myself.

Here's what I found after a full day of real use — not benchmark scores, but actual workflow wins and friction points.

What It Actually Is

Aside is a Chromium desktop app that pairs a normal daily browser with a built-in AI agent. The agent can:

Log into any site using its built-in password manager
Navigate dashboards, email, spreadsheets, and internal tools
Take autonomous actions through "Ultrabrowse" mode
Give you a clean interactive snapshot of any page

The key insight: a browser agent reaches anything you can open a tab to. Integration-based agents only touch tools someone wired up. That's a fundamentally bigger surface area.

What Stood Out

The Snapshot View Is Genuinely Useful

Open any page and Aside returns a structured tree of every interactive element — buttons, links, inputs, textareas — each with a unique ref ID. No more digging through Chrome DevTools or guessing CSS selectors.

This alone makes it worth trying if you do any kind of browser automation.

Local-First Memory Is the Right Move

Browsing history stays on your machine as markdown. No cloud upload, no "your data trains our model" fine print. For developers who care about privacy, this is how you build trust with a tool that has access to everything you do online.

It Has MCP Support

This is the hidden superpower. Tools like Hermes Agent can drive Aside programmatically through the Model Context Protocol. That means your AI agent can browse the web, interact with pages, and execute multi-step workflows — all through a standardized interface.

No API keys. No OAuth setup. Just login once and go.

What Needs Work

Shared REPL Scope

In the REPL environment, every JavaScript call shares the same persistent scope. Use const x once, and you can't redeclare it in the next call.

// Call 1 — works fine
const pageTitle = await page.evaluate(() => document.title);

// Call 2 — SyntaxError: Identifier 'pageTitle' has already been declared
const pageTitle = await page.evaluate(() => document.body.innerText);

// Workaround: use `let` instead
let data = await page.evaluate(() => document.title);
data = await page.evaluate(() => document.body.innerText); // OK

Minor friction, but it compounds when you're iterating quickly.

UI Automation Is Inherently Fragile

This isn't specific to Aside — it's the fundamental tradeoff of operating at the UI layer vs. the API layer. If a site changes a button's data-testid, your script breaks. The snapshot approach helps, but it's not a silver bullet.

Learning Curve for Agent-Driven Mode

The REPL and MCP interface is powerful, but documentation assumes you already know browser automation. A library of common workflow templates would lower the barrier significantly.

The Verdict

Would I use it daily? Yes, for specific workflows.

Use Case	Recommendation
Browser automation scripts	⭐⭐⭐⭐⭐
Research + data extraction	⭐⭐⭐⭐
Building AI agent pipelines	⭐⭐⭐⭐
General daily browsing	⭐⭐⭐

The team is shipping fast — they just hosted an ICML 2026 afterparty with researchers from OpenAI, Anthropic, and DeepMind. For a first-gen product from a 3-person team, this is impressive.

If you build agents or automate anything in the browser, give it a spin.

I share daily build logs and tool reviews on X. Follow me @ezraphm if you're into agent systems, automation, and shipping things anyway.

DEV Community: Ezra Pham