Browser Use Orchestrator icon

Browser Use Orchestrator

Pilote browser-use-cdp-orchestrator (agents web, sessions, jobs, profils)

Actions16

Agent → Run Session Step

AI-generated

Summary

Runs one natural-language step of a multi-step browser-use agent in a persistent browser session, keeping agent memory and browser state (page, cookies, logins) between steps. Returns the step result, or a job state with a job_id if it isn't finished.

Inputs

  • Task (required) — The step to run, in natural language, with the expected result. You can refer to earlier steps. Include a URL on the first step or when switching sites.
  • Session ID — Session returned by the previous step. If omitted, a new persistent session opens. Use it to chain steps.
  • Vision — Sends screenshots to the LLM on each step. Slower and costlier. Fixed at the session's first step and ignored afterwards.
  • Screenshot — Captures a full-page JPEG at the end of the step and adds its URL as result.screenshot_url. Can change per step.
  • Model — LLM model id for the session. Fixed at the first step. Defaults to OPENAI_MODEL from the server's .env.
  • Fields — JSON object { field_name: description } defining the structure of this step's results (max 12 fields; optional type prefix such as "number: price"). If omitted, results are free text.
  • Secrets — Names of server-side secrets (see List Secrets) made available to the agent. Values never pass through the LLM. Fixed at the first step.
  • Allowed Domains — Domains the browser may navigate to. Fixed at the first step, when the browser is created.
  • Profile ID — Browser-use profile to load, for persistent cookies and logins. Fixed at the first step.
  • Proxy Country — Proxy country code, e.g. "fr". Fixed at the first step.
  • Wait Seconds — Maximum server-side synchronous wait (0–600). 0 launches without waiting. In async mode it is forced to 0.
  • Execution Mode — Synchronous (default) launches the job and then internally awaits it until it finishes or Max Wait expires. Asynchronous returns immediately with a job_id to follow via Await Job.
  • Max Wait (Seconds) — Synchronous mode only (default 300). Total time to wait. If the job isn't done, the current state with job_id is returned and the job keeps running server-side.

Output shape

One JSON item per input item: the job state/result object from the orchestrator. It includes the result (free text or structured per Fields) and the job_id, and normally the session_id to pass to the next step. The exact result fields are defined server-side and don't appear in the source.

Job statuses are done, error and cancelled. If Max Wait expires in sync mode, or in async mode, the output is a non-terminal state with a job_id; fetch the result with Await Job. With Continue On Fail, errors are returned as {error: message}. The step needs the browserUseOrchestratorApi credential (base URL).

Examples

Example 1: Open a site and extract structured data in the first step of a session

Task: "Go to example.com/shop and list the first 3 products"; Fields: {"title":"product name","price":"number: price in euros"}; Allowed Domains: ["example.com"]; Execution Mode: Synchronous.

Example 2: Follow up in the same browser

Task: "Now open the second result and give its price"; Session ID: {{ $json.session_id }} from the previous step.

Example 3: Launch a long step without blocking

Execution Mode: Asynchronous; then pass the returned job_id to Await Job.

Links

Discussion