Open Source
Open Source

ego-lite: A Browser Built for AI Agents (9.3k stars)

citrolabs/ego-lite is an open-source Chromium-based browser built for human + agent sharing (MIT, JavaScript, 9,335 GitHub stars, #9 on Trending weekly). Core selling point: agents reuse your real login state and tabs via the ego-browser skill and run tasks in parallel in their own Spaces, differentiating it from automation frameworks like browser-use that spin up a clean separate browser. macOS only today; Windows/Linux on roadmap.

Published August 9, 20268 min read
<!-- ego-lite-resource | open-source | ego-lite: A Browser Built for AI Agents (9.3k stars) -->

When an AI agent tries to do work in a browser, the awkward part usually isn't whether the model is smart enough--it's "whose browser is this." Have the agent hit a site that needs login and it either spins up a clean browser and logs in from scratch (cookies gone) or hijacks the tab you're actively using, mouse and keyboard flying around. citrolabs/ego-lite bets on a third path: a browser designed from day one for "human + agent sharing," where the agent reaches your real logged-in state and tabs through the ego-browser skill and runs tasks in parallel inside its own Spaces. 9,335 GitHub stars, #9 on this week's Trending weekly chart (+4,912 this week), MIT license, primary language JavaScript, last push 8/8.

Boundaries first: this piece is assembled from ego-lite's GitHub README and official docs, with facts as of 2026-08-09. Star counts move in real time--treat GitHub as the source of truth. This is a standalone deep dive on ego-lite, complementary to but independent of this site's prior browser-use resource (the browser-use automation framework) and AI browser automation tools compared. ego-lite's README explicitly positions itself against browser-use, but the two are different shapes (framework vs shared browser)--broken down below.

1. Why "a browser for agents" is needed

Anyone building agents has hit this wall. You want the agent to check a logged-in admin console, fill a form, or scrape a page only visible after login, and the two traditional automation routes each have their own pain.

The first route is "a framework drives a standalone browser." browser-use, agent-browser and the like are browser automation frameworks--they spin up their own browser instance to drive. Three problems stack up: that new browser is clean, so the cookies, sessions, and 2FA states you carry in your own browser don't transfer, and the agent has to log in again, which typically dies on OAuth/SMS codes/risk control; it's a separate process from the browser you're using, so you can't work in your own browser at the same time; and the framework drives via DOM screenshots plus coordinate clicks, which burns tokens, is brittle, and is slow. This site broke that path down in the browser-use resource.

The second route is "AI browsers." Products like ChatGPT Atlas embed AI into the browser itself, but they lean toward a "browses for you" consumer shape rather than a "programmable by an agent" engineering shape, so agents can't drive them for automation tasks smoothly.

ego-lite bets on a third path: not another automation framework, but a browser designed from the ground up for "human + agent sharing" (Chromium-based). You keep using it to browse, log in, and keep cookies; the agent plugs in through the ego-browser skill and reuses the real logged-in state and tabs of this browser, not a fresh clean instance. One line: a framework "gives the agent a new browser," ego-lite "lets the agent into the one you're already using."

2. ego-lite vs browser-use: framework vs shared browser

The README states the positioning plainly: existing tools like browser-use and agent-browser are browser automation frameworks that need a separate browser, can't carry login state over, and have humans and agents fighting over the same tab; ego-lite is designed for "human + agent sharing" from the start. Lay the key dimensions side by side. This table is based on ego-lite's GitHub README and official descriptions, as of 2026-08-09, subject to real-time change.

Dimensionego-lite (citrolabs)browser-use (framework)
ShapeShared browser (Chromium, macOS app)Browser automation framework (library)
Login stateReuses real human login state/cookiesClean browser, must log in again
TabsHuman and agent share one browserIsolated instance, separate from human
ParallelismAgents run in parallel in their own SpacesConcurrency controlled at framework layer
TokensOfficial claim: fewer tokens, fasterDOM/screenshot-driven, high token cost
Integrationego-browser skill / macOS appPython library, pip install
LanguageJavaScriptPython
PlatformmacOS only (Win/Linux on roadmap)Cross-platform
LicenseMITMIT

A few points to clarify. First, the shape difference is the root: browser-use is a library you call from code to drive a browser it spun up; ego-lite is a browser app itself that the agent plugs into to reuse its state. One is "framework drives browser," the other is "browser natively supports agents." Second, login-state sharing is ego-lite's core selling point and its biggest engineering difference from the framework route--reusing real cookies means the agent can walk straight into sites you've already logged into, skipping the OAuth/code hurdle. Third, token efficiency: ego-lite officially claims fewer tokens and faster completion, the logic being it uses closer-to-native browser calls rather than a "screenshot plus coordinate click" loop, but "fewer" is the README's wording--this site did not independently stress-test it, so don't treat it as an SLA. Fourth, platform is ego-lite's current hard limit: macOS only, with Windows/Linux on the roadmap, so cross-platform teams can't use it today. This site compared more tools side by side in AI browser automation tools compared.

3. How Spaces parallelism and login-state sharing work

This is the part of ego-lite most worth breaking out. Two mechanisms hold up "human + agent sharing."

Login-state sharing. You log into sites normally in the ego-lite browser, and the cookies/sessions stay in the browser. When the agent plugs in, what the ego-browser skill reaches is that real state, not an isolated clean profile. That means the agent can get to "only visible after you log in" pages: enterprise admin consoles, subscribed content sites, services behind 2FA. The framework route either can't do this or needs you to manually export cookies and feed them in--brittle and prone to tripping risk control.

Spaces parallelism. The agent doesn't take over the tab you're using; it runs tasks in its own Spaces--multiple agents can each run in their own Space in parallel without clashing and without grabbing your active browsing. The README's design intent: the human browses normally up front while agents work in parallel in background Spaces and hand results back when done. This solves the old pain of "as soon as the agent runs, it hogs my browser"--traditional automation, once started, hijacks the mouse and keyboard and floods the screen with popups, leaving you unable to do anything else.

One engineering note: because the agent reuses the real browser state rather than a fresh instance, it skips the "log in again and maintain a separate set of credentials" overhead, which is part of why the README says "fewer tokens, faster." But the trade-off is just as clear: the agent gets your real login state, so the permission boundary is yours to manage--covered under limitations below.

4. Getting started: install the app or npx skills add

Two paths.

Install the macOS app. Download the .dmg from the GitHub repo, in separate Apple Silicon and Intel builds. Once installed it's a browser you can use, log in with, and keep cookies in.

Install the skill (lightweight). npx skills add citrolabs/ego-lite installs only the ego-browser skill, the integration point for agents. Suited to cases where you already have a browser workflow you like and just want the agent to plug in and run tasks.

How the two connect: if you installed only the skill and not the app, the first time an agent runs a browser task it will guide you to install the app. Re-emphasize the platform limit: macOS only today, with Windows and Linux on the roadmap and no usable build as of 8/9. Docs live at lite.ego.app/document/, community on Discord. The README explicitly names Codex and Claude Code as supported agents--if you're already on one of those, integration is smoother.

5. Use cases and limitations

Suited for: you're on macOS daily and have a set of logged-in sites you want an agent to automate (admin patrol, data scraping, form filling, cross-site operations); you want multiple agents running browser tasks in parallel without them clashing or grabbing your tabs; you use an agent like Codex or Claude Code and want to give it a browser that can use your real logged-in state.

Four limitations to state clearly. First, macOS only. Windows/Linux teams can't use it today--wait for the roadmap. Second, the agent gets your real login state, so the permission boundary is yours to manage. Don't hand an agent your production admin cookies to run untrusted tasks; exposing high-privilege site logins to an agent equals handing over the keys. Tier by site and isolate sensitive operations with a Space. Third, 9,335 stars with a +4,912 weekly jump is impressive, but the project was created 2026-04-16 and is still young--interface stability and doc completeness are early-stage, so don't bolt it straight onto production. Fourth, "fewer tokens, faster" is the README's wording; this site didn't independently benchmark it--run your own tasks before concluding.

6. FAQ

Q: Are ego-lite and browser-use the same thing? A: No. browser-use is a browser automation framework (a Python library) that spins up a clean browser to drive, and login state doesn't carry over; ego-lite is a browser app designed for "human + agent sharing," where the agent reuses your real login state and tabs and runs in parallel in its own Spaces. Different shapes--the README explicitly positions ego-lite against browser-use.

Q: Can the agent really use sites I've already logged into? A: Yes. ego-lite's core mechanism is login-state sharing: the cookies/sessions you log in with in the ego-lite browser are directly reachable by the agent through the ego-browser skill, no re-login. Enterprise consoles, subscription sites, 2FA services all work. But correspondingly, the agent gets your real credentials, so the permission boundary is yours to manage.

Q: Does it support Windows/Linux? A: As of 2026-08-09, no. macOS only today (.dmg in Apple Silicon and Intel builds). Windows and Linux are on the roadmap with no usable build.

Q: How do I install it, and is it paid? A: Two paths: download the macOS app (.dmg from the GitHub repo), or npx skills add citrolabs/ego-lite to install only the ego-browser skill. MIT license, open source and free.

Q: How many stars, and is the project mature? A: 9,335 GitHub stars, #9 on this week's Trending weekly chart (+4,912), but the project was created 2026-04-16 and is still young; interfaces and docs are early-stage. Run it through your own tasks before production integration.


References

This article is AI-assisted and human-edited. Last updated: 2026-08-09

FAQ

Are ego-lite and browser-use the same thing?
No. browser-use is a browser automation framework (a Python library) that spins up a clean browser to drive, and login state doesn't carry over; ego-lite is a browser app designed for human + agent sharing, where the agent reuses your real login state and tabs and runs in parallel in its own Spaces. Different shapes--the README explicitly positions ego-lite against browser-use.
Can the agent really use sites I've already logged into?
Yes. ego-lite's core mechanism is login-state sharing: the cookies/sessions you log in with in the ego-lite browser are directly reachable by the agent through the ego-browser skill, no re-login. Enterprise consoles, subscription sites, 2FA services all work. But correspondingly, the agent gets your real credentials, so the permission boundary is yours to manage.
Does it support Windows/Linux?
As of 2026-08-09, no. macOS only today (.dmg in Apple Silicon and Intel builds). Windows and Linux are on the roadmap with no usable build.
How do I install it, and is it paid?
Two paths: download the macOS app (.dmg from the GitHub repo), or npx skills add citrolabs/ego-lite to install only the ego-browser skill. MIT license, open source and free.
How many stars, and is the project mature?
9,335 GitHub stars, #9 on this week's Trending weekly chart (+4,912), but the project was created 2026-04-16 and is still young; interfaces and docs are early-stage. Run it through your own tasks before production integration.

Related

Open Source

Sequence In, Structure Out: Why Did a 62-Star Repo Make Nature?

structRFM is a structure-guided RNA foundation model from the MIRACLE lab at USTC; the paper appeared in Nature on November 5, 2026 (DOI 10.1038/s41586-026-11210-2). The repository heqin-zhu/structRFM is MIT licensed and showed 62 stars in a November 6, 2026 GitHub API check; it was created in February 2025 and its README still points to the August 2025 bioRxiv preprint. The team says the roughly 21 million sequence-structure pairs, model weights and source code are public: the dataset is built from RNAcentral and BPfold with a 512-token length cap and is 4.5 GB on Zenodo. SgMLM structure-guided masked language modeling and MUSES multi-source structure ensembling let the model emit sequence-level, nucleotide-level and pair-level features from sequence alone, supporting secondary and tertiary structure prediction, splice site prediction, IRES identification, ncRNA classification and zero-shot homology classification.

Nov 6, 20269 min read
Open Source

17 Seconds on a CPU: Can an API Prove Which Model Answered?

anthropics/FlashProve is an open-source research prototype for zero-knowledge verifiable inference from Anthropic (Apache-2.0, 843 stars per a November 5, 2026 GitHub API check, version 0.16.0, officially labeled a research artifact: not maintained and not accepting contributions). The three-step protocol: the provider publishes a weight commitment ID offline (about 39 minutes on one H100), attaches a GPU-generated proof to each reply (about 19MB per 1024 tokens), and anyone verifies with cp_verify on a plain CPU in seconds. It runs Llama 4 Maverick at full 400B size on a single H100 at about 110 tokens per second of proof prefill with about 17-second CPU verification, at a cost of roughly 70 to 95 times normal inference GPU time; soundness is formally proved in Lean 4. It proves provenance, not correctness, and no major provider ships proofs yet.

Nov 5, 20269 min read
Open Source

Even Failed Experiments Trained It: An MIT Science LLM

On November 2, Deep Principle released and open-sourced Hydrogen-1.0: a 753B science LLM post-trained from GLM-5.3, under MIT, live on ModelScope and HuggingFace (deep-principle/Hydrogen-1.0, about 1.51TB BF16, plus NVFP4/MXFP4 variants, tested on 8 GPUs with SGLang). Vendor-reported FSR 38.3% beats GPT-6-Astra at 33.3% and GPQA-Diamonds 93.0% nears Kimi-K3; its signature is training on its own lab records including failed experiments, with a Lab Compile step that catches faulty experiment instructions.

Nov 4, 20269 min read