Workflow Use turns a browser task into a stored, replayable workflow. You record the task once with a browser extension, or describe it in natural language and let Browser Use complete it a single time; the tool then converts the execution history into a structured workflow file with variables extracted from forms. Later runs replay that file deterministically without an LLM, and fall back to a Browser Use agent when a step fails.
Workflows are JSON files run from a Python CLI or the workflow_use library. A workflow can run with predefined variables or as a tool, where a prompt supplies the inputs. Generated workflows are stored under workflows/storage/ with a searchable metadata.json index, and can run on Browser-Use Cloud browsers using semantic mapping with no LLM calls. A FastAPI backend plus a web UI visualizes workflows as graphs, runs them with custom parameters, and streams execution logs.
The project is in early development; the roadmap lists self-healing that updates the workflow file when a step fails, chaining step outputs into later steps, and exposing workflows as MCP tools.
Features
- Record once, replay: a browser extension captures interactions and converts them into a deterministic workflow file
- Generation mode: describe a task in natural language; Browser Use runs it once and the history becomes a parameterized workflow
- Agent fallback: a failed step hands over to a Browser Use agent instead of aborting the run
- Variable extraction: form fields become workflow variables that you fill from predefined values or from a prompt
- Workflow storage: generated workflows are saved with metadata and managed through list, info, and delete commands
- Cloud browsers: workflows run on Browser-Use Cloud with semantic abstraction and no AI calls
- Web GUI: a FastAPI backend and frontend show workflows as interactive graphs, execute them, and stream logs
- Configurable models: separate agent, extraction, and workflow-generation models per generation run