RunDrill — Cursos Interativos para Agentes IA
rundrill/rundrill
Cursos com drills, progresso e memória de erros dentro do seu agente IA
Ver no GitHubPlataforma de cursos executáveis dentro de agentes Claude que fornece drills programados, rastreamento de progresso e memória de erros. Funciona como marketplace de plugins para Claude Code, Codex e Antigravity, com cursos em programação (Python, etc.) e disciplinas profissionais.
Instalação
npx convez-ai rundrill/rundrill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Treinar modelos de IA em linguagens de programação e soft skills
- Rastrear progresso de aprendizado e erros cometidos durante o uso
- Instalar cursos como plugins em Claude Desktop, Codex e Antigravity
Veredito

Solução robusta para integrar treinamento estruturado em agentes IA, com sistema de cursos modular e suporte multi-plataforma. Ideal para equipes que querem aprendizado contínuo acoplado ao fluxo de trabalho do agente.
README
A tradução para o português deste README ainda está na fila.
RunDrill plugin marketplace
The public install index for RunDrill's programming & professional courses. Each course ships
as one plugin — a <subject>-coach skill backed by an MCP drill engine (curriculum, progress,
mistake memory).
Natural-language courses (English, Kazakh, …) live in the sibling
rundrill/rundrill-lang marketplace.
This repo holds only the catalogs; the course plugins live in their own repos under the
rundrill org and are referenced by Git source. Two catalogs cover
the two marketplace-based hosts (one entry per course in each):
.claude-plugin/marketplace.json— Claude Code / Desktop,githubsource..agents/plugins/marketplace.json— OpenAI Codex,urlsource.
Authoring scaffolding (_template/) lives in the private development monorepo, not here.
Installing (for users)
Claude Code / Claude Desktop — via this marketplace:
/plugin marketplace add rundrill/rundrill # the org/repo of this index
/plugin install rundrill-python@rundrill # <plugin-name>@<marketplace-name>
/reload-plugins # → /python-coach is live
@rundrill is the marketplace name field (set in marketplace.json), not the repo name.
(Language courses install from the other index: rundrill-english@rundrill-lang.)
OpenAI Codex — via this marketplace too:
codex plugin marketplace add rundrill/rundrill # GitHub shorthand; reads .agents/plugins/marketplace.json
Then open the plugin directory, pick the RunDrill marketplace, and install the course.
Google Antigravity — no marketplace; Antigravity scans plugin dirs. Drop a course folder in:
~/.gemini/config/plugins/rundrill-python/(global, all workspaces), or<workspace>/.agents/plugins/rundrill-python/(workspace-scoped).
Registering a published course
A course plugin lives in its own repo (e.g. rundrill/rundrill-python). Add one entry to each
catalog, pinned to a released tag (ref, and optionally a full-commit sha) so users don't pull
unfinished work.
.claude-plugin/marketplace.json (Claude — github source):
{
"name": "rundrill-python",
"source": { "source": "github", "repo": "rundrill/rundrill-python", "ref": "v0.2.11" },
"description": "Python coach — read and review code (incl. AI-generated) with drills, progress, and mistake memory."
}
.agents/plugins/marketplace.json (Codex — url source; always include policy + category):
{
"name": "rundrill-python",
"source": { "source": "url", "url": "https://github.com/rundrill/rundrill-python.git", "ref": "v0.2.11" },
"policy": { "installation": "AVAILABLE", "authentication": "ON_INSTALL" },
"category": "Education"
}
Naming convention
| Layer | Pattern | Example | Notes |
|---|---|---|---|
| Plugin (install unit) | rundrill-<subject> | rundrill-english | unique within the marketplace; the Antigravity plugin dir name too |
| Skill / slash command | <subject>-coach | /english-coach | subject-distinct so the bare command works; Claude falls back to /rundrill-english:english-coach on a clash |
| MCP server | rundrill-<subject> | server key in both .mcp.json and mcp_config.json | per-course name; one shared backend behind per-course endpoints |
Rules: lowercase kebab-case everywhere (not underscores); hyphenate multi-word subjects
(system-design, soft-skills); lowercase acronyms (toefl, ielts, owasp). The SKILL.md
name: field, its folder name, and the slash command must all match.
Planned regional bundles use a distinct word so they never collide with a single course:
rundrill-pack-west, rundrill-pack-asia (thin meta-plugins listing course plugins as deps).
Notes
- A plugin folder must be self-contained — its skills and MCP config can't reference files outside the plugin, because users install one plugin at a time. Shared harness logic lives server-side in the MCP backend, not in the plugin.
- Remote MCP shapes differ: Claude uses
{ "type": "http", "url": … }in.mcp.json; Antigravity uses{ "serverUrl": … }inmcp_config.json. If RunDrill's OAuth server supports dynamic client registration (DCR), Antigravity needs onlyserverUrl— it handles the OAuth flow itself. Otherwise addoauth: { clientId, clientSecret }and registerhttps://antigravity.google/oauth-callbackas a redirect URI.
Authoring a new course and the submodule layout: see the private monorepo's plugins/README.md
and Plugin Packaging & Distribution.