alinaqi/maggy

What started as an opinionated Claude Code setup kit is now an autonomous AI engineering command center

What it solves

Maggy and Claude Bootstrap provide a structured engineering system for users of Claude Code, solving common AI-assisted development issues such as high API costs, context window saturation, and a lack of enforced quality standards. It prevents the AI from using expensive models for trivial tasks and ensures that code is not considered "done" until it passes tests and architectural reviews.

How it works

The project is split into two components: Claude Bootstrap, a configuration pack (skills, hooks, and rules) installed directly into the Claude Code environment, and Maggy, a local FastAPI server with a web dashboard.

Key mechanisms include:

  • 13-Tier Routing: A local classifier (Qwen3) scores tasks by complexity and routes them to the most cost-effective model capable of handling the request (ranging from local models to Claude Opus).
  • TDD Enforcement: Stop hooks prevent the AI from completing a task until tests pass.
  • Memory Systems: Mnemos and Engram provide task-scoped and cross-session memory to prevent state loss during context compaction.
  • Polyphony: Uses Docker/OrbStack to isolate parallel agent sessions in separate containers to avoid file conflicts.
  • Cortex MCP: A code intelligence server that uses a SQLite graph store for AST extraction and bidirectional traversal.
  • Telos: A testing framework that measures "Intent Fidelity" by combining test conformance, architectural drift, and symbol integrity.

Who it’s for

  • Solo engineers who want TDD enforcement and persistent memory within their Claude Code workflow.
  • Teams needing a centralized dashboard to route work across multiple LLM providers (Claude, DeepSeek, Gemini, etc.) while managing costs.
  • Platform engineers looking for a reference implementation of intent tracking and plugin-based AI developer tooling.

Highlights

  • 73 built-in skills covering various languages and frameworks (React, Python, Flutter, etc.).
  • Visual validation for web projects using Playwright to record proof-of-work videos.
  • Context Shunt: Trims large file reads by using a cheap worker model to summarize content before it enters the main context.
  • Council PR Review: A multi-model adversarial review process for GitHub PRs to reduce false positives.
  • Security Audit: A structured, multi-phase adversarial audit process integrated into the harness.

Related

  • Dispatch
    Scaling AI Coding: How Claude Code Navigates Large Codebases

    An exploration of agentic search and the 'harness' architecture used by Claude Code to manage multi-million line monorepos and legacy systems without relying on traditional RAG indexing.

  • Project
    Chachamaru127/claude-code-harness

    Claude Code Harness is a Go‑native plugin that orchestrates Claude‑based (or compatible) AI agents through a disciplined loop—plan, work, review, sync, release—while enforcing strong guardrails and human approvals at each step. It supports Claude Code, Codex CLI, Cursor, and Grok, generating specs, running AI‑driven implementations, independently reviewing changes, and packaging verified releases.

  • Project
    revfactory/harness

    A team-architecture factory for Claude Code that automatically transforms domain descriptions into structured multi-agent teams and skills.

  • Project
    shareAI-lab/learn-claude-code

    A 17‑lesson tutorial that teaches you how to build the surrounding “harness” code for an LLM‑based coding agent (Claude). It walks through adding tools, permissions, context management, memory, task scheduling, multi‑agent coordination, and goal‑driven termination, all around a simple agent‑loop that lets the model decide when to call tools.

  • Project
    Arch1eSUN/Arcgentic

    A harness engineering layer for Codex and Claude Code that transforms ad-hoc AI prompting into a gated, role-based engineering workflow with audits and stop states.