🛡️ OpsSwarm: Autonomous Infrastructure Remediation Agent

A production-grade Agentic System designed to detect, diagnose, and resolve infrastructure incidents autonomously—with strict Human-in-the-Loop governance.

---

📖 Overview

OpsSwarm is not just a chatbot; it is a state-aware Site Reliability Engineering (SRE) Agent. It connects to live infrastructure logs, reasons about root causes using Groq's LPU, and executes remediation tools via the Model Context Protocol (MCP).

Crucially, it implements a Safety-First Architecture:

1. Cyclic Reasoning: Uses LangGraph to plan and verify fixes before executing.
2. Tool Isolation: Tools run on a separate MCP Server, decoupling the AI from the OS.
3. Deterministic Guardrails: NeMo Guardrails intercept and block destructive commands (like rm -rf or DROP DB) at the embedding level.

---

🚀 Key Features

• 🧠 Cognitive Architecture: Implements a Diagnose -> Plan -> Approval -> Execute loop.
• ⚡ Sub-Second Inference: Powered by Llama-3-70b on Groq, enabling real-time log parsing.
• 🔌 Model Context Protocol (MCP): Standardized interface for connecting LLMs to local/remote tools (Docker, K8s, CLI).
• 👨‍💻 Human-in-the-Loop (HITL): Critical actions require explicit operator approval via the UI.
• 🛡️ Enterprise Security: Custom Colang flows prevent prompt injection and unauthorized actions.

---

📸 Project Walkthrough

Scenario 1: The "Happy Path" (Auto-Remediation)

The system detects a critical failure, diagnoses the root cause, and proposes a fix. It then pauses for human verification.

1. Incident Detection & Diagnosis The agent parses raw server logs, identifies Database Shard 04 Connection Refused, and formulates a plan.

2. Human-in-the-Loop Approval The workflow halts at a "Conditional Edge." The agent cannot proceed without a state update from the operator.

3. Execution & Resolution Once approved, the agent executes the restart_resource tool via MCP and verifies system health.

---

Scenario 2: The "Security Test" (Guardrails)

Demonstrating resilience against malicious prompts or accidental destructive commands.

1. Simulated Attack An operator (or prompt injector) attempts to force the agent to delete the production database.

2. Guardrail Interception The NeMo Guardrails layer intercepts the intent before it reaches the planner. The command is blocked deterministically.

---

🛠️ Tech Stack

Component	Technology	Role
Orchestrator	LangGraph	Manages the cyclic state machine and agent memory.
Inference Engine	Groq API	Provides Llama-3-70b inference at ~300 tokens/sec.
Tooling Layer	MCP (Model Context Protocol)	Standardizes tool execution (Server/Client architecture).
Safety Layer	NeMo Guardrails	Enforces security policies using Colang definitions.
Data Validation	Pydantic	Ensures strict schema compliance for all agent outputs.
Frontend	Streamlit	Provides the interactive SRE Dashboard.

---

📂 Project Structure

opsswarm/
├── config/
│   ├── settings.py         # Centralized App Configuration
│   └── rails/              # NeMo Guardrails Configs
│       ├── config.yml      # Model & Flow Definitions
│       └── security.co     # Colang Security Rules
├── src/
│   ├── graph.py            # LangGraph State Machine (The Brain)
│   ├── mcp_server.py       # Tool Provider (The Hands)
│   ├── mcp_client.py       # Tool Connector
│   └── state.py            # Pydantic Data Models
├── results/                # Demo Screenshots
├── app.py                  # Streamlit Dashboard Entry Point
├── requirements.txt        # Dependencies
└── .env                    # Secrets (API Keys)

---

Installation

1. Install Dependencies Open a terminal in the project root (opsswarm/) and run:

   pip install -r requirements.txt

2. Configure Environment Create a .env file in the root directory and add your API keys:

   PROJECT_NAME="OpsSwarm Enterprise"
   GROQ_API_KEY=gsk_your_key_here
   # Optional: LangSmith Keys
   # LANGCHAIN_TRACING_V2=true
   # LANGCHAIN_API_KEY=lsv2_your_key_here

3. Run the Application

   streamlit run app.py

---

🔮 Future Roadmap

• Kubernetes Integration: Move MCP server to a sidecar pod for K8s cluster management.
• Slack Integration: Allow approval/rejection directly via Slack channels.
• RAG Knowledge Base: Indexing past incident reports (Post-Mortems) to improve diagnosis accuracy