🦊 AGIAAS — Autonomous Agent Platform

Your AI-powered on-call engineer that never sleeps.

How It Works · Features · Quick Start · Architecture · Structure

---

💡 What is AGIAAS?

AGIAAS (Autonomous GitHub Incident Analysis & Automation System) is an intelligent agent platform that monitors your GitHub infrastructure, analyzes anomalies using state-of-the-art AI, and facilitates automated fixes via a Telegram-integrated human-in-the-loop workflow.

Think of it as an on-call SRE that watches your repos 24/7 — triages incidents, writes patches, opens PRs, and waits for your approval before merging. All from your phone.

---

🧠 How It Works

When an incident hits your infrastructure, AGIAAS runs a 6-step autonomous pipeline:

┌────────────────────────────────────────────────────────────────────┐
│                    AGIAAS ON-CALL PIPELINE                        │
├────────────────────────────────────────────────────────────────────┤
│                                                                    │
│  ① MONITOR        GitHub webhook fires (issue, PR fail, CI error) │
│       ↓                                                            │
│  ② REASON         Mixture-of-Agents debate for root-cause analysis│
│       ↓                                                            │
│  ③ RESEARCH       Web search + session history for context        │
│       ↓                                                            │
│  ④ REMEDIATE      Parallel task execution (patch, restart, cache) │
│       ↓                                                            │
│  ⑤ REPORT         Telegram UI Card with diff, cost, and actions   │
│       ↓                                                            │
│  ⑥ LEARN          Auto-generated runbook for future incidents     │
│                                                                    │
└────────────────────────────────────────────────────────────────────┘

You stay in control: Every proposed fix is sent to Telegram for your explicit approval before any code is merged.

---

🚀 Key Features

Incident Intelligence

• 🎯 Proactive Monitoring — Real-time tracking of GitHub Issues, PRs, and Workflow failures via webhooks
• 🧠 Mixture-of-Agents Debate — Multi-agent reasoning for deep root-cause analysis
• 🔀 Dynamic Model Selection — Choose AI models per-repo from the dashboard (Hermes-4-405B, Claude-3.5-Sonnet, Gemini Flash) with 3-tier priority: Project → Global → Auto-detect
• 🔍 Contextual Research — Combines web search with historical session data to inform diagnostic decisions
• 📓 Auto Runbook Generation — Every resolved incident produces a structured runbook to prevent recurrence

Automation & DevOps

• 🛠️ Autonomous Patching — Branch creation, code patching, and PR generation without human intervention
• ⏪ CI/CD Rollback — One-tap rollback to the last successful workflow run
• 📡 Zero-Config Webhooks — Auto-provisions Cloudflare tunnels and syncs GitHub webhook URLs automatically
• 🔄 Multi-Provider Webhooks — GitHub, Sentry, Vercel, and Cloudflare event ingestion

Security & Observability

• 📱 Human-in-the-Loop — Interactive Telegram approval flow (✅ Approve / ❌ Reject / 💬 Re-Analyze)
• 🔐 AES-256-GCM Encryption — All tokens and webhook secrets encrypted at rest
• 💰 Real-Time Cost Tracking — Dynamic token usage and model cost calculation per incident
• 📋 Full Audit Trail — Every AI suggestion, shell command, and token cost is logged

---

🔄 Architecture

graph TD
    subgraph GitHub ["GitHub Infrastructure"]
        Events[Issues / PRs / Workflows]
        GH_API[GitHub API & CLI]
    end

    subgraph Providers ["External Providers"]
        Sentry[Sentry Alerts]
        Vercel[Vercel Deploy Fails]
        CF[Cloudflare Events]
    end

    subgraph Core ["AGIAAS Core (Monorepo)"]
        WS["Webhook Server\n(Cloudflare Tunnel)"]
        Agent["Python Agent\n(6-Step Pipeline)"]
        DB[(SQLite / Drizzle\nAES-256-GCM)]
        API["oRPC Backend\n(Type-Safe)"]
    end

    subgraph Interaction ["User Interface"]
        Web["React Dashboard\n(TanStack Start)"]
        Bot["Telegram Gateway\n(UI Cards)"]
    end

    subgraph AI ["AI Models"]
        Claude[Claude-3.5-Sonnet]
        Hermes[Hermes-4-405B]
    end

    Events -- "Webhooks" --> WS
    Sentry -- "Alerts" --> WS
    Vercel -- "Deploy Status" --> WS
    CF -- "Health" --> WS
    WS -- "Trigger" --> Agent
    Agent -- "Analysis" --> Claude
    Agent -- "Analysis" --> Hermes
    Agent -- "Approval Request" --> Bot
    Bot -- "User Action" --> Agent
    Agent -- "Execute Fix/PR" --> GH_API
    Web -- "Management" --> API
    API -- "Data" --> DB

Loading

---

🔌 Project Structure

This is a monorepo powered by Turborepo and pnpm.

Package	Path	Description
@agiaas/agent	packages/agent	Python-based AI agent with 6-step incident pipeline
@agiaas/webhook-server	packages/webhook-server	Cloudflare tunnel + webhook ingestion & auto-sync
@agiaas/web	apps/web	React dashboard (TanStack Start + Tailwind CSS)
@agiaas/docs	apps/docs	Documentation site (Fumadocs + TanStack Start)
@agiaas/api	packages/api	Type-safe backend API (oRPC + Hono)
@agiaas/db	packages/db	Database schema & migrations (Drizzle ORM + SQLite)
@agiaas/auth	packages/auth	Authentication (better-auth)
@agiaas/ui	packages/ui	Shared UI components (Radix UI + Tailwind)
@agiaas/env	packages/env	Zod-validated environment variable management
@agiaas/config	packages/config	Shared TypeScript, ESLint, and Biome configurations

---

🛠️ Tech Stack

Layer	Technology
AI Models	Hermes-4-405B (Nous Research) · Claude-3.5-Sonnet (OpenRouter) · Gemini Flash — per-repo configurable
Agent Engine	Custom Python agent with Mixture-of-Agents reasoning + dynamic model routing
Frontend	TanStack Start + Tailwind CSS + Radix UI + WebGL (ASCIIText)
Backend API	oRPC (end-to-end type safety) + Hono
Database	SQLite + Drizzle ORM + AES-256-GCM encrypted secrets
Authentication	better-auth (session-based)
Messaging	python-telegram-bot (structured UI Cards)
Infrastructure	Docker Compose, Cloudflare Tunnels (auto-provisioned)
Monorepo	Turborepo + pnpm workspaces
Code Quality	Biome (lint + format) + TypeScript strict mode

---

🚀 Quick Start

Prerequisites

• Runtimes: Node.js 18+, Python 3.12+
• Tools: pnpm, gh (GitHub CLI), sqlite3, cloudflared
• Accounts: GitHub access, Telegram Bot (via @BotFather)

Option A — Docker (Recommended)

git clone https://github.com/mehmetkr-31/agiaas.git
cd agiaas
pnpm install
pnpm setup        # Interactive — generates security keys & .env
docker-compose up --build

Dashboard → http://localhost:3678 Agent API → http://localhost:8678

Option B — Local Development

git clone https://github.com/mehmetkr-31/agiaas.git
cd agiaas
pnpm install
pnpm setup        # Interactive — generates security keys & .env
pnpm db:migrate   # Initialize SQLite database
pnpm python:install  # Set up Python venv for the agent
pnpm dev          # Start all services in parallel

Dashboard → http://localhost:3678 Agent API → http://localhost:8678

---

🔑 Environment Variables

Create a .env file in the project root (or use pnpm setup to generate one interactively).

Variable	Required	Description
DB_ENCRYPTION_KEY	✅	32-character key for AES-256-GCM secret encryption
BETTER_AUTH_SECRET	✅	32+ character key for session management
NOUS_API_KEY	✅*	Nous Research API key (or use OPENROUTER_API_KEY)
OPENROUTER_API_KEY	✅*	OpenRouter API key (or use NOUS_API_KEY)
GITHUB_TOKEN	✅	PAT with repo and workflow scopes
GITHUB_USERNAME	✅	Your GitHub username
TELEGRAM_BOT_TOKEN	📱	From @BotFather
TELEGRAM_CHAT_ID	📱	Your personal chat ID
WEBHOOK_PORT	No	Webhook receiver port (default: 8678)
PORT	No	Dashboard port (default: 3678)

*At least one AI provider key is required.

---

📱 Telegram Workflow

AGIAAS sends structured UI Cards to Telegram with rich formatting:

┌──────────────────────────────────────┐
│ ⚠️ INCIDENT DETECTED                │
│ ──────────────────────               │
│ 📂 Repo: your-org/your-repo         │
│                                      │
│ 💬 Response:                         │
│ CI pipeline failed on main branch.   │
│ Root cause: missing env variable     │
│ STRIPE_KEY in deployment config.     │
│                                      │
│ ──────────────────────               │
│ 💰 Cost: $0.0142  |  ⚡ Tokens: 3,847│
└──────────────────────────────────────┘

  ✅ Approve   ❌ Reject   💬 Re-Analyze

---

🤖 Security & Principles

1. Explicit Approval — No code is modified or pushed without a direct "Approve" from Telegram.
2. Minimal Footprint — Agents operate in isolated .tmp directories with no persistent disk access.
3. Encrypted Secrets — All tokens and webhook secrets stored with AES-256-GCM encryption.
4. Audit Logs — Every AI suggestion, command execution, and token cost is logged with timestamps.

---

📊 Cost Tracking

AGIAAS provides real-time financial tracking for AI operations:

• Per-incident cost breakdown — Input/output tokens × model pricing
• Dynamic pricing — Fetched from OpenRouter model metadata
• Budget alerts — Set AI_BUDGET_LIMIT to receive Telegram notifications when daily spend exceeds threshold

Tip

Use lighter models (Hermes-4-405B, Gemini Flash) for initial triage and reserve Claude-3.5-Sonnet for complex root-cause analysis to optimize costs.

---

🪟 Windows Users

• Inside Docker: All apt-get commands in Dockerfiles run inside Linux containers — fully compatible via Docker Desktop.
• Line Endings: If you encounter \r: command not found, ensure your IDE saves files with LF line endings.
• Paths: pnpm setup automatically detects Windows paths for GitHub credentials.

---

📖 Documentation

Full documentation is available in the apps/docs package, built with Fumadocs:

• Installation Guide
• Configuration Reference
• Architecture Deep Dive
• Webhook Integrations
• Telegram Bridge
• Cost Management

---

📝 License

Distributed under the MIT License. See LICENSE for more information.

---

Built with ❤️ by the AGIAAS Team