ego lite is a Chromium-based browser built specifically for AI agents to work with websites. Its main difference from regular browser automation is that the agent gets a separate workspace, can use existing authenticated sessions, and reads the page through a compact Snapshot instead of sending huge HTML to the language model.
The practical result is simple: Codex, Claude Code, Cursor, Hermes Agent, and other agents can open websites, read data, click buttons, fill out forms, and download files while the user continues working in their own tabs. This significantly reduces the amount of data that has to be sent to the model.
What ego lite is in simple terms
A regular AI agent cannot use a browser on its own. It needs a software layer that opens the page, passes information about it to the model, performs the selected action, and returns the result. This architecture is covered in more detail in the article about Agent Harness and AI agent architecture.
ego lite combines the browser with such a control layer. The connecting component is ego-browser: through it, the agent gets functions for navigation, page reading, clicking, filling in fields, waiting for loading, downloading files, and running JavaScript.
User task → AI agent → ego-browser → Agent Space → page Snapshot → browser actions → result
The AI agent itself is not tightly built into the browser. You can use the tool you already work with: Codex, Claude Code, Cursor, OpenCode, Hermes Agent, DeepSeek Harness, or your own agent harness.
Why AI agents spend so many tokens on browsers
For a person, a web page is a visual interface. For a language model, it can become a huge array of HTML, styles, scripts, hidden elements, and the page's technical state.
According to ego lite documentation, the contents of a typical admin page, when transferred almost in full, can reach about 30,000 tokens. And after the page changes, the agent often has to read it again.
The problem is not just API cost. A large amount of technical clutter takes up the model's context window and makes it harder to find truly important elements: the “Submit” button, a search field, a link to the next page, or values in a table.
How Snapshot reduces token usage
Instead of full HTML, ego lite uses Snapshot — a compact structured representation of the page. It is built mainly on the browser's accessibility tree, that is, the semantic representation of the interface.
The Snapshot retains what the agent really needs:
- page title and URL;
- visible text and the main structure;
- buttons, links, input fields, and other interactive elements;
- the role and name of each element;
- a temporary reference such as @1, @2, @3, which the agent can use to perform an action.
In the official documentation, the developers give an estimate: instead of tens of thousands of tokens, a full page Snapshot can take a few hundred tokens.
This is especially important for multi-step tasks. If an agent needs to open a website, search, click a result, change a filter, and then read a table, it has to retrieve the page state several times. The more compact each state is, the less context is consumed.
A similar principle is used at other levels of agent architecture: it is more efficient for an agent to receive only the data needed for the current step. Read more in the article how an AI agent manages context through Skills, RAG, and routing.
What Spaces are and why the agent no longer takes over the browser
In ego lite, each agent task can run in a separate Space. This is an isolated workspace within the same browser process.
The user continues working in their own tabs, while the agent opens pages and performs actions in its own Space. The user's cursor, active tab, and current window should not switch with every agent action.
The Agent Space remains visible. You can open it and see which page the agent is on and what it is doing.
Spaces also allow several independent tasks to run. For example, one agent can analyze competitors' websites, another can work with the CRM, and a third can gather information for a report.
This differs from launching a separate Chrome instance for each operation: Spaces use shared browser infrastructure while keeping task work contexts separate.
Takeover: when control needs to be handed over to a person
A fully autonomous browser is not always needed. Sometimes a website shows a CAPTCHA, additional authorization, login confirmation, or another action that must be performed by the user.
For this, ego lite supports a control handoff mechanism. The agent can pause a task and hand the Space over to a person. The user performs the necessary action directly in the browser, after which control can be returned to the agent.
This creates a convenient human-in-the-loop flow:
The agent handles routine tasks → manual confirmation appears → the user takes control → confirms the action → the agent continues working.
This is useful for more than just CAPTCHA. The same approach works for final form submission, publishing content, placing an order, or any other step where it is more convenient to stop automation before the final action.
How ego lite works with existing logins
ego lite is based on Chromium and offers to transfer data from an existing browser on first launch. If the user confirms the migration, history, cookies, extensions, profile, and login state can be transferred.
This means the agent does not have to start with an empty browser every time and log in again to a CRM, social network, or internal dashboard.
It is important to clarify: migration is performed at the user's discretion. This is not an automatic import of all data without confirmation.
According to the developers, browser data is stored locally. When performing a task, the contents of the required page may, of course, be sent to the AI agent and language model selected by the user to process the task.
Which AI agents can be connected
The official ego lite documentation includes instructions for several popular agent environments:
- OpenAI Codex;
- Claude Code;
- Cursor;
- OpenCode;
- Hermes Agent;
- OpenClaw;
- DeepSeek Harness;
- local models via a compatible agent harness;
- custom AI agents capable of running ego-browser on the user's computer.
For example, for Codex the browser is connected as a Skill. If you have not worked with this mechanism before, see the separate guide Skills and Plugins in Codex.
The setup for Hermes is similar: the agent receives ego-browser as a tool and uses it for web tasks. Hermes itself and its use cases are described in the article Hermes Agent for business.
How to install ego lite
As of September 30, 2026, the full ego lite app is available for macOS. The developers list Windows as a closed beta and Linux as on the roadmap.
Step 1. Install the browser
Download ego lite from the official website and install the app on your Mac.
Step 2. Complete the initial setup
On first launch, the browser will offer to transfer data from an existing browser. If you need to work in services where you are already signed in, select the appropriate profile and confirm the migration.
During onboarding, ego lite also detects installed AI agents and can add the ego-browser Skill to them.
Step 3. Install the Skill manually if needed
You can add the Skill via npm:
npx skills add citrolabs/ego-lite
After that, the AI agent receives instructions for using ego-browser.
Step 4. Set your first task
In a compatible agent, you can call the Skill and describe the task in plain language. For example:
/ego-browser Open the websites of three competitors, collect the names and prices of new products, and save the result in Markdown. Do not change anything on the websites.
The agent then creates a Space, opens pages, gets a Snapshot, and performs the necessary actions on its own.
Why ego-browser uses JavaScript instead of a long chain of commands
Another feature of ego lite is that an agent can combine several actions into a single JavaScript scenario.
In the traditional approach, the workflow may look like this: the model called a tool → got a result → thought again → called a second tool → got a result → called a third.
Each such cycle requires information exchange between the model and the tool.
ego-browser lets you move part of the logic directly into the executed code. For example, an agent can open a page, wait for it to load, find elements, go through several entries, and return only the final result.
This reduces the number of model-tool round trips — cycles of “model → tool → model” — and further lowers token usage.
What can be automated with ego lite
The main use case for ego lite is websites without a convenient API or where the task requires working with a regular user interface.
| Task | What the agent does | Result |
|---|---|---|
| Competitor monitoring | Opens websites, checks prices, new products, and changes | Markdown, CSV, or report |
| CRM and internal dashboards | Finds records, collects data, applies filters | Structured extraction |
| Content monitoring | Reviews multiple sources and selects relevant publications | Summary or content database |
| Working with forms | Opens a page, fills in fields, and prepares the form | Completed draft for review |
| Data collection | Navigates pages, filters, and tables | CSV, JSON, Markdown, or local file |
| Internal SaaS | Works with the service through an existing login | Completed operation or report |
Essentially, the browser becomes another tool for an AI agent alongside the terminal, files, APIs, and knowledge base. To understand the overall difference between a chatbot and a system that actually takes action, see the overview how AI agents work for business.
How much faster is ego lite really
The developers publish their own ego lite tests on multi-step browser tasks. The current project README states that in four test scenarios, ego lite completed tasks up to 2.5 times faster than Vercel agent-browser and used fewer tokens.
This should be viewed as a developer benchmark, not a universal guarantee. The actual difference depends on the website, interface complexity, chosen model, number of actions, and the task itself.
A more important architectural advantage can be assessed without a benchmark: Snapshot reduces page size, while executing multiple actions with a single script reduces the number of model calls. For long browser workflows, these two factors directly affect context usage and execution time.
Is ego lite fully open source?
Not exactly. The ego-lite project is published on GitHub with open-source code for ego-browser, Skill, and related infrastructure under the MIT license. But the ego lite browser itself is distributed separately as a free app.
So it is more accurate to say that ego-browser and the agent part of the project are open source, rather than calling the entire browser open source.
By the end of September 2026, the project repository had gained around 16,600 stars on GitHub.
What limitations does ego lite have
The main limitation right now is the platform. The full version works on macOS. If your work machine runs only Windows, it makes sense to wait for a public Windows release or use another browser tool.
Snapshot also does not replace computer vision in every situation. The documentation explicitly states that images, complex canvas interfaces, content outside the accessible area, and some cross-origin iframes may require a screenshot, additional data extraction, or manual action.
In addition, ego lite saves tokens for browser automation, but does not make the language model itself free. If a paid API from OpenAI, Anthropic, or another provider is used, its cost remains.
For simply reading a public page, ego lite is often overkill too. If you don't need login, clicks, forms, or a full browser, regular web search or an HTTP request will be simpler.
Which scenario is best to start with
Don't start exploring a browser agent with a large process across dozens of websites. A good MVP is one recurring task whose result is easy to verify.
- Choose one website or a small set of websites.
- Describe what data the agent should retrieve.
- Specify separately what it is allowed to modify and what it should only read.
- Set the result format: table, Markdown, CSV, or a brief report.
- Run the task several times and compare the results.
- Only after that should you add a schedule, new websites, or integration with CRM and other systems.
For example, the first scenario could be a daily check of five competitor websites: the agent opens them, looks for changes in pricing and products, and saves the result in a structured Markdown file.
After validating this MVP, the browser can be embedded into a larger workflow: data source → AI analysis → browser action → CRM or spreadsheet → manager notification.
Who ego lite is really for
ego lite is especially useful for developers and companies that already use AI agents and want to give them access to real web interfaces without constantly launching separate browsers and passing huge pages into the model's context.
The tool is worth considering if an AI agent regularly works with CRM, SaaS dashboards, social media, internal admin panels, forms, catalogs, or websites without a full API.
For such scenarios, ego lite solves three practical tasks at once: it reduces the amount of data sent to the model, separates the agent's work from the user's tabs, and lets it use an existing browser session.
If you need not just to test the browser but to integrate an AI agent into a company's workflow — for example, connect a web interface with CRM, Telegram, Google Sheets, or a reporting system — you can discuss an automation MVP with CenterAI.
FAQ
Does ego lite work on Windows?
As of September 30, 2026, the full application is available for macOS. The official README lists Windows as a closed beta and Linux as being on the roadmap.
Is ego lite free?
Yes, the browser itself is distributed for free. But the AI agent and the model it uses are paid for separately if the selected provider or service is paid.
Can ego lite be used with a local model?
Yes. The official documentation includes an example of working through OpenCode and a local model in Ollama. Another agent harness will also work if it can launch ego-browser on the computer where the browser is installed.
Can ego lite be connected to Codex or Claude Code?
Yes. Separate instructions have been published for Codex, Claude Code, Cursor, Hermes Agent, and several other agent environments. The connection is made through the ego-browser Skill.
Does ego lite replace Playwright, APIs, and regular web search?
No. For a public page without authorization, regular search or an HTTP request is often simpler. APIs are best used when a service provides a stable programmatic interface. ego lite is especially useful where an agent needs to work with the real web interface.
Sources and verification
Last verified: 30.09.2026