ZeroDOM

Phân tích một trang web trực tiếp thành đồ thị tương tác tối ưu token cho các tác nhân trình duyệt AI — giảm 71% token so với chế độ ảnh chụp ARIA của Playwright. Mọi bộ chọn CSS đều được kiểm tra tính duy nhất trên toàn tài liệu thực tế trước khi trả về, không giả định từ một thuộc tính.

Xem trên vexralabs.com

Tài liệu

Suggest an edit

99.0%

selector resolution accuracy

98.9%

median token savings vs raw HTML

10-30ms

deterministic parse, no model call

Two workflows, one graph

The parse is identical either way, a flat list of every element the page can act on. What you pipe it into is the difference.

AI agent & automation engineers

Feed a token-dense graph into your LLM loop (LangChain, LlamaIndex, or a hand-rolled loop) as compact text or a JSONL node stream. Numbered handles stay stable across turns; the selector is resolved on your side, never in context.

98.9% fewer tokens than raw HTML
JSONL straight into the loop: zerodom inspect --pipe
MCP server for Claude / Cursor, no headless bot walls

Security operators & bug-bounty hunters

Drive your real, logged-in Chrome through the extension to reach authenticated surfaces, or spawn a zero-footprint throwaway Chrome over a raw CDP pipe. Match the parsed surface against deterministic rules that make no network calls of their own.

Real session via the relay + extension (sidesteps bot walls)
Zero-footprint recon: zerodom inspect --stealth
Deterministic client-side rules: zerodom scan

From the terminal

Unix-native. Read a page, stream its nodes as JSONL, or scan a list of authorized targets against your ruleset, stdin in, structured lines out, ready for jq or an agent loop.

# Token-dense interaction graph + a savings report
zerodom inspect https://example.com

# Flat node array as JSONL, one node per line — straight into an agent loop
zerodom inspect --pipe https://example.com

# Scan targets you are authorized to test against surfaces.yaml; findings as JSONL
cat targets.txt | zerodom scan -

# Zero-footprint recon: throwaway-profile Chrome over a CDP pipe, no debug port
zerodom inspect --stealth https://example.com

--stealth spawns a fresh, throwaway-profile Chrome and speaks CDP over inherited file descriptors 3 and 4, no localhost debugging port, nothing left on disk. To drive your real, logged-in browser instead (authenticated pages, existing sessions), use the Chrome extension and relay.

The problem

An agent driving a browser gets two action spaces today. Both are bad. Pixels are slow and produce coordinates that go stale on scroll. The accessibility tree is cheaper but enormous, with no stable handles. ZeroDOM is a third option.

See it in action

Parse any page and get a flat graph of every clickable and fillable element. The model sees numbered handles, not CSS selectors, not raw HTML, not coordinates.

<form class="login-form" id="auth">
  <div class="field">
    <label for="email">Email</label>
    <input type="email" id="email" required>
  </div>
  <div class="field">
    <label for="pass">Password</label>
    <input type="password" id="pass">
  </div>
  <button class="btn-primary">Sign In</button>
  <a href="/forgot">Forgot password?</a>
</form>
<nav class="sidebar">
  <a href="/dashboard">Dashboard</a>
  <a href="/settings">Settings</a>
</nav>
<!-- 200+ more lines -->

[01] textbox 'Email'
[02] textbox 'Password'
[03] button 'Sign In'
[04] link 'Forgot password?'
[05] link 'Dashboard'
[06] link 'Settings'

The model sees [03] and clicks. ZeroDOM resolves button.btn-primary on your side, selector never enters context.

97%fewer tokens for the same six actions (~3,000 → ~80)

How it compares

Feature	Raw HTML	ARIA snapshot	ZeroDOM
Tokens per page (typical)	~120,000	~8,000	~2,000
Unique node ids	No	No	Yes
Stable across refactors	No	Rarely	Yes (#id)
Selectors in context	Yes	Yes	Never
Deterministic output	N/A	No	Yes
Latency	N/A	~50ms	10-30ms

How it works

The selector never leaves your process. A model only ever sees a node id like [03], the CSS path that actually addresses the element lives in selector_map(), on your side, and gets resolved right before the click.

Live page

a Playwright Page, or HTML you already have

ZeroDOMParser

one DFS pass: prune, collect, label

InteractionGraph

compact text or JSON, sent to the model

Model picks a node

e.g. click [03], never a selector

selector_map()[node]

resolved on your side, not the model's

Act

click/fill via Playwright, then re-read