Skip to content

armgit / moli

Public Imported from GitHub
0

Best headless browser for AI agents. Lite, Fast, High-Compatibility. Built in Rust

Better README

Donough Liu796401abd8ee
diff --git a/README.md b/README.md
index 457060898..fef8f0d50 100644
--- a/README.md
+++ b/README.md
@@ -1,50 +1,25 @@
<p align="center">
<img
- src="assets/moli-browser-banner.png"
- alt="Moli Browser — Fast, Lightweight, Headless"
+ src="assets/moli-browser-banner.jpg"
+ alt="Moli Browser — Structure first. Pixels on demand. Open source browser for AI agents."
width="1086"
/>
</p>
<h1 align="center">Moli</h1>
-<p align="center">
- <strong>A browser engine built for AI agents.</strong>
-</p>
-
-<p align="center">
- Real JavaScript, DOM, and browser APIs · One-shot, non-retained layout and CPU rendering ·
- CLI, CDP, WebDriver Classic, and WebDriver BiDi
-</p>
+Moli is a production-ready, structured-first browser engine for AI agents.
-Moli is a headless browser built from the ground up in Rust for AI agent.
-It runs real page JavaScript on V8, maintains a native DOM and CSS
-state, and handles storage and networking — exposing the page through
-structured data first, and pixels only when you ask for them.
+It runs real JavaScript, DOM, and browser APIs by default, and computes layout
+or pixels only when requested.
-**DOM and structured page data are on by default; pixels are opt-in.**
-Reading a document, extracting Markdown, walking an accessibility tree, or
-running JavaScript doesn't require keeping a GUI browser's layout, paint, and
-compositor machinery alive. When geometry or pixels do matter, Moli runs a
-fresh layout pass and renders a frame on demand.
+Use it through CLI, MCP, CDP, WebDriver Classic, or WebDriver BiDi.
## Why Moli
-Conventional browsers assume any loaded page might need to be displayed,
-animated, or scrolled at any moment, so they keep visual machinery warm at all
-times — even when an automation task just wants a title, a list of links, or
-a JavaScript result.
-
-Agents typically work in three stages:
-
-1. **Understand the page** — DOM, text, links, forms, network responses,
- storage, JavaScript state.
-2. **Act on the page** — fill, select, submit, navigate, or execute code via
- DOM-backed controls.
-3. **Look at the page, when needed** — geometry for spatial reasoning, pixels
- for a screenshot.
-
-Moli only does the visual work a given request actually needs:
+Most browser automation needs page structure, not a continuously rendered
+visual world. Moli keeps the native DOM and style state as its source of truth,
+then runs layout or software paint only for operations that need them.
| Agent request | What Moli does |
| --- | --- |
@@ -52,21 +27,6 @@ Moli only does the visual work a given request actually needs:
| Read an element's box, hit-test a point, send coordinate input | Runs one layout pass, keeps only the latest geometry snapshot |
| Capture a screenshot or refresh a screencast | Rebuilds from current DOM/style, renders one fresh frame, discards it |
-This isn't a browser with rendering stripped out — Moli still has a real DOM,
-V8, CSS, layout, text shaping, hit-testing, and software paint. What changes
-is *when* those systems run and how long their state sticks around: DOM state
-is the default; layout and pixels are paid for only on request.
-
-That cost model matters for high-density crawling, browser-use agents,
-retrieval pipelines, eval environments, and RL workloads — where startup time,
-idle memory, and per-page footprint decide how many sessions a machine can run.
-
-## On-demand, non-retained rendering
-
-Moli never keeps a visual world alive "just in case." Native DOM and Stylo
-state are the single source of truth; layout and pixels are built only when
-an operation actually needs them.
-
<p align="center">
<a href="assets/moli_ondemand_rendering_flow.svg">
<img
@@ -77,20 +37,10 @@ an operation actually needs them.
</a>
</p>
-By default, Moli uses `LayoutPolicy::Mock` and never touches the real layout
-or paint pipeline. `--layout` switches to `LayoutPolicy::OnDemand`, unlocking
-real geometry, hit-testing, coordinate input, screenshots, and screencast —
-but it still isn't a continuously rendering browser.
-
-A cold geometry request rebuilds layout once and keeps only the latest
-snapshot. Later geometry reads may reuse that snapshot; screenshots and
-screencast frames always rebuild fresh from the current DOM and style. Even
-screencast is a low-frequency repeat of the same one-shot pipeline, not a
-retained 60 FPS compositor.
-
-The payoff: extraction, DOM inspection, JS execution, and most agent actions
-skip both the CPU cost of rendering unused frames and the memory cost of
-retained layout/paint/compositor state.
+Moli still includes V8, CSS, layout, text shaping, hit-testing, and software
+paint. The difference is when visual work runs and how long its state is kept.
+This cost model targets crawling, browser-use agents, retrieval pipelines,
+evaluation environments, and reinforcement-learning workloads.
## What works today
@@ -115,9 +65,9 @@ retained layout/paint/compositor state.
overrides, structured logging, and network diagnostics.
<p align="center">
- <a href="assets/moli-devtools-rust-lang.png">
+ <a href="assets/moli-devtools-rust-lang.jpg">
<img
- src="assets/moli-devtools-rust-lang.png"
+ src="assets/moli-devtools-rust-lang.jpg"
alt="Chrome DevTools connected to Moli while inspecting and rendering rust-lang.org"
width="1200"
/>
@@ -128,6 +78,13 @@ retained layout/paint/compositor state.
<sub>Chrome DevTools connected to Moli: rendered page, live DOM, CSS, and geometry from the same browser runtime.</sub>
</p>
+## Moli and Lexmount
+
+Moli is Lexmount’s open-source browser engine. Lexmount Browser is the managed
+cloud runtime and control plane built around it.
+
+**The open-source engine is fully usable without Lexmount Browser.**
+
## Quick start
Build from the workspace root:
@@ -248,7 +205,7 @@ count.
| Lightpanda | 85 | 44.3% | 0.97 s | 40 MiB |
| Obscura | 57 | 29.7% | 1.30 s | 39 MiB |
-### Sampled internal agent episode
+### Sample agent workload
| Metric | Moli | Chromium |
| --- | ---: | ---: |
@@ -262,8 +219,8 @@ full run recorded **1.612 million passing tests**.
## Project scope
-Moli is in active development — a practical agent browser, not a drop-in
-replacement for every Chrome feature.
+Moli is production-ready for its documented agent-browser scope and remains
+under active development.
Current intentional boundaries:
diff --git a/assets/moli-browser-banner.jpg b/assets/moli-browser-banner.jpg
new file mode 100644
index 000000000..675dec169
Binary files /dev/null and b/assets/moli-browser-banner.jpg differ
diff --git a/assets/moli-browser-banner.png b/assets/moli-browser-banner.png
deleted file mode 100644
index 7ce0130c4..000000000
Binary files a/assets/moli-browser-banner.png and /dev/null differ
diff --git a/assets/moli-devtools-rust-lang.jpg b/assets/moli-devtools-rust-lang.jpg
new file mode 100644
index 000000000..5551e1c28
Binary files /dev/null and b/assets/moli-devtools-rust-lang.jpg differ
diff --git a/assets/moli-devtools-rust-lang.png b/assets/moli-devtools-rust-lang.png
deleted file mode 100644
index 759881265..000000000
Binary files a/assets/moli-devtools-rust-lang.png and /dev/null differ
diff --git a/moli/src/cli.rs b/moli/src/cli.rs
index 6ef9696e0..e2519ad20 100644
--- a/moli/src/cli.rs
+++ b/moli/src/cli.rs
@@ -48,7 +48,7 @@ const DUMP_MODES: &[&str] = &[
#[command(
name = "moli",
version,
- about = "A Rust headless browser runtime scaffold",
+ about = "A structured-first headless browser engine for AI agents",
subcommand_required = true,
disable_help_subcommand = true
)]
@@ -502,6 +502,16 @@ fn normalize_dump_flag(args: &mut Vec<OsString>) {
mod tests {
use super::*;
+ #[test]
+ fn help_uses_the_public_product_description() {
+ let help = Cli::try_parse_from(["moli", "--help"])
+ .unwrap_err()
+ .to_string();
+
+ assert!(help.contains("A structured-first headless browser engine for AI agents"));
+ assert!(!help.contains("scaffold"));
+ }
+
#[test]
fn delay_ms_infers_fetch_command() {
let args =