Skills MCP — Servidor de Habilidades para Agentes
klutometis/skills-mcp
MCP server para descoberta, criação e execução de skills de agentes
Ver no GitHubServidor MCP agnóstico que expõe Anthropic Agent Skills como recursos descobríveis e executáveis. Permite que qualquer agente liste, crie e execute skills definidas em SKILL.md diretamente no bare metal, com execução via bash e suporte a variáveis de ambiente.
Instalação
npx convez-ai klutometis/skills-mcp1 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
- Orquestração de agentes com skills customizadas e reutilizáveis
- Integração de agentes Claude com ferramentas locais e APIs externas
- Desenvolvimento iterativo de automações onde o agente aprende skills em tempo real
Veredito

Solução técnica e minimalista para quem quer arquitetura agnóstica de skills em agentes. Excelente para casos de orquestração complexa, mas requer controle total do ambiente (sem sandbox) — adequado só para ambientes controlados e single-tenant.
README
A tradução para o português deste README ainda está na fila.
skills-mcp
Agent-agnostic MCP server for Anthropic Agent Skills (SKILL.md) on
bare metal. Bring your own agent; this brings the skills + execution.
It solves the three problems of a self-authoring skills layer, with no bundled agent and no identity of its own:
| Problem | How |
|---|---|
| Discovery | FastMCP's built-in SkillsDirectoryProvider exposes every skills/<name>/SKILL.md as an MCP resource (skill://<name>/SKILL.md). reload=True makes new skills appear instantly. Plus list_skills / read_skill tools for tools-only clients. |
| Creation | create_skill(name, skill_md) writes a new SKILL.md — live on the next call. This is the "learn a skill live" beat. |
| Execution | bash(command) runs arbitrary shell on the box (apt, curl, python...). Skills are instructions the agent runs; the agent reads a SKILL.md, then executes it via bash. |
The box's environment inherits into bash, so secrets (e.g. LIFX_TOKEN)
set on the box are available to the commands skills run.
No sandboxing. This is a bare-metal execution surface by design — run it as a single-user, single-tenant box you control.
Tools
list_skills() -> [{name, description}]read_skill(name) -> strcreate_skill(name, skill_md) -> uriwrite_skill_file(name, relpath, content) -> pathbash(command, timeout=120) -> str
Plus MCP resources via the Skills Provider: skill://<name>/SKILL.md,
skill://<name>/_manifest, and supporting files.
Run
uv sync
# HTTP (default; for Railway / remote agents)
PORT=8000 uv run skills-mcp
# or stdio (local agent)
MCP_TRANSPORT=stdio uv run skills-mcp
Env:
SKILLS_DIR— where skills live (default./skills).MCP_TRANSPORT—http(default) orstdio.PORT/HOST— HTTP bind (default0.0.0.0:8000).- any secrets your skills need (e.g.
LIFX_TOKEN).
The "learn a skill live" loop
- Agent:
list_skills()→ gap (nolifx-control). - Agent:
create_skill("lifx-control", "<SKILL.md with the curl>"). - Agent:
bash('curl -X PUT https://api.lifx.com/v1/lights/all/state -H "Authorization: Bearer $LIFX_TOKEN" -d power=on'). - Light turns on. The orchestrator agent is the only brain — no second agent, no injected identity.
Skill format
A skill is a directory with a SKILL.md (Anthropic Agent Skills open
standard). Optional YAML frontmatter description:; optional supporting
files. See skills/hello/.