AI NewsOpen sourceAnnouncement
Moli, a Rust browser for AI agents, renders pixels only on request
Moli is a Rust-built headless browser aimed at AI agents that generates layout and pixels only on request, so a page fetch that never needs a screenshot skips the paint stage entirely.
Image: GitHub
Why it mattersA team running a crawler, an evaluation harness or a browser-use agent at scale sees the same 192-page workload run in a tenth of the memory Chromium uses, according to the project's numbers.
Moli is an Apache-2.0 headless browser written in Rust and designed around AI-agent workloads. The repository was opened on 10 August 2026 by lexmount and has 1,602 stars in the three weeks since. It supports Linux, macOS and Windows, and exposes the same instance over CDP, WebDriver Classic and WebDriver BiDi from one binary.
The design assumes that most browser work an agent does is structure-only. A fetch, a DOM query, a script call or a Markdown extraction does not need pixels, so Moli defaults to a mock layout policy that returns deterministic geometry in a compatible format and never runs paint. Passing --layout opts into real box construction, hit testing, coordinate input and screenshots. Passing --resource adds image, font, video and audio fetches. Nothing costly is on by default.
The numbers the project publishes
The project posts three sets of measurements, all conducted by the authors on their own test setup. On a 192-URL public-web crawl covering large Chinese and international sites, moli says it counted 103 useful pages against Chrome Headless at 101, at the same 1.43 second median time, but using 73 MiB of resident memory against Chromium's 773 MiB. On a sample agent workload the project reports the CDP-ready time at 34.85 ms against Chromium's 169.37 ms and peak process-set memory at 102 MiB against 349 MiB.
On the wider Lexbench-Headless-Browser benchmark of 1,308 comparable tasks, the project reports Chrome at 99.8 percent success, Moli at 81.9 percent, and three other agent-oriented browsers at 62.1, 53.3 and 44.9 percent. These are the vendor's own numbers on a benchmark the vendor authored, and the strongest of them show a headless browser purpose-built for a smaller set of tasks beating browsers built for the whole open web. An independent run is the check that matters, and the Apache-2.0 licence and public benchmark corpus make one possible.
What runs the browser
Moli is a standalone kernel rather than a Chromium wrapper. It uses libcurl for network, html5ever for parsing, rusty_v8 for JavaScript, Servo's Stylo for CSS cascade, Taffy for box layout and Parley for text layout, and AnyRender or Vello CPU for the paint stage when it runs. A single moli serve command exposes CDP, WebDriver Classic and WebDriver BiDi at the same endpoint, so a Playwright script that already connects over CDP works with no code change.
An agent workflow that opens ten pages to answer one question is running Chromium ten times, and most of what Chromium spends its memory on is the visual state the agent will never look at. A browser that lets that state stay off until it is asked for is the sort of infrastructure change that lowers the cost of running agents at scale. A single machine that could hold ten Chromium instances holds many more Moli instances, on the project's numbers.
The trade-off is that the wider web will render at 81.9 percent rather than 99.8 percent, so a full Chromium has to be kept available for the pages that break. For a controlled crawl or a documented target, this is fine. For an agent working on the open web against strong anti-bot protection, Chromium is still the right answer.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.