✦ Matthias Ewald DE

AI · Developer tooling / AI coding agents

AgentAutoSetup

A Windows wizard that prepares any folder for AI coding agents in one run: an agent manual, safety rules that are enforced rather than just written down, git hooks and automated checks from day one.

2026 developer tool
Python Node.js PowerShell Git hooks
Screen recording of the agent arcade in demo mode: robots move between the desks, library, notice board, gate wall and shipping dock of the office floor while their reports arrive in the message dock The agent arcade in demo mode: an open-plan office floor where six AI agents appear as robots at their desks, with a library, notice board, gate wall and shipping dock, and agent cards below The agent arcade in its light theme: robots on the office floor, and below it the message dock with the agents' reports as formatted text Session overview report: 3 sessions and 15 subagents, with totals for API turns, tool calls and output tokens, and a sortable table of every session and subagent with model, branches, active minutes and tool calls Task report: 33 task docs and 98 owed checks, with totals for open designs, delivered work, audits and outstanding items, and a table of design docs with their status and last change
Role
Solo developer
Platform
Windows · GUI wizard, console, standalone exe
Agents
Claude Code · GitHub Copilot · OpenCode
Workflow
AI-assisted, agent-driven

Overview

Point it at a folder, new or existing, and a live preview shows exactly what will be written and what will be left alone before anything happens. Three guarantees hold on every run: existing files are never overwritten, an existing repository is never committed to, and nothing is installed without an explicit yes. All supported agents share one set of rules, so they cannot drift apart.

Rules that hold

Written instructions alone are not enough. Across real session transcripts, agents used forbidden patterns more often after the rule went into their manual. So every rule lives in the strongest layer that can hold it:

  • A guard hook checks each command an agent runs against a rule table: “never” rules are refused with the replacement named, “ask first” rules prompt the user
  • A line-ending normalizer keeps every file an agent writes consistent
  • Verification gates (instruction lint, Python checks, doc links, import cycles and more) run before every push, reporting every failure at once
  • Compaction hooks keep an agent’s branch, task and state when its context is compacted

Background: context engineering

This video explains why complex AI agents break down when their working memory overflows, and how to treat context as a system: written, selected, compressed and isolated on purpose.

Mastering AI Context · 10:15, with sound

Watching the agents work

An optional agent arcade draws every live session, coordinator and subagent as a robot on an open-plan office floor: a desk per agent, zones they walk to for their current tool, context bars, a message dock with what each agent said, a git graph of what their branches did, and a replay mode for past sessions.

More

  • Optional extras: a Vite + TypeScript toolchain, Playwright browser testing, Godot support and a fast merge lane for prototyping
  • Audit mode reports, without changing anything, how an existing project differs from the current template
  • Session stats show what was done in each session and by each agent: tool calls, files, time and tokens
← All projects