aegisdesk

AegisDesk

Enterprise Autonomous IT Intelligence — Multi-Agent Swarm for the Modern Service Desk

AegisDesk is a next-generation, multi-agent AI system purpose-built for Enterprise IT Service Desks. It goes far beyond conventional RAG chatbots by combining a deterministic intent router, an ACID-compliant semantic graph memory, and a hardened security layer — trading raw response latency for security guarantees and routing accuracy to protect mission-critical infrastructure.

---

Table of Contents

• Why AegisDesk?
• Architecture Overview
• Key Features
• Tech Stack
• Project Structure
• Getting Started
• Configuration
• Usage
• Security Model
• Development
• Testing
• Contributing
• Roadmap
• Author

---

Why AegisDesk?

Traditional IT helpdesk bots suffer from three fundamental problems:

• Routing itself is zero-token; end-to-end latency scales with tool-call depth and API tier.

AegisDesk is engineered for enterprises where accuracy, auditability, and security are non-negotiable.

---

Architecture Overview

User Query
    │
    ▼
┌─────────────────────────────────────────────────┐
│          Zero-Token Semantic Router              │
│   (Deterministic intent routing — no LLM call)  │
└──────────┬───────────────────┬──────────────────┘
           │                   │
   ┌───────▼──────┐   ┌────────▼──────────┐   ┌───────────────────┐
   │   Network    │   │  Cloud/Atlassian   │   │  Web Intelligence │
   │  Ops Agent  │   │   Infra Agent      │   │     Agent         │
   │ (ping, port  │   │ (Azure/AWS/Jira)   │   │ (Internal wikis,  │
   │  scan, proc) │   │                   │   │  HR portals)      │
   └──────┬───────┘   └────────┬──────────┘   └────────┬──────────┘
          │                    │                        │
          └────────────────────▼────────────────────────┘
                               │
               ┌───────────────▼───────────────┐
               │  ACID SQLite Semantic Graph   │
               │  + ChromaDB Vector Store      │
               │  + BGE CrossEncoder Reranker  │
               └───────────────┬───────────────┘
                               │
               ┌───────────────▼───────────────┐
               │    FastAPI SSE Streaming API  │
               │  (JWT Auth + RBAC + TTLCache) │
               └───────────────────────────────┘

The LangGraph Supervisor orchestrates all agent activity. Recursive tool calls are capped at n=5 to prevent runaway loops and protect your API budget — any breach triggers escalation to a human IT agent.

---

Key Features

🤖 Multi-Agent Swarm Architecture

Incoming queries are routed directly to specialized worker agents, eliminating the "monolithic prompt" anti-pattern:

• Network Operations Agent — OS-level diagnostics (ping, port scans, process enumeration) with strict regex-based RCE sanitization.
• Cloud Infrastructure Agent — Direct integration with Azure, AWS, and Atlassian toolchains via secured REST APIs.
• Web Intelligence Agent — Headless scraping of internal wikis and HR portals, protected against SSRF via pre-flight DNS resolution.

🧠 ACID-Compliant Semantic Graph Memory

Unlike systems that lose context on reboot, AegisDesk uses a custom SQLite-backed Semantic Graph (sqlite-vec + langgraph-checkpoint-sqlite):

• Entities and relational edges are stored persistently across sessions.
• Context is assembled recursively via Waggle-inspired edge traversal.
• A BAAI/bge-reranker-base PyTorch CrossEncoder injects the most hyper-relevant subgraph into the LLM context window without overflow.

⚡ Server-Sent Events (SSE) Streaming API

• FastAPI backend with native text/event-stream responses for real-time, ChatGPT-like UX.
• JWT Authentication and Role-Based Access Control (RBAC) on all endpoints.
• Context-Aware Global Caching: Utilizes a dynamic TTLCache key that blends query hashes with user-specific state mutations, eliminating stale responses and cache stampedes without losing performance.
• Thread-Safe Concurrency: Strict asyncio.Semaphore(10) combined with a Sliding Window Token Bucket prevents LLM API Rate Limits and OS file handle exhaustion.
• Non-Blocking Architecture: CrossEncoder inference is fully decoupled from the ASGI event loop via asyncio.to_thread, guaranteeing zero deadlocks and zero event loop saturation under high concurrent load.

🛡️ Zero-Trust Security Protocols

• RCE Prevention — shell=True is explicitly disabled. All OS inputs are stripped of shell metacharacters (&, |, ;, $, <).
• SSRF Mitigation — All web scraper requests undergo pre-flight DNS resolution. Loopback, link-local, or private subnet targets trigger an immediate block.
• Denial-of-Wallet Protection — LangGraph Supervisor counts recursive tool_calls; infinite loops are caught at n=5 and escalated to a human.
• Human-in-the-Loop — Pipeline uses interrupt_before=["dangerous_tools"] so critical OS commands are never executed blindly.
• Information Disclosure Sanitization — Raw API exceptions, internal paths, and organizational IDs are systematically masked from user output, securely logging stack traces internally while returning sanitized, generic error codes to the client.
• Async Rate Limit Resilience — The architecture utilizes custom asyncio.sleep() exponential backoffs, ensuring transient HTTP 429 errors from LLM providers are handled silently without blocking the concurrent FastAPI SSE stream.

💻 Rich CLI Interface

A typer-powered, Rich-rendered interactive CLI for headless server deployments — no browser required.

---

Tech Stack

---

Project Structure

AegisDesk/
├── src/aegisdesk/
│   ├── app/                  
│   │   ├── api/              # FastAPI endpoints (SSE streams, JWT auth, RBAC)
│   │   ├── memory/           # SQLite Semantic Graph + context assemblers
│   │   ├── rag/              # LangGraph pipeline, graph definitions, reranking
│   │   └── db/               # ChromaDB vector store (Singleton-managed)
│   ├── core/
│   │   ├── tools.py          # Sanitized subprocess tooling
│   │   ├── integration_tools.py  # Cloud & Atlassian integrations
│   │   ├── pipeline.py       # Execution engine (Human-in-the-Loop interrupt)
│   │   └── llm_factory.py    # LLMFactory — unified LLM instantiation
│   ├── cli/
│   │   └── main.py           # Rich-rendered Typer CLI
│   └── observability/
│       └── logger.py         # Structured logger (get_logger())
│
├── docs/
│   ├── architecture.md       # Deep-dive architecture documentation
│   ├── roadmap.md            # Development roadmap
│   ├── context-map.md        # File-to-task mapping for contributors
│   └── adr/                  # Architecture Decision Records (ADRs)
│       └── 0002-sqlite-plus-chroma.md
│
├── tests/                    # Pytest test suite
├── examples/                 # Usage examples
├── Dockerfile                # Container build definition
├── pyproject.toml            # Project metadata, dependencies, tooling config
├── requirements.txt          # Pinned dependency list
├── AGENTS.md                 # AI agent collaboration guide
├── CONTRIBUTING.md           # Contributor guidelines
└── test_security.py          # Standalone security validation suite

---

Getting Started

Prerequisites

• Python 3.10 or higher
• git
• (Optional) Docker

Installation

The recommended way to install AegisDesk is globally via pip or pipx:

pip install aegisdesk
# OR
pipx install aegisdesk

Developer Installation (Source)

If you want to contribute or run the test suites:

git clone https://github.com/sitanshukr08/Aegisdesk.git
cd Aegisdesk

python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate

pip install -e ".[dev]"

Docker (Alternative)

docker build -t aegisdesk .
docker run --env-file .env -p 8000:8000 aegisdesk

---

Configuration

Copy the environment template and fill in the required keys:

cp .env.example .env

Key variables:

# LLM Provider (choose one or more)
OPENAI_API_KEY=sk-...
GROQ_API_KEY=gsk_...
GOOGLE_API_KEY=...

# JWT Secret
JWT_SECRET_KEY=your-secret-key

# Cloud Integrations (optional)
AZURE_SUBSCRIPTION_ID=...
AWS_ACCESS_KEY_ID=...
ATLASSIAN_API_TOKEN=...

Only fill the keys needed for the features you are running. Unused providers can be left blank.

---

Usage

Initialize AegisDesk

# Initialize data structures, logs, and environments
aegisdesk init

Ingest IT Documentation

# Ingest HR / IT documentation into the ChromaDB vector store
aegisdesk ingest ./docs/vpn_troubleshooting.pdf
aegisdesk ingest ./docs/onboarding_guide.pdf

Ask a Question (CLI)

aegisdesk ask "Can you ping the corporate gateway and check if my Okta token expired?"

Run the API Server

uvicorn aegisdesk.app.main:app --reload --host 0.0.0.0 --port 8000

The SSE endpoint will be available at http://localhost:8000. Authenticate with a JWT token and stream responses in real time.

Diagnostics

aegisdesk doctor

---

Security Model

AegisDesk is hardened against common Red Team attack vectors:

---

Development

Coding Standards

• Keep business logic out of API route handlers and CLI commands; place shared behavior in src/aegisdesk/core/.
• Never instantiate ChatGroq or ChatOpenAI directly — always use get_llm() from src/aegisdesk/core/llm_factory.py.
• Never use print() — use get_logger() from src/aegisdesk/observability/logger.py.
• Use typed Pydantic models for all public request/response shapes.
• Prefer repository classes for persistence over direct database access in services.

Branch Naming

feature/cli-init-command
fix/cache-isolation
docs/rag-boundaries
refactor/memory-graph-edges

Linting & Type Checking

ruff check .
mypy src/

---

Testing

# Run the full test suite
python -m pytest

# Run only security tests
python test_security.py

# Run with verbose output
python -m pytest -v tests/

Tests cover cache keys, persistence, retrieval, and ticket escalation logic. All runtime behavior changes must be accompanied by tests.

---

Contributing

AegisDesk is moving from prototype to product in small, reviewable phases. Contributions should be narrow in scope and reviewable.

Every PR must explain:

1. What changed
2. Why it changed
3. What behavior changed (if any)
4. What was intentionally left out
5. What validation was run via the aegisdesk CLI

See CONTRIBUTING.md for full guidelines and see AGENTS.md for AI agent collaboration rules.

---

Roadmap

See docs/roadmap.md for the full roadmap. Current version: v0.1.6 (Phase 16).

Planned improvements include:

• Additional worker agents (HR, Procurement, Identity Management)
• OpenTelemetry-based observability tracing
• Multi-tenant RBAC with department-level scoping
• GitHub Actions CI/CD pipeline
• Formal open-source license

---

Author

Sitanshu Kumar GitHub: @sitanshukr08

AegisDesk — Turning the IT Service Desk from a cost center into an autonomous intelligence layer.