I Upgraded Open PAIAgent: Modular AI Agents Configured via Plain Markdown

Let’s be honest: the modern AI agent ecosystem has a serious bloat problem.

Half the frameworks out there require a 50-node DAG, 400 MB of Python dependencies, and three levels of LangChain abstractions just to fetch a web page and save a markdown note.

I wanted something radically different: sub-second boot times, ~50-100 MB RAM usage, total data privacy, and a modular architecture configured entirely through plain Markdown files.

That’s why I built Open PAIAgent — and the latest update brings multiple ready-to-use agents and an intelligent dynamic skills system.

⚡ What makes Open PAIAgent different?

🚫 1. No LangChain, No LlamaIndex — Pure TypeScript

Instead of wrestling with black-box agent loops, Open PAIAgent's cognitive loop (think-act-evaluate) is written in clean, auditable TypeScript:

⏱️ Sub-second cold starts
📉 ~50-100 MB RAM footprint (runs comfortably on the cheapest VPS or your local machine)
🔍 Zero magic: You can read, debug, and understand every prompt and tool call.

🧠 2. Modular Prompts & Multi-Agent Architecture

Your agent’s persona, identity, and behavior aren't hardcoded in monolithic configs. They are dynamically compiled from simple Markdown files:

Agent.md — Role and core instructions
Identity.md — Persona and tone of voice
User.md — Information about you and your preferences
Memory.md — Working context

Want to switch personalities or roles? Just create a new agent folder under agents/<agentId>/.

⚡ 3. Dynamic Skills System (Context Token Saver)

Instead of dumping 50 tool descriptions and instructions into the LLM context window on every request, skills live in agents/<agentId>/skills/*.md.

Each skill defines keyword triggers in its header:

# Keywords: wildberries, ozon, amazon, temu, shopping

When analyzing products on e-commerce platforms:
1. Extract price, rating, and review count.
2. Structure comparisons in a markdown table.

When your prompt mentions e-commerce, the skill is dynamically injected into the system prompt. Otherwise, it uses zero tokens.

Out of the box, it comes loaded with:

🛒 E-commerce & Search: Price sorting & parsing rules for Amazon, Wildberries, Ozon, Temu, List.am.
💻 Software Engineering: Strict style rules, refactoring, and debugging workflows.
🎨 Content Writing & Design: Copywriting templates, UI/UX guidelines, prompt optimizers.

🛠️ Built-in Superpowers & Tools

Open PAIAgent isn't just a chatbot; it's a full-fledged local operator:

💬 Dual Interfaces: Cyberpunk-themed Web UI + full remote access via a Telegram Bot.
👁️ AI Vision: Attach images in chat with automatic resizing and visual analysis.
📄 Document & Spreadsheet Generator: Generates styled PDFs (via HTML/CSS templates), formatted Excel sheets (.xlsx), and structured Word docs (.docx).
🌐 Smart Web Scraper: Handles both fast static scraping and dynamic SPA rendering (React/Vue).
🎨 Image Generation: Native support for Together AI and X.AI (Grok) with smart fallback logic.
💾 Semantic Memory: SQLite + sqlite-vec for fast local vector search.
⏰ Task Scheduler: Background scheduler (checks every 60s) with real-time UI tracking and instant Telegram notifications.

🚀 Get Started in 2 Minutes

# Clone the repository
git clone https://github.com/nordevelopment/OpenPAIAgent.git
cd OpenPAIAgent

# Install dependencies & run
npm install
npm run dev

Create your .env, define your agents in agents/, and your personal AI assistant is ready to roll.

🤝 Open Source & Data Sovereignty

Your data stays on your machine. No telemetry, no vendor lock-in, and full control over your models and workflows.

⭐ GitHub Repo: https://github.com/nordevelopment/OpenPAIAgent
🌐 Project Page: https://nordevelopment.github.io/OpenPAIAgent

If you like the "keep it lightweight and transparent" philosophy in AI, check out the repo, star it ⭐, and let me know what skills or agents you'd love to see next!

DEV Community