SAL — Skill Adaptation Layer
nathanchouck/sal
Converta skills de Claude em MCP servers com dependências inter-skill automáticas
Ver no GitHubMCP server que converte Claude Agent Skills em servidores MCP, agregando múltiplas skills em um único hub com dependências cruzadas resolvidas. Expõe ferramentas, recursos e relacionamentos entre skills como uma camada de abstração unificada.
Instalação
npx convez-ai nathanchouck/sal1 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
- Integrar repositórios de skills em um único servidor MCP
- Mapear e resolver dependências cruzadas entre múltiplas skills
- Executar e rotear skills dinamicamente via requisição em linguagem natural
Veredito

Ferramenta de arquitetura avançada para quem precisa orquestrar múltiplas skills Claude como servidores MCP. Solução robusta com gestão automática de dependências, mas requer compreensão clara da estrutura de skills e bom design de repositórios para máxima efetividade.
README
A tradução para o português deste README ainda está na fila.
SAL — Skill Abstraction Layer
Ingest a repository of interlinked Claude Agent Skills and expose it as a single MCP hub server — every skill's tools, their reference data, and the cross-skill dependencies wired together.
- A bundled
.pyscript becomes a tool that runs the script. - Any other file (and the
SKILL.mdbody) becomes a tool that returns its text, and a first-class resource reachable viasal_read_resource. - A skill's
skills/<other>/…references and`other-skill`handoffs become cross-skill dependencies that resolve against the other skills in the repo. - Prose has no automatic mapping — it's listed in
lossiness.json, never dropped.
For how it all works, see docs/ARCHITECTURE.md.
1. Install
pip install -e ".[dev]" # SAL + pytest + ruff
pip install -e ".[server]" # fastmcp — needed to run emitted hubs
pip install -e ".[router]" # anthropic — optional LLM routing fallback
2. Build a hub from a repo of skills
sal build path/to/skills --out out/hub
Point build at a directory of skill folders (each with a SKILL.md). For a
single skill, point it at one skill directory — it's just a repo of size one.
Authoring issues (e.g. hardcoded absolute paths in scripts — which break a
script tool at runtime) are reported in lossiness.json but don't fail the build,
so a dropped-in library still produces a working hub. Add --strict to fail the
build on them instead (good for CI). Scripts should resolve paths relative to
__file__, never hardcode /Users/… or C:\Users\….
This writes into out/hub/:
| File | What it is |
|---|---|
server.py | the runnable MCP hub (all skills + sal_* API tools) |
manifest.json | every skill, its tools, resources, and dependencies |
lossiness.json | what mapped, unmapped prose, and unresolved cross-skill deps |
3. Inspect the cross-skill graph
sal graph path/to/skills
4. Run the hub
python out/hub/server.py # stdio (Claude Desktop)
Or over HTTP:
fastmcp run out/hub/server.py:mcp --transport http --host 127.0.0.1 --port 3000
5. Add it to Claude Desktop
Settings → Developer → Edit Config, then restart:
{
"mcpServers": {
"sal-hub": {
"command": "python",
"args": ["C:\\path\\to\\out\\hub\\server.py"]
}
}
}
6. Route / execute without a client
sal route "<natural-language request>" path/to/skills
sal run path/to/skills <skill> <tool> [--args "..."]
Develop
pytest # run all tests
ruff check . && ruff format --check . # lint + format
The sal_* API tools (on every hub)
| Tool | Purpose |
|---|---|
sal_read_resource(skill, resource) | read any skill's resource (cross-skill) |
sal_resolve_dependencies(skill) | the skills + resources injected for a skill |
sal_execute(skill, tool, input) | run any tool by name, normalized output |
sal_route(user_request) | deterministically pick a skill + tool |