Open-Source AI Agent Brings Visual SRE Diagnostics to Self-Hosted Teams

Overview

A new open-source project is aiming to bring visual Site Reliability Engineering (SRE) diagnostics to self-hosted teams, reducing the time it takes to identify and resolve infrastructure incidents. The tool, named AIOps Cockpit, combines real-time telemetry visualization with an AI-driven diagnostic agent that operates within strict safety boundaries.

Rather than relying on proprietary SaaS platforms, this solution is designed for organizations that require full control over their data and operational environment. By leveraging large language models such as Claude 3.7, the agent can interpret natural language incident reports and correlate them with live system metrics.

The project emphasizes a human-in-the-loop workflow, ensuring that while AI assists in analysis, no remedial actions are applied without explicit human review. This approach addresses growing concerns about autonomous AI agents executing potentially destructive commands in production environments.

Core Features and Visualization

At the heart of AIOps Cockpit is its 3D Spatial Topology Map, which renders infrastructure as an interactive, living environment. Instead of parsing static text logs or 2D graphs, operators can view their systems as a dynamic space where health status is immediately apparent through color and motion.

Green indicators represent healthy nodes.
Red indicators signal critical failures or "fires" within the stack.
Spinning animations denote components that are actively failing or restarting.

This visual language allows SRE teams to grasp the scope of an incident at a glance. The interface also supports hand gesture controls via MediaPipe, enabling users to pinch to zoom, rotate the view using wrist movements, or select specific nodes by pointing. While optional, these interactions aim to reduce cognitive load during high-pressure outage scenarios.

Behind the visualization, the system provides real-time telemetry updates. Using Socket.IO, it streams live host metrics—including CPU, RAM, and disk usage—as well as Docker container states directly to the frontend. This ensures that the 3D map reflects the current reality of the infrastructure, not a cached snapshot.

AI-Driven Diagnostics with Human Oversight

The primary value proposition of the platform lies in its AI Agent Diagnostics module. Users can describe an incident in plain language, such as "checkout API 500s, DB restarting," and the LLM will investigate using a set of predefined tools. It reads logs, checks metrics, and attempts to pinpoint the root cause.

However, the system is built around a two-person rule for any proposed fixes. The AI never applies changes directly. Instead, it generates a unified diff of the suggested remediation. A human operator must review this diff and approve it before any action is taken. Furthermore, the requester of the fix cannot be the same person who approves it, adding a layer of accountability similar to financial or military protocols.

"The AI never applies a fix directly. It generates a unified diff. A human must review and approve it."

This design is a direct response to the risks associated with autonomous AI agents in production. By restricting the AI to an observational and advisory role, the platform mitigates the risk of hallucinated commands or prompt-injection attacks leading to service disruption.

Security and Defense-in-Depth

Security is treated as a foundational pillar rather than an afterthought. AIOps Cockpit implements several defense-in-depth measures to protect against both internal errors and external attacks.

Append-only audit logs: Every action, query, and suggestion is recorded immutably, providing a clear chain of custody for incident response and post-mortem analysis.
Prompt-injection defenses: The system includes safeguards to prevent malicious actors from manipulating the AI’s behavior through crafted inputs.
Strict tool allowlist: Only pre-approved commands and APIs can be accessed by the agent. Destructive actions, such as deleting resources or terminating processes, are disabled by default.

These constraints ensure that even if the AI misinterprets an incident, its ability to cause harm is technically bounded by the configuration.

Deployment and Accessibility

The project is distributed as a self-hosted application, requiring Node.js 20+ and pnpm for installation. It is structured as a monorepo with separate frontend and backend packages, allowing teams to tailor the deployment to their existing DevOps workflows.

To get started, users clone the repository, install dependencies, and configure environment variables, including their Anthropic API key or AWS credentials for telemetry integration. Once configured, a single command launches both the frontend interface and the backend services.

Component	Requirement
Runtime	Node.js 20+
Package Manager	pnpm
AI Model	Claude 3.7 (via Anthropic API)
Telemetry	Socket.IO, Docker stats

By keeping the stack open-source and self-hosted, AIOps Cockpit appeals to organizations in regulated industries or those with strict data sovereignty requirements. It offers a middle ground between expensive commercial AIOps suites and the manual toil of sifting through logs during an on-call rotation.

As the AI landscape evolves, tools that prioritize observability, safety, and user control are likely to gain traction among engineering teams wary of black-box solutions. AIOps Cockpit positions itself as a transparent, auditable, and visually intuitive companion for modern SRE practices.