PySpur — Visual Playground for Agentic Workflows

Introduction

PySpur is an open-source visual development environment for building AI agent workflows. It provides a node-based graph editor where developers can wire together LLM calls, tool invocations, conditional logic, and human-in-the-loop steps, then run and debug the entire pipeline interactively.

What PySpur Does

• Offers a drag-and-drop canvas for composing agent workflows as directed graphs
• Supports branching, looping, and parallel execution paths within a single workflow
• Integrates with multiple LLM providers including OpenAI, Anthropic, Google, and Ollama
• Provides step-level tracing and debugging with input/output inspection at every node
• Includes built-in evaluation nodes for scoring agent outputs against criteria

Architecture Overview

PySpur uses a React-based frontend with a node graph editor built on React Flow. The backend is a Python FastAPI service that executes workflow graphs by traversing nodes in topological order, handling branches and loops. Each node type (LLM call, tool, conditional, human-in-the-loop) is a plugin class. Execution state is persisted in a database so workflows can be paused, resumed, and replayed. The system supports async execution for long-running agent tasks.

Self-Hosting & Configuration

• Deploy with Docker Compose for a batteries-included setup with PostgreSQL and the web UI
• Configure LLM provider API keys in the .env file or through the web interface
• Custom tool nodes can be added by writing a Python class implementing the node interface
• Adjust concurrency limits and timeout settings in the application config
• Supports SSO and team collaboration features for multi-user deployments

Key Features

• Visual graph editor for rapid prototyping of complex agent pipelines
• Real-time execution tracing shows data flowing through each node as it runs
• Human-in-the-loop nodes for approval gates and manual input during agent execution
• Version control for workflows with diff views between iterations
• Evaluation framework built directly into the workflow for continuous quality monitoring

Comparison with Similar Tools

• Flowise — visual LangChain builder; PySpur focuses on agentic workflows with loops and human-in-the-loop
• Langflow — drag-and-drop LLM chain builder; PySpur adds execution tracing and evaluation as first-class features
• n8n — general workflow automation; PySpur is specialized for AI agent development and debugging
• Rivet — visual prompt IDE; PySpur covers the full agent workflow lifecycle, not just prompt design
• Dify — LLMOps platform; PySpur emphasizes rapid iteration speed with its visual debugging tools

FAQ

Q: Do I need Docker to run PySpur? A: Docker Compose is the recommended setup. Local development without Docker is possible but requires manual database and service configuration.

Q: Can I export workflows for production deployment? A: Yes. Workflows can be exported as JSON and loaded via the API for headless execution in production environments.

Q: Which models work with PySpur? A: Any model accessible via OpenAI, Anthropic, Google, or Ollama APIs. Custom providers can be added through the plugin system.

Q: Is PySpur suitable for production agent systems? A: PySpur is designed primarily as a development and prototyping tool. Production deployments are supported but may require additional infrastructure for scaling.

Sources

• https://github.com/PySpur-Dev/pyspur
• https://docs.pyspur.dev