A Browser for AI Agents: ego (lite)

Trying to delegate browser tasks to AI agents often comes with more friction than you might expect.

Browsers used for automation often start in a clean state without login information, forcing you to re-login, perform 2FA, or go through SSO on every site.

Conversely, connecting an agent to your active browser often leads to conflicts, as the agent keeps opening new windows and tabs or steals your mouse focus while you're working.

“Is there a way to share my existing login environment while letting an AI agent work without interrupting my own browsing?”

ego (lite) is a Chromium browser for AI agents born from this exact question.

It can import your Chrome bookmarks, extensions, cookies, and login sessions so you can use it like your everyday browser, while being designed to allow external AI agents like Codex, Claude Code, and Cursor to perform web automation in a separate workspace.

You can keep working in your own tabs, while the agent reads pages, clicks buttons, fills out forms, and downloads files in an isolated Space.

What makes it different?

ego (lite) takes a different approach from browsers with built-in AI or automation frameworks that run separate headless browsers.

The key is for humans and AI agents to share a single browser environment while keeping their respective workspaces separated.

One browser for both humans and agents

Existing browser automation tools typically launch a new Chromium instance.

Because these browsers lack your login information and extensions, it's difficult to start tasks on services that require authentication, like Gmail, Notion, LinkedIn, or internal admin pages.

During onboarding, ego (lite) can copy your open Chrome tabs, tab groups, bookmarks, saved passwords, extensions, cookies, login sessions, and profiles.

This allows agents to work in the actual web environment you are already logged into, rather than starting from an empty browser every time.

The import process is optional and does not move or delete your existing Chrome data. Once imported, Chrome and ego (lite) are managed separately, so new logins or extensions in one will not automatically sync to the other.

Space: A dedicated workspace for agents

Just because users and agents share the same browser doesn't mean they compete for the same tabs.

ego (lite) provides a separate workspace called Space for each AI agent task.

A Space is not a new browser window, separate profile, or cloud browser. It works by creating isolated browser contexts for each task within the same ego (lite) process, segregating the tabs and storage for each job.

You can keep your tabs and mouse focus exactly where they are while the agent navigates pages or fills out forms in its Space.

You can even run multiple agents or tasks simultaneously in their respective Spaces. It's also possible to switch to an active Space to check progress or take direct control when needed.

Tabs remain even after a task finishes, allowing you to review which pages the agent visited and what it performed.

Connect the AI agents you already use

ego (lite) does not restrict you to using specific built-in AI models.

ego-browser through a skill called

During installation, it locates the agent CLI on your Mac to link the ego-browser skill. Afterward, simply describe your desired task to the agent in natural language along with /ego-browser.

For example, you can ask it to organize data on a logged-in admin page, compare prices of various products, collect social media posts, or run repeated web tests.

A key feature is that it doesn't lock you into a specific model or agent platform, but rather adds browser execution capabilities to the coding agents you already use.

A structure that executes multiple actions at once

Many browser automation tools repeat the process of checking a page, clicking one button, and sending the result back to the model.

Since each step involves a round-trip between the model and the browser, execution time and token usage increase as the task gets longer.

ego (lite) provides browser features such as snapshot, fill, click, wait, navigate, and capture as JavaScript functions.

Agents can write multiple actions as a single, short JavaScript flow and execute them sequentially on the page. Instead of calling simple commands one by one, this method allows the entire task, including conditional statements and loops, to be constructed as code.

In the production team's own benchmarks comparing four complex tasks against Vercel's agent-browser, it is reported to be up to 3.45 times faster and consume fewer tokens, depending on the task.

Agent-friendly Snapshot

The full HTML of a webpage contains information not needed for actual tasks, such as styles, scripts, and hidden elements.

Sending this to the model at every step makes the input excessively large and makes it harder for the agent to find the elements it needs to click.

ego (lite)'s Snapshot converts the page into compressed text based on its accessibility tree, including titles, body structure, buttons, links, and input fields.

Each interactive element is assigned a temporary reference number, allowing the agent to select elements without relying on screen coordinates or frequently changing CSS classes.

The Snapshot feature is implemented within the custom Chromium engine, and the team explains that it was designed to access cross-origin iframes, Shadow DOM, React Portals, and external SDK widgets that are typically difficult for general JavaScript-based extraction methods to capture.

Locally-focused data management

Sharing a logged-in browser with an agent makes privacy and permission management a crucial concern.

ego (lite) stores browsing data such as imported passwords, cookies, history, bookmarks, extensions, and profiles on the user's local computer. There is no need to create a separate ego (lite) account or register an email.

However, when performing actual tasks, the AI agent chosen by the user will read and process the page content necessary for that task.

Therefore, it is important to connect trusted agents and set limits in advance for irreversible actions such as payments, publishing, deletion, or sending messages.

The documentation also recommends setting work boundaries so that the agent pauses and waits for user confirmation for steps involving authentication codes, CAPTCHAs, payment verification, or changing account permissions.

Future possibilities

Browser automation is moving beyond simple web scraping to encompass tasks behind logins, such as email, CRM, applicant tracking systems, accounting tools, and internal operational screens.

However, the difficulty of passing along actual login states, combined with the separation between the automation browser and the user's primary browser, remains a significant hurdle.

ego (lite) can be viewed as a project that attempts to redesign the browser not as a tool that agents briefly borrow, but as “a workspace for both humans and AI agents”.

Currently, ego (lite) is a free beta version available for Intel and Apple Silicon Macs. Windows and Linux support is not yet provided but is included in the public roadmap.

Development is also underway for Skills to accumulate successful web tasks into reusable workflows, Browser ACP for connecting diverse agent harnesses, and enhanced browser control and side tab panels. Electron and native application manipulation are planned for later stages.

If you want to delegate repetitive browser-based tasks to an AI agent or have experienced the limitations of existing headless browsers while trying to automate logged-in SaaS and internal tools, this is a project worth checking out.

Key Features

🔗 Related Links

Website: https://lite.ego.app/

Download: https://lite.ego.app/download

Korean Documentation: https://lite.ego.app/document/ko/docs/product-introduce

GitHub: https://github.com/citrolabs/ego-lite

Roadmap: https://lite.ego.app/roadmap

📣 Are you the project creator?

Launch Curator showcases new indie projects and builders.

If you would like to introduce your own project, please join the channel and feel free to post about it.