Self-hosted multi-agent autonomous coding pipeline. Five specialised LLM agents (Orchestrator, Coder, Reviewer, QA, PM) collaborate to produce tested, reviewed pull requests, all controlled from Telegram.
Most AI coding tools run as single-agent assistants. I wanted to see what happens when you split the work across specialised agents with deterministic orchestration and a human approval gate, on a single VPS, with no cloud dependencies beyond LLM API calls.
Architected a 5-agent pipeline on Hetzner CPX32, orchestrated through Lobster YAML so prompt injection cannot derail the workflow. Built ClawRouter for USDC micropayment-based model routing across multiple LLM providers, plus a Playwright visual QA module for automated accessibility and computed-style auditing.
385+ passing tests on the primary testbed project (KPI-Canvas). Dozens of end-to-end feature builds shipped through the pipeline. I use it as my development acceleration infrastructure.