lexmount/moli

The problem

Agents that browse the web usually get there through Playwright or Puppeteer driving a full Chromium. That works, but every page pays for a visual world to exist even when the agent only parses the DOM, walks links, or extracts text. At crawl scale the cost is memory and CPU that go to work nobody asks for. Moli is a headless browser that treats the DOM and style state as the source of truth and only builds layout or paint when an operation actually requires them.

What's different

The usual route is a general-purpose browser plus an automation driver layered on top, with the rendering pipeline always running. Moli inverts that: page extraction, DOM queries, JS evaluation, network and storage inspection read runtime state directly, screenshots rebuild a frame and then discard paint state, and screencast polling compares generation metadata before emitting anything. It also speaks CDP, WebDriver Classic and WebDriver BiDi from the same endpoint, so Playwright can connect over CDP rather than requiring a bespoke client. The trade-off is that anything needing a persistent visual state is rebuilt rather than maintained, which is fine for extraction and awkward for workloads that poll visual output constantly.

Who it's for

Teams running browser-use agents, crawling pipelines, retrieval jobs, evaluation harnesses or RL environments where the page is read far more often than it is seen. If your agent only needs Markdown, a semantic tree, or a screenshot on demand, this fits. If you need pixel-perfect rendering fidelity, heavy video/animation testing, or a long-lived rendered session pushed to a human, stick with stock Chromium or a cloud browser offering; the README gives no compatibility matrix, so verify your target sites yourself.

Cost of adoption

Apache-2.0, with prebuilt binaries installed via a one-line curl or PowerShell installer; Linux, macOS and Windows are claimed. The README is detailed on the CLI and protocol surfaces but thin on performance data, compatibility caveats and production hardening, so the "Lite, Fast, High-Compatibility" claims need your own benchmarking. Optional resource fetching and layout are explicit flags, meaning defaults may surprise anyone expecting Chrome parity. It is also very young — created 2026-08-10 — so expect API and behavior churn.

Verdict

Moli targets a real inefficiency: rendering the web for agents that only read it. The on-demand layout model is a legitimate architectural bet, not a wrapper, and shipping CDP plus both WebDriver dialects lowers the switching cost considerably. But the signals are those of a spike rather than an institution: 75 days old, roughly 20k stars, and +18,590 in a single month on the Rust board without any Hacker News discussion. Treat it as promising and benchmark it on your own crawl targets before putting it in a production pipeline; watch whether the project survives its own growth.