About

maestria is a set of free, open-source plugins that give AI coding agents a shared engineering workflow: choose a route, delegate focused work, and check the result independently.

How the Workflow Fits Together

An orchestrator coordinates seven specialists: adventurer (reconnaissance), architect (approach and trade-offs), builder (atomic implementation), diagnose (root-cause analysis), planner (phased plans), reviewer (independent review), and writer (documentation).

Two patterns guide their work: pipeline composition (select the thinker, worker, and verifier stages needed and pass context and evidence between them) and the maker/checker split (a separate reviewer checks the maker’s result; no agent approves its own output).

Choose a route with When to Use maestria; How It Works covers benefits and limits, Pipeline & Roles sequencing, and Specialist Reference each role.

Install for Your Platform

Native integrations support OpenCode, Kimi Code, Pi, Hermes, Claude Code, Codex CLI, Cursor, Prime Agent, and Oh My Pi. Install one with npx maestria install <platform>; the CLI guide lists platform IDs and prerequisites.

For clients that support Agent Plugins v1, @maestria/agent-plugin provides portable skills: run npx maestria plugin install to stage the package, then activate it through the client’s installer. Runtime permissions and delegation behavior depend on the host.

Why maestria Exists

I kept repeating the same personal defaults across projects: check claims against code and docs, match effort to what is at stake, reuse before inventing, and keep output short and self-contained. maestria turns those defaults into shared guardrails, called Universal Floors, with guidance for delegation and review. Safety and authorization come first, and the host keeps control of tools.

Projects still get local say, within bounds. Two project files carry it. They sit below the shared core, live at the root, and never override safety or host permissions. That boundary keeps tailoring useful without quietly rewriting the rules.

The rest is small taste that compounds: do one atomic task per run, show evidence before claims, and never let an agent approve its own work. Delivery closes through review gates, with one canonical source projected per platform and checked by sync.

Maintainer and License

maestria is open source under the MIT license, maintained by Agustinus Nathaniel (agustinusnathaniel), with contributions welcome through the GitHub repository. To reach the maintainer, use the channels on the Contact page; for how this site handles data, see Privacy.