Automating The Browser

Automating the browser

crawlberg interact <url> --actions '[...]' drives a real headless browser through an ordered list of actions, then captures the resulting page. Reach for it when a static scrape is not enough: the content lives behind a "Load more" button, an infinite scroll, a form submission, or JS that only runs on interaction.

Quick recipe

crawlberg interact https://example.com \
  --actions '[{"type":"click","selector":"#load-more"},
              {"type":"wait","milliseconds":500},
              {"type":"scrape"}]'

Actions run in order. The result wraps the final page state under interaction (see Output below).

Flag surface

There is no --respect-robots-txt flag on interact; it targets the one URL you point it at.

Action schema

Each action is a JSON object tagged by type (camelCase). Unknown fields are rejected, so match the shapes exactly:

There is no select action and no wait_for_selector action: to wait for an element, use wait with a selector field. Scroll direction is up or down only.

Limits (enforced — exceeding them errors)

• Max 100 actions per call.
• Single wait: max 300000 ms; total wait across all wait actions: max 300 s.
• selector max 4096 bytes; text and executeJs script max 1 MB each.
• scroll amount max 100000 px (absolute).

Output

JSON mode (default)

The result is wrapped under an interaction key. It carries the final page HTML (interaction.final_html) and a per-action record under interaction.action_results[...] — including screenshot data and executeJs return values where applicable.

crawlberg interact https://example.com \
  --actions '[{"type":"type","selector":"#q","text":"rust"},
              {"type":"press","key":"Enter"},
              {"type":"wait","selector":".results"},
              {"type":"scrape"}]' \
  --format json | jq '.interaction.final_html | length'

Markdown mode

Prints interaction.final_html directly — the raw final DOM, not converted Markdown. Use JSON mode and feed final_html to a scrape step when you need clean Markdown.

Patterns

Expand lazy content, then capture

crawlberg interact https://example.com/feed \
  --actions '[{"type":"scroll","direction":"down"},
              {"type":"wait","milliseconds":800},
              {"type":"scroll","direction":"down"},
              {"type":"wait","milliseconds":800},
              {"type":"scrape"}]'

Capture a full-page screenshot after login

crawlberg interact https://app.example.com \
  --actions '[{"type":"type","selector":"#email","text":"me@example.com"},
              {"type":"type","selector":"#password","text":"..."},
              {"type":"click","selector":"button[type=submit]"},
              {"type":"wait","selector":".dashboard"},
              {"type":"screenshot","fullPage":true}]'

Never hardcode credentials — pass them from the environment or a secrets store.

Point at an external Chrome

crawlberg interact https://example.com \
  --browser-endpoint ws://browser.internal:9222/devtools/browser/<id> \
  --actions '[{"type":"click","selector":"#go"},{"type":"scrape"}]'

See the headless-fallback skill for CDP endpoints, wait strategies, and persistent browser profiles set via --config.

When to reach for scrape or crawl instead

If the page yields its content on a plain fetch (or via --browser-mode always) with no clicking or typing, use crawlberg scrape — see scraping-html-to-markdown. Use interact only when ordered actions must run before the content exists.

Source: hashgraph-online/awesome-codex-plugins → plugins/kreuzberg-dev/plugins/plugins/crawlberg/skills/automating-the-browser/SKILL.md