Skip to main content

Architecture

Claritty is designed as a decentralized, multi-agent system. It leverages a modern "Hub and Spoke" architecture to scale across dozens of clusters while keeping AI reasoning close to the metal.

Hub and Spoke Model​

At the core of Claritty is the separation between the SRE Agent (the spoke) and the Centralized Hub.

Cluster A (prod) ──► claritty-agent ─┐
Cluster B (dev)  ──► claritty-agent ─┼──► Centralized Claritty Hub (UI + DB)
Cluster C (eu)   ──► claritty-agent ─┘
  • Agents: Lightweight daemons running inside your Kubernetes clusters. They are stateless, completely isolated, and execute the heavy AI reasoning locally using your configured LLM (Ollama, Groq, Mistral, OpenAI, Anthropic).
  • Hub: A centralized dashboard that aggregates the incident reports generated by the agents. The Hub does not have access to your raw cluster metrics or logs, guaranteeing a Zero-Trust security boundary.

The 6-Stage AI Pipeline​

When an agent detects a failure (e.g., a CrashLoopBackOff, a network partition, or API throttling), it triggers Claritty's sophisticated 6-stage AI pipeline.

Each stage represents a specialized, autonomous agent:

  1. 🩺 Triage Agent
    Analyzes high-level cluster states, identifies failing nodes or pods, and scopes the incident.
  2. 📊 Metrics Agent
    Gathers granular CPU, memory, and custom metrics for the affected resources to detect spikes or resource starvation.
  3. 📜 Log Analysis Agent
    Parses raw container logs, extracting stack traces, fatal errors, and contextual warnings.
  4. 🏗 Infrastructure Agent
    Evaluates the underlying Kubernetes architecture, checking for misconfigured Deployments, PVC capacity issues, or Service routing failures.
  5. 📖 Runbook Agent
    Consults built-in, battle-tested YAML runbooks to map the discovered symptoms to known failure modes.
  6. ⚡ Commander Agent
    Synthesizes the findings from all previous agents to generate a human-readable RCA and proposes exact kubectl remediation commands.
info

Safety Guarantee: The Commander Agent validates all proposed commands against a strict, predefined allowlist. Destructive actions are flagged and heavily restricted.