OpenAI Agent SDK Real-Time Monitoring Dashboard: A Self-Hosted AI Workflow Visualization Solution

This article introduces an open-source OpenAI Agent SDK real-time monitoring dashboard designed to address the observability challenges of AI agent workflows. The dashboard receives SDK tracking events via Redis Pub/Sub, uses WebSocket for real-time push, and provides hierarchical permission control. It is a self-hosted AI workflow visualization solution that fills the gap in observability for the OpenAI Agents SDK in production environments.

Traditional monitoring tools target deterministic systems and focus on metrics like CPU and memory, but AI agent behavior is non-deterministic, with execution paths varying based on input. There is a need to understand business logic-level information such as "what the agent is doing". Additionally, AI agents may handle sensitive data, so self-hosted solutions can avoid compliance risks associated with third-party SaaS services, which is an essential need for enterprises.

The core design concept is "operational visibility without data leakage", allowing operations staff to grasp agent status while controlling access to sensitive information through hierarchical permissions. The project is positioned as a lightweight real-time monitoring dashboard, focusing on visualizing the current running status. It does not replace professional tracking and analysis tools, nor does it store long-term historical data or customer information.

System data flow: OpenAI Agents SDK generates tracking events → custom processor normalizes them and sends to Redis Pub/Sub → Dashboard FastAPI service subscribes to Redis channels and maintains an in-memory replay buffer → pushes events to browser UI via WebSocket. Architecture advantages: Redis decouples the SDK and dashboard, the in-memory buffer avoids long-term storage issues, and WebSocket ensures real-time updates.

Two access levels are distinguished: Viewer Token (read-only view for operations staff, showing operational metrics like active status and execution progress); Developer Token (detailed view for developers, allowing viewing of span details for debugging). The hierarchical design balances the needs of operations and development teams while controlling access to sensitive information.

Docker Compose configuration is provided for quick system startup, with default binding to 127.0.0.1:8090. For production environments, it is recommended to place a TLS reverse proxy in front of the frontend to ensure WebSocket security, and all settings are configured via environment variables. Workflow visualization can be customized via JSON files, and tenants can specify the configuration path.

Operations teams can grasp agent running status in real time; development teams can debug and optimize workflows (identify bottlenecks, optimize prompts and tool calls); compliant enterprises can meet data sovereignty and privacy protection requirements through self-hosting.

This dashboard accurately addresses the core need of "real-time understanding of what the agent is doing" and does not aim to be an all-in-one APM system. As AI agent production applications become more prevalent, such tools will grow in importance, and teams already using the OpenAI Agents SDK should give it a try.