OpenLIT Controller

The OpenLIT Controller is a lightweight agent that discovers and instruments AI applications without requiring any code changes, SDK integration, or container image modifications. It runs alongside your applications and uses eBPF for LLM Observability and Python SDK injection for Agent Observability.

​

Why the Controller?

Traditional observability requires adding SDKs to every service — changing code, rebuilding images, and redeploying. The Controller eliminates all of that:

​

How It Works

The Controller provides two types of observability:

​

LLM Observability (eBPF-based)

Uses eBPF to intercept LLM API calls at the network level — no application changes required. Captures:

• Model name, provider, and endpoint
• Token usage (prompt + completion)
• Request latency and error rates
• Cost estimation

​

Agent Observability (SDK injection)

Automatically injects the OpenLIT Python SDK into running Python applications to capture:

• Agent framework spans (LangChain, CrewAI, LangGraph, etc.)
• Tool calls and chain-of-thought traces
• Vector database operations

​

Deployment Modes

The Controller auto-detects its environment and adapts accordingly:

​

Architecture

1. Service Discovery — The Controller scans for processes making LLM API calls (OpenAI, Anthropic, Gemini, Bedrock, etc.)
2. Agents — Discovered services appear in the OpenLIT dashboard where you can enable/disable observability
3. LLM Observability — One-click eBPF instrumentation for LLM traffic metrics
4. Agent Observability — One-click SDK injection for agent framework traces
5. Reconciliation — The Controller automatically restores your desired state after pod restarts, container recreates, or process restarts

​

The Agents Page

The Agents page is the dashboard page where you manage all Controller-discovered services. Once the Controller is running, navigate to Agents in the OpenLIT sidebar.

​

Services view

The main view shows every service the Controller has discovered across your infrastructure:

​

Enabling observability

Click Enable next to any service. The Controller will:

1. LLM Observability — Attach eBPF probes to intercept LLM API traffic. Takes effect within seconds, no restart required.
2. Agent Observability — Inject the OpenLIT Python SDK. In Kubernetes this triggers a rolling update; in Docker the container is recreated; on Linux the systemd service is restarted.

​

Disabling observability

Click Disable to remove instrumentation:

• LLM Observability — eBPF probes are detached immediately. No application restart.
• Agent Observability — The injected SDK environment variables are removed. In Kubernetes a rolling update restores the original pod spec. In Docker the container is recreated without the SDK. On Linux the systemd drop-in is removed and the service is restarted.

​

Service detail page

Click on a service name to see its detail page:

• Controller instance — Which Controller node is managing this service
• Workload info — Namespace, deployment name, container image (Kubernetes), container ID (Docker), or PID/executable (Linux)
• Instrumentation status — Current state for both LLM and Agent Observability
• SDK version — The OpenLIT Python SDK version injected (for Agent Observability)
• Last seen — When the Controller last reported this service

​

Controller instances

The Controller tab shows all running Controller instances with their:

• Node/host name
• Deploy mode (Kubernetes, Docker, Linux)
• Poll interval and last poll time
• Number of services discovered
• Version

​

Troubleshooting

​

Getting Started