Koko-only SDK APIs live behind the LP.* CDP namespace — designed for AI agents and high-density crawling where Playwright's DOM queries are too verbose or fragile.

Token-efficient extraction

const md = await page.markdown();
const tree = await page.semanticTree({ format: "text", maxDepth: 4 });
const meta = await page.getStructuredData();
const forms = await page.detectForms();
const links = await page.links();

NodeHandle (stable backend nodes)

Semantic tree and form detection return backendNodeId handles — more stable than CSS for agent loops.

const [search] = await page.findElement({ role: "combobox", name: "search" });
const input = page.node(search.backendNodeId!);
await input.fill("Koko browser");

const submit = await page.waitForSelectorHandle('input[name="btnK"]');
await submit.click();

NodeHandle actions

Method	CDP backing
`fill(text)`	`LP.fillNode`
`click()`	`LP.clickNode` ⚠️ hang risk on some buttons
`hover()`	`LP.hoverNode`
`press(key)`	`LP.pressKey`
`selectOption(v)`	`LP.selectOptionNode`
`check()` / `uncheck()`	`LP.setCheckedNode`

Google SERP workflow

const serp = await page.searchGoogle({
  query: "zig language tutorial",
  limit: 5,
});
console.log(serp.results); // { title, url }[]
// Includes TTFX probe, block detection (/sorry, captcha), pathHint diagnostics

Session persistence

import { captureSessionState, restoreSessionState } from "koko-sdk";

const state = await captureSessionState(page, "https://example.com");
await restoreSessionState(page, state);

JavaScript dialogs

Pre-arm dialogs before triggering JS — headless auto-dismiss breaks reactive CDP:

await page.armDialog({ accept: true, promptText: "hello" });
await page.evaluate("prompt('Name?')");

Crawl extract

const result = await page.extract();
// TTFX probe + structured extract optimized for crawlers

Agent extraction APIs