Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
ego lite is a macOS browser built so that a person and their AI agents can use the same browser at the same time without colliding. Released under MIT by Citro Labs in April 2026, it has climbed to roughly 6,080 stars and 300 forks in about three months, ships 13 tagged releases with v1.2.5 landing on 2026-07-17, and carries only 42 open issues. Its premise is a rejection of how browser automation has been packaged until now. ## The Problem It Names Browser automation for agents currently splits into two shapes, and ego lite argues both are wrong. Automation frameworks — Browser-Use, Vercel's agent-browser — are libraries the agent calls. They ship no browser of their own, so they need a separate one to drive, and the logged-in session you actually care about rarely carries over cleanly. You end up either handing the agent a blank profile that hits a login wall, or letting it drive your real browser and fighting it for tabs and focus. AI browsers — ChatGPT Atlas, Perplexity Comet — go the other way. They ship a built-in agent that inherits your Chrome data and your logins work fine, but only that agent can drive the browser. If your work already lives in Claude Code or Codex, you cannot bring it. ego lite is neither. It is one browser, designed from the start to be shared between you and whatever agent you already use. ## Spaces and the ego-browser Skill The mechanism is Spaces: fully isolated parallel workspaces inside the same browser process. Each agent, or each task, gets its own Space. You browse in the front; agents work in the background; neither steals the other's tabs and your cursor stays where you left it. The UI shows which Space currently has an agent running, and you can take over or stop it at any point. The documented pattern is genuinely parallel — Claude Code enriching ten leads in ten Spaces while Codex scrapes five competitor sites in five more. Any agent drives it through `ego-browser`, a skill that acts as the connection layer between an agent CLI and the browser. Installing the .dmg drops the skill into every agent's skills directory on the machine; alternatively `npx skills add citrolabs/ego-lite` installs just the skill, and the app install is walked through on first use. On first launch ego lite asks one question: migrate your Chrome data? Say yes and agents inherit your existing logins, cookies, extensions, and bookmarks. That is the whole login story — no session export, no cookie juggling. ## Code, Not Command Loops The design decision with the clearest performance argument is that `ego-browser` exposes the browser as in-page JavaScript functions — snapshot, fill, click, wait, navigate, capture — rather than as CLI commands. The agent writes a JavaScript snippet calling those tools and `ego-browser` executes it on the page in one pass. This matters because the conventional CLI shape traps an agent in a call-two-commands, read-the-result, call-two-more loop, and every cycle costs a round trip and a chunk of context. Letting the agent compose a multi-step task into a single code output plays to what coding models are actually good at. The project reports complex workflows finishing up to 2.5× faster with higher success rates and far fewer tool calls per task, benchmarked against Vercel's agent-browser across four complex automation tasks, with the gap widening as tasks get harder. The other technical claim is snapshot quality. Snapshots are the text view a model uses to "see" a page, and ego lite says kernel-level customization of the browser lets it handle deeply nested iframes — a case where framework-level approaches consistently break. That claim is plausible given it controls the browser rather than driving one from outside, though it is not independently verified. A planned "experience accumulation" feature would distill successful actions into reusable tools so repeat tasks run up to 5× faster. It is marked coming soon and should be read as a roadmap item, not a shipped capability. ## Pros and Cons The strengths are concrete. Owning the browser instead of driving one solves login inheritance and tab contention in a single move rather than papering over them. Spaces make parallelism real instead of a race condition. Agent-agnostic control through a skill means the project is not betting on any one vendor's CLI surviving. The code-over-CLI interface is a defensible efficiency argument, not just a benchmark artifact. MIT licensing, local data storage, and 13 releases in three months on a project with 42 open issues all point at maintainers who are shipping and closing things. And it is free, which the comparison table pointedly notes the AI browsers are not. The limitations are real, and one deserves care rather than a bullet point. Handing an agent your migrated Chrome profile means handing it your authenticated sessions — email, cloud storage, banking, anything you stay logged into. Combined with an interface that executes agent-written JavaScript on live pages, a misinterpreted instruction or a prompt injection embedded in a visited page can act as you on any site where you hold a session. Space isolation separates agents from each other and from your tabs; it does not narrow what an agent's credentials can reach. This is the cost of the login problem being solved this way, and it argues for treating the Chrome-migration prompt as a real decision rather than a setup step. Beyond that: macOS only, with Windows and Linux on the roadmap, which excludes a large share of the agent-tooling audience today. The app is distributed as a .dmg from the project's own CDN rather than through GitHub releases, so verifying what you install takes an extra step. The 2.5× and token-reduction figures are self-reported and measured against a single competitor on four tasks, so treat them as directional. The project is about three months old, and browser internals are unforgiving territory to be young in. Headline features are still marked coming soon. And because the browser is the product, adopting it is a bigger commitment than adding a library — it means changing the browser you actually work in. ## Outlook The interesting bet here is architectural rather than competitive: that the correct unit of agent browser tooling is a shared browser, not a driving library or a walled AI browser. If agents become routine coworkers on browser tasks, contention over tabs and sessions stops being an annoyance and becomes the central design constraint, and Spaces is a credible answer to it. The remaining questions are platform reach and whether the security model matures past the current all-or-nothing profile inheritance — per-Space credential scoping would make the whole approach considerably easier to recommend for sensitive accounts. ## Conclusion ego lite is worth trying for macOS users already running browser tasks through Claude Code, Codex, or Cursor, particularly anyone doing parallel work — lead enrichment, multi-site research, repetitive form filling — where tab contention and re-authentication are the actual bottleneck. The parallel Spaces model and the code-over-CLI interface are genuine improvements over driving a separate browser. Before migrating Chrome data, consider which logged-in accounts an agent would then be able to act on, and whether a fresh profile with only the sessions a task needs is the safer starting point.
OpenClaw is an open-source, local-first AI gateway with 366K GitHub stars that routes AI responses through WhatsApp, Telegram, Slack, Discord, iMessage, Teams, and 15+ other platforms — zero cloud dependency.
OpenClaw
Open-source personal AI assistant connecting to 13+ messaging platforms with local gateway architecture, voice support, and multi-agent routing.