AI agents

An agent that cannot load the page cannot finish the task. A site that detects a stock automated browser answers with a challenge page, a block page or an empty result, and the agent reads that page and acts on it as if it were the real one.

One HTTP call gets a browser that loads the page:

curl -i -G "https://browser.rayobyte.com/connect" \
  --data-urlencode "os=windows" \
  --data-urlencode "maxLifetime=86400" \
  --data-urlencode "proxy=http://USERNAME:[email protected]:8000" \
  -H "x-api-key: rb_live_YOUR_KEY"

The response body is a CDP WebSocket URL and the x-session-id header is the session id. Any agent framework that takes a CDP URL works, including OpenClaw, which has a full config walkthrough.

Long-running agents

Sessions survive a CDP disconnect, so an agent that restarts can pick up the same browser with its cookies and its logins intact. Keep the x-session-id value, not the br_ id inside the CDP URL, and reconnect:

curl "https://browser.rayobyte.com/connect?sessionId=br_3f8a1c9d2e4b5a60" \
  -H "x-api-key: rb_live_YOUR_KEY"

maxLifetime caps at 86,400 seconds. A browser held open for a full day burns 24 of the 48 free browser hours, so an always-on agent runs out of the included allowance in two days. Plan the budget before you build on it, and email [email protected] to raise the cap.

Reconnect rather than relaunch

An agent that creates a new browser on every task gets 429 Concurrent browser limit reached on the sixth, because the old ones are still running. Either close each browser when the task ends or hold one session and reconnect to it.