Koolie

Governance for AI coding assistants – one rule source for every client, enforced technically wherever the client allows it.

Deutsch: README.md. The German documentation is authoritative; further documents are in German.

A team working with AI coding assistants needs more than an instruction file: shared rules for every assistant, technical blocks wherever a rule must not be broken, and clear points where a human decides. Koolie brings exactly that into an existing Git repository – and states openly, for each client, which rule it really enforces and which only acts as an instruction.

Quick start

In your project directory:

uvx koolie                  # or: pipx run koolie · npx @renoxar/koolie

The command fetches Koolie, offers the current directory as the project and asks for your AI client. Afterwards the project holds the instruction file, the permissions and a guard hook for that client. Without a package registry, use the release archive and the starters install.cmd / install.command.

Want to try it first? The quickstart installs Koolie into an empty practice repository – in ten minutes, without starting an AI client.

How it works

  Core  .koolie/core/            Project overlay  .koolie/project-overlay/
  rules, skills, checklists,     paths, approval routes, project documents
  guard hook, validator          – owned by the project
            │                               │
            └───────────────┬───────────────┘
                            ▼
                Client pack per AI client
       instruction file · permissions · hooks · skills
       (e.g. CLAUDE.md and .claude/ for Claude Code)

• One core for all clients. Rules and skills are written once, tool-neutral; a team with mixed tools shares the same rules.
• The project stays in charge. Project values live in the overlay. A new release updates a project with the same command: it detects the existing core and offers to upgrade it; the overlay is never touched.
• Control levels instead of trust. The assistant handles small tasks on its own; from "medium" it needs a confirmed plan, at "high" an explicit approval.

What is enforced – and what is not

Each client pack keeps a capability matrix that assigns every promise to one of these kinds – backed by a measurement on the client. Koolie complements a good standard setup (branch protection, CI, managed settings); it does not replace it.

Supported clients

Pilot means: fully built, measured on real installations, meant for accompanied use. The limits of each client are listed in its pack (German).

Who is it for?

• Development teams that use AI assistants every day and want shared, verifiable rules.
• Technical leads – team leads, architecture, security, data protection – who must show what a tool enforces technically.

Koolie is not meant for running an agent fully autonomously without human approvals.

Documentation

Why "Koolie"?

A Koolie is an Australian herding dog. It does not drive the herd and does not replace the shepherd – it keeps the flock together, on course and within bounds. That is what this framework is meant to do with an AI assistant.

License

GPL-3.0 with an additional permission: files created from the templates and any output of the tools belong to the project that creates them. Merely using Koolie – also commercially – puts no obligations on you; the details are in LICENSE-HINWEIS.md (German). Copyright © 2026 <FRAMEWORK_OWNER>.