L2M

Appearance

Web Browse Node ​

The web_browse node drives a headless Chromium (via Playwright) to fetch a URL, wait for JS to execute, and extract structured page data: rendered HTML, page text, screenshot, title, meta tags, links, and images.

Use it whenever http_request can't see the page — single-page apps, content hidden behind client-side rendering, sites that require viewport-based layout calculations — or whenever an LLM needs to reason about a page's actual look (e.g. extracting branding).

Browser binaries must be installed once: pnpm exec playwright install chromium (same Chromium the existing pdf_output node uses).

Config ​

FieldTypeDescription
urlstringTarget URL. Required (or urlTemplate).
urlTemplatestringTemplated alternative — {{...}} resolved against upstream context.
waitUntilenum"load" / "domcontentloaded" / "networkidle" / "commit". Default domcontentloaded. Use networkidle for SPAs that hydrate after the initial HTML lands.
timeoutMsnumberTotal navigation + extraction timeout. Default 30000.
userAgentstringCustom UA string.
viewportWidth / viewportHeightnumberDefault 1280×800. Larger viewport → larger screenshot, larger HTML for some responsive sites.
extraHeadersJsonstringJSON object of extra HTTP headers (auth tokens, etc.).
screenshotboolean / "fullPage"Default true (captures viewport). "fullPage" captures the entire scrollable page. false skips entirely (smallest output, fastest).
extractTextbooleanRun document.body.innerText. Default true.
extractLinksbooleanCollect <a href> elements. Default true. Capped at 500 to bound output.
extractImagesbooleanCollect <img src> elements. Default true. Capped at 200.
waitForSelectorstringCSS selector to wait for before extraction. Use for SPAs where domcontentloaded fires before the real content lands.
outputKeystringWhere to nest the page object. Default "page".

Output ​

json
{
  "page": {
    "url": "https://example.com",
    "finalUrl": "https://www.example.com/",
    "status": 200,
    "title": "Example Domain",
    "html": "<!DOCTYPE html>...",
    "text": "Example Domain. This domain is for use in...",
    "meta": {
      "description": "Example domain for documentation",
      "og:title": "Example",
      "og:image": "https://example.com/cover.png",
      "twitter:card": "summary_large_image"
    },
    "linkTags": [
      { "rel": "icon", "href": "/favicon.ico", "type": "image/x-icon" },
      { "rel": "canonical", "href": "https://www.example.com/", "type": null }
    ],
    "links": [
      { "href": "https://...", "text": "About", "rel": "nofollow" }
    ],
    "images": [
      { "src": "https://images.example.com/logo.png", "alt": "Logo", "width": 200, "height": 80 }
    ],
    "screenshot": "data:image/png;base64,...",
    "durationMs": 1247
  },
  "url": "https://www.example.com/",
  "title": "Example Domain",
  "text": "Example Domain. This domain is for use in...",
  "html": "<!DOCTYPE html>...",
  "screenshot": "data:image/png;base64,...",
  "meta": { ... },
  "links": [ ... ],
  "images": [ ... ]
}

outputKey lets you customize the nested key. The most-used fields (url, title, text, html, screenshot, meta, links, images) are also mirrored at the top level so downstream prompt_template nodes can reference them without nesting.

Examples ​

Brand-asset extraction (the CipherTrust scenario) ​

json
{
  "type": "web_browse",
  "config": {
    "urlTemplate": "{{brandUrl}}",
    "waitUntil": "networkidle",
    "screenshot": true,
    "viewportWidth": 1440,
    "viewportHeight": 900
  }
}

Pair with a prompt_template + llm_call to extract structured branding data. Full sample at samples/workflows/branding-extract-flow.json.

Lean text extraction (no screenshot) ​

json
{
  "type": "web_browse",
  "config": {
    "url": "https://docs.example.com/api",
    "screenshot": false,
    "extractLinks": false,
    "extractImages": false
  }
}

Use for "fetch this docs page and answer a question about it" workflows where the screenshot bloats the LLM context budget.

Wait for a SPA to hydrate ​

json
{
  "type": "web_browse",
  "config": {
    "url": "https://app.example.com/dashboard",
    "waitUntil": "networkidle",
    "waitForSelector": "[data-loaded='true']",
    "timeoutMs": 60000
  }
}

waitForSelector is the reliable way to handle apps that report networkidle before their actual content lands.

Security: protocol allowlist ​

Only http, https, and data: URLs are accepted. file:// is explicitly blocked — letting workflow authors drive the headless browser at arbitrary local files would expose the API host's filesystem to anyone with workflow-edit access. (Standard SSRF threat model; same reasoning as the audit-export file-root sandbox.)

Localhost and private-network http(s) URLs are allowed — use cases include local dev and intra-cluster scraping. A hardening follow-up could add a WEB_BROWSE_URL_ALLOWLIST env var if operators need to restrict targets.

Tradeoffs ​

  • Slow per call. Launching Chromium adds ~500–800ms of overhead before the actual navigation. A persistent browser pool would cut that materially; out of scope for the MVP since most workflows visit one URL per execution. If you visit many URLs, batch them into a loop_node or accept the per-iteration cost.
  • Memory-heavy. Each invocation spawns a new browser process. Not suitable for high-RPS scenarios — keep web_browse to lower-volume report / extraction flows.
  • Screenshot bloat. A 1440×900 PNG base64 is ~80–200 KB. Set screenshot: false for high-volume text-extraction flows.
  • No JS execution. We don't expose page.evaluate() — workflow authors can't run arbitrary scripts on the target page. This is intentional (security + reproducibility); a separate web_eval node could expose it later for trusted use cases.
  • No form filling / clicking. Single-pass fetch only. Multi-step flows (login, click-through, extract) would warrant a richer web_session node — also a reasonable follow-up.