lexmount/moli

Moli – A Headless Browser Built for AI Agents

What it is – Moli is an open‑source, production‑ready headless browser written in Rust. It provides a full web‑runtime (HTML, CSS, JavaScript, networking, storage, etc.) but is designed to keep the resource footprint tiny by rendering only when a visual output is explicitly requested. The binary can be driven via a simple CLI, the Chrome DevTools Protocol (CDP), or the standard WebDriver Classic / WebDriver BiDi interfaces, so existing automation tools (Playwright, Selenium, Puppeteer, etc.) can talk to it without extra adapters.

---

Why it matters for AI / LLM agents

• Structure‑first workflow – Most agent tasks need the DOM, CSS cascade, or network data, not a continuously painted page. Moli reads the browser state directly, skipping layout and paint unless you ask for them.
• Fast & lightweight – The default mode never does a real layout, keeping median memory around 70 MiB and start‑up latency under 35 ms, far lower than Chrome headless (≈ 770 MiB, > 150 ms).
• On‑demand visuals – When a screenshot, PDF, or geometry query is needed, Moli builds a single layout tree, renders the frame, then discards the paint state. No persistent compositor or GPU work.
• Unified protocol endpoint – One binary serves CDP, WebDriver Classic, and WebDriver BiDi, so you can pick the automation library you already use.

---

Core features (as described in the README)

Feature	What it does
Complete web runtime	HTML5 parsing, native DOM, V8 JavaScript, CSS cascade, fetch/XHR/WebSocket, cookies, WebCrypto, IndexedDB, OPFS, iframes, workers, etc.
Extraction‑optimized outputs	moli fetch can dump HTML, Markdown, JSON, a semantic‑tree text format, or raw screenshots/PDFs.
CLI, CDP, WebDriver Classic & BiDi	Same kernel behind all three; Playwright can connect via chromium.connectOverCDP.
On‑demand layout & paint	--layout enables real geometry, hit‑testing, coordinate input, screenshots, and screencast frames; otherwise only a deterministic mock layout is kept.
Fine‑grained resource control	Flags to enable optional families (--image, --font, --audio, …) and to point to custom profile, cache, or cookie files.
Cross‑platform	Pre‑built binaries for Linux, macOS, and Windows.
Open‑source licensing	Dual‑licensed under Apache‑2.0 or MIT.

---

Typical use cases for AI agents

1. Web‑search & crawling – Fetch a page, wait for JavaScript to settle, and return a clean Markdown or JSON representation for downstream LLM reasoning.
2. Data extraction – Query the DOM, run a small script (--eval) to click a button or expand a section, then export the resulting structure without ever rendering a picture.
3. Interactive automation – Use the layout tree to compute element boxes, perform coordinate‑based clicks, or generate a single screenshot when visual confirmation is needed.
4. Evaluation environments – Run reinforcement‑learning or tool‑use loops where many thousands of page loads are required; the low memory/CPU cost keeps the experiment affordable.
5. Tool‑integration – Connect Playwright, Selenium, or any CDP client to the moli serve endpoint and treat it like a normal browser, but with a much smaller footprint.

---

Quick start (from the README)

# Linux / macOS – one‑liner installer
curl --proto '=https' --tlsv1.2 -fsSL \
  https://github.com/lexmount/moli/releases/latest/download/moli-installer.sh | sh

# Windows – PowerShell installer
powershell -ExecutionPolicy ByPass -c "irm https://github.com/lexmount/moli/releases/latest/download/moli-installer.ps1 | iex"

Fetching a page as Markdown

moli fetch --dump markdown --wait-until done https://example.com

Getting a semantic tree (model‑friendly)

moli fetch --dump semantic_tree_text --wait-selector body https://example.com

Screenshot on demand

moli fetch --layout --dump screenshot https://example.com > page.png

Running an automation server

moli serve            # DOM‑first only
moli serve --layout   # enable geometry & screenshots

---

How it works under the hood (high‑level)

• Rust kernel – The browser is a native Rust program, not a Chromium wrapper.
• Networking – libcurl handles HTTP(S) and multi‑request concurrency.
• HTML parsing – html5ever builds the DOM tree.
• JS engine – rusty_v8 embeds V8 for full JavaScript support.
• CSS – Servo/Stylo provides selector matching and cascade computation.
• Layout & text – Taffy (box layout) + Parley (text shaping) generate geometry only when --layout is requested.
• Software rendering – AnyRender/Vello (CPU) plus usvg and the Rust image ecosystem produce PNG/PDF frames.
• State model – The DOM + computed style are the single source of truth. When a visual operation is needed, a temporary working layout tree is built, frozen into a compact FrozenLayoutTree, used for the operation, then discarded. No persistent compositor or GPU pipeline exists.

---

Benchmarks (as reported)

Benchmark	Moli	Chrome Headless
Public‑web crawl (192 URLs)	53.6 % useful pages, median 1.43 s, 73 MiB RSS	52.6 % useful, 1.43 s, 773 MiB RSS
Sample agent workload	CDP ready 34.85 ms, peak 102 MiB PSS	CDP ready 169.37 ms, peak 349 MiB PSS
Lexbench‑Headless‑Browser (1,308 tasks)	81.9 % success, median CPU 100.6 ms, median memory 92 MiB	99.8 % success, median CPU 687 ms, median memory 697 MiB
WPT test suite	1.612 M tests passed (full run)	–

These numbers illustrate that Moli achieves ≈ 15 % of Chrome’s CPU time and ≈ 13 % of its peak memory while still handling the majority of real‑world pages.

---

Project status & ecosystem

• Production‑ready – The README states the browser is ready for production workloads and is under active development.
• Relation to Lexmount – Moli is the open‑source core; Lexmount also offers a managed cloud service (Lexmount Browser) that builds on top of it, but the binary works fully standalone.
• Community – Discord community link, Trendshift badges, and a star‑history graph indicate an engaged user base.
• License – Dual‑licensed Apache‑2.0 / MIT, with third‑party components listed separately.

---

TL;DR for a non‑expert

Moli is a fast, low‑memory headless browser you can control from code or the command line. It gives AI agents the full power of a modern web page (JavaScript, CSS, networking) but skips the heavy visual rendering unless you explicitly ask for a screenshot or geometry. This makes it ideal for large‑scale web crawling, data extraction, or any AI workflow that needs to “read” the web without the overhead of a full Chrome instance.