MAS
A custom terminal-driven agentic orchestration framework built to execute multi-file repository refactorings with RunPod cloud GPU or local LLM compute offloading.
Autonomous Agent Framework · Heavy Refactoring Engine
01. Introduction
Core Objectives
MAS is a custom multi-agent orchestration tool designed to process multi-file refactoring tasks without vendor lock-in or recurring SaaS API costs. It provides full control over context parsing and offloads heavy compute to on-demand RunPod cloud GPUs or local LLM instances.
02. Tech Stack
Engine & Tools
03. Background
Why This Exists
Commercial AI coding tools struggle with full-repository context graphs and generate high monthly subscription bills. MAS solves this by deploying specialized worker bots that analyze dependency trees, execute code transformations, run local test suites, and commit clean feature branches.
Transparent CLI & Context Dependency Parser
Tasks are submitted via CLI and converted into structured file context graphs. Agents work deterministically on isolated tasks with full execution logging.
RunPod Cloud & Local vLLM Offloading
Lightweight routing models manage task distribution while heavy contextual refactorings are dispatched dynamically to cost-effective RunPod GPU instances or local vLLM nodes.
Agent Task Pipeline
From CLI Ticket Input to Verification & Commit
Agents coordinate through a master terminal orchestrator, delegating file transformations to specialized worker bots.
CLI Task Ingestion
User submits refactoring spec via terminal
Context Dependency Graph
Orchestrator parses impacted repository files
Worker Execution (Docker / vLLM)
Worker bot implements multi-file changes in container
Automated Test & Git Commit
Executes test suite & commits clean git branch
Roadmap & Direction
Meilensteine & Systemfortschritt
Meilenstein 1: Agent Core
Terminal client interface, local repository context parser & autonomous agent tool loop.
Meilenstein 2: GPU Offloading
Dynamic spin-up of cost-effective RunPod GPU workers for heavy contextual refactorings.
Meilenstein 3: Local vLLM Pooling
Local vLLM pooling integration for zero external API costs during extended coding runs.