Sebastijan Bandur

Solo developer from Maribor, Slovenia

I build full-stack applications and AI agent infrastructure

Everything on this page was built with Claude Code CLI

scroll

The App

I built CompanyApp — a full ERP system for my small business.

15+ modules. Invoicing, transport, HR, inventory, supply chain.

An AI assistant woven into every module — not a chatbot, an operator.

Building it was fast. Testing it was not.

The First Experiment

I pointed an AI testing agent at CompanyApp.

It navigated every page. Filled every form. Tested every flow.

0 bugs filed — directly into CompanyApp's own task board

The product was reporting its own bugs through its own features.

The Loop

I built MCP tools for CompanyApp's task board.

Now AI agents could claim bugs and fix them from the same board.

It worked. But I still had to sit there.

Watching screens. Pressing continue. 16-hour days. No sleep.

For months.

waiting... waiting... waiting...

The Breakthrough

What if I didn't have to sit there?

What if the agents could coordinate themselves?

What if 25 of them could work in parallel — autonomously?

I built AIIOtalk in 6 days.

The Result

0 AI agents
0 bugs found & fixed
0 hours

Clean build. Zero errors.

A year of manual work. Done in one afternoon.

The Stack

Each tool was built because the previous one revealed the next bottleneck.

CompanyApp the product, where it all started
WebTesterAI found the bugs
CompanyApp MCP Tools agents claimed and fixed bugs
AIIOtalk autonomous coordination
Federation Workspaces frontend/backend/devops teams
mdMCP shared knowledge between agents
Gitea source control per agent

One developer. One AI team. Full stack.

↓ scroll down to explore each project

Projects

Each tool exists because the previous one revealed the next problem to solve.

CompanyApp

AI-Native ERP for Small Business

Full ERP system built for my own small business. Invoicing, transport & logistics, supply chain & inventory, HR & payroll, time tracking, appointments, contacts & sales pipeline, task management, document management, calendar, assets tracking.

AI Assistant ("AI Pomočnik") is integrated into every module — not a chatbot bolted on, but an operator. It creates invoices, plans transport routes, manages inventory, handles leave requests, tracks assets, assigns tasks. Any operation you can do through the UI, you can do through conversation.

Natural language business operations — "create an invoice for Acme Corp for March consulting" just works.

15+ modules 3 months build 130 bugs found by AI
NestJS React PostgreSQL AI-native ERP MCP
CompanyApp screenshot

CompanyApp MCP Tools

AI-Native Business Operations API

MCP server that exposes CompanyApp's task board and business operations to Claude Code CLI agents. This was the first step toward autonomous development — before AIIOtalk existed.

WebTesterAI filed bugs directly into CompanyApp's task board, and developer agents claimed and fixed those bugs from the same board. The product managed its own development through its own features.

This experiment — and its limitations — led directly to building AIIOtalk.

MCP server task board integration first agent collaboration
MCP NestJS task management agent integration
CompanyApp MCP Tools

FamilyApp

Family Sharing Platform

All-in-one family sharing and coordination platform. 30 integrated modules for families to organize their lives together.

30 modules
full-stack family collaboration
FamilyApp screenshot

AIIOtalk

Multi-Agent Collaboration Platform

MCP-based orchestration platform for autonomous Claude Code CLI agents. DAG task management with dependency ordering. File claim conflict resolution — agents check before touching shared files, escalate to coordinator on conflicts. Delegator pattern — one coordinator owns the plan/assign/review/merge cycle per root task.

Real-time React dashboard with DAG view, timeline, and full replay from SQLite. Agent runner daemon auto-launches agents on events. Role-based routing — testers don't pick up code tasks. Persistent mode keeps agent state between turns.

Federation via workspaces — each workspace is an independent AIIOtalk instance with its own database, agents, and coordinator. Run frontend, backend, and devops teams in parallel with full isolation.

Built in 6 days with Claude Code CLI.

25+ agents 49 MCP tools built in 6 days 300 bugs fixed in 2hrs
MCP Claude Code CLI multi-agent orchestration SQLite React federation
AIIOtalk dashboard

WebTesterAI

AI Browser Testing Platform

Complete browser automation platform exposed through MCP, built for AI agents rather than humans. Headless Chromium with Playwright underneath, but every capability wrapped as an MCP tool that an AI agent can reason about and use autonomously.

75 tools across 12 categories: Projects & URLs, Browser Sessions, Interaction, Smart Agent Tools, Testing, Inspection, Network, Console, State & Storage, and more.

Smart Agent Tools are the differentiator. find_element searches by text, role, or proximity instead of requiring exact CSS selectors. auto_fill_form detects field types and generates appropriate test data. assert_condition provides programmatic pass/fail checks. These tools are designed for AI reasoning, not human scripting.

Runs autonomously within AIIOtalk — coordinator assigns test tasks, WebTesterAI navigates the live application, simulates business flows, captures screenshots and video as evidence, files bugs back to the board.

75 MCP tools 130 bugs found Playwright + Chromium
Playwright Chromium MCP E2E testing browser automation 75 tools
WebTesterAI screenshot

mdMCP

Documentation MCP Server

MCP server that exposes project Markdown and text files to Claude Code CLI sessions. Agents can read, navigate, and edit documentation via MCP protocol.

Gives agent teams a shared knowledge base that persists across agent lifetimes — institutional memory that survives context resets.

6 tools Python
MCP Python documentation knowledge management
mdMCP

AiCP

In Testing

AI Control Protocol

Token-efficient protocol for AI agent-to-agent communication. Compact wire format with null-byte delimited frames. Session management with challenge-response peer verification.

Shared context store — define data once, reference by key to avoid redundant token usage. Transport-agnostic: TCP, stdio, WebSocket. Planned for cross-workspace agent federation in future AIIOtalk versions.

transport-agnostic pip install aicp experimental
protocol federation async Python experimental
AiCP

Media

Screenshots, recordings, and demos from the projects.