Problem

Browser automation is fragile. Scripts break constantly and agents waste tokens getting stuck in loops.

Today there are two options:

Selector-based scripting like Playwright and Puppeteer: these require you to target specific DOM elements.

First-generation browser agents (Stagehand, browser-use): use natural language interfaces but still resolve instructions into selectors under the hood.

Selector-based scripts can break every time the UI changes. You end up maintaining selectors instead of building features. First-gen agents inherit the same brittleness, especially when they misidentify the right element.

Solution:

Lumen is vision-first. It sees the screen and acts like a human. Every natural language instruction resolves into an x,y coordinate on the screen. Three layers of stuck detection keep it on track and a dual-history system with context compaction lets it handle 20+ step workflows without blowing up the context window.

We ran a WebVoyager eval (25 tasks across 15 sites, scored by LLM judge, 3 trials per task, all frameworks on Claude Sonnet 4.6):

Lumen: 100% success rate, 77.8s avg time, ~104K tokens.

browser-use: 100% success rate, 109.8s avg time.

Stagehand: 76% success rate, 207.8s avg time, ~200K tokens.

Lumen matches browser-use on accuracy while completing tasks ~30% faster, and beats Stagehand on every metric.

Get Started

Start using Lumen today:

Docs: https://lumen.omlabs.xyz/

Support us: GitHub star: https://github.com/omxyz/lumen