convez.ai
MCP ServerDevOps & InfraestruturaIntermediário
infragate

CAPA - Gerenciador de Capacidades para Agentes de IA

infragate/capa

Package manager unificado para capacidades de agentes de IA via capabilities.yaml

Ver no GitHub
BomPluginScore · de 0 a 100

CAPA é um gerenciador de pacotes unificado que consolida skills, tools, rules, sub-agentes, MCP servers e plugins em um único arquivo `capabilities.yaml`. Ele automatiza a instalação e configuração dessas capacidades em Cursor, Claude Code, Windsurf, GitHub Copilot e 35+ outros agentes de codificação com IA. Elimina a necessidade de espalhar configurações por múltiplos arquivos e garante que toda a equipe tenha a mesma base de conhecimento versionada.

Instalação

$npx convez-ai infragate/capa

1 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

Estrelas
748
Forks
96
Acompanhando
1
Saúde
62%
Último commit
há 1 semana
Versão
v2.4.0
Issues abertas
20
Contribuidores
4

Casos de uso

  • Centralizar configuração de skills e tools para múltiplos agentes de IA
  • Sincronizar rules, MCP servers e plugins entre Cursor, Claude Code e outras IDEs
  • Versionamento e distribuição consistente de capacidades de IA para equipes
  • Reduzir fragmentação de config espalhada por diferentes arquivos e plataformas

Veredito

CAPA resolve um problema real de fragmentação em ambientes multi-agente, oferecendo uma abordagem declarativa e versionada. Ferramenta sólida para equipes que usam múltiplos agentes de IA e precisam de consistência operacional, com UI local e suporte a 35+ plataformas.

README

A tradução para o português deste README ainda está na fila.

CAPA is the package manager for AI coding agents. Declare your skills, tools, rules, sub-agents, MCP servers, and plugins once in capabilities.yaml, run capa install, and CAPA writes them into Cursor, Claude Code, Codex, Windsurf, GitHub Copilot, and 35+ other agents.

Why CAPA?

An AI coding agent needs to know when to act and to have the tools to act with. Wiring that up usually means scattering config across half a dozen files: MCP servers in one place, skill markdown in another, team conventions in CLAUDE.md, Cursor rules in .cursor/rules/, a half-finished onboarding doc somewhere in Notion. Nobody's setup matches anyone else's.

CAPA collapses that into one capabilities.yaml next to your code. You list your skills (markdown that tells the agent how to do something), the tools each skill needs (MCP calls and shell commands), your rules, your sub-agents, and any plugins you want pulled in. capa install does the rest. Cursor gets .cursor/rules/. Claude Code gets .claude/agents/ and CLAUDE.md. Codex gets AGENTS.md. Each provider on your list gets the files it expects, with capa-managed marker blocks for the parts it owns.

One file, version controlled, pinned by capabilities.lock, cached by SHA. The teammate who clones tomorrow gets the same bytes you got today.

Screenshots

CAPA is equipped with local web UI. You can visualize your capabilities.yaml, browse registries, manage credentials, and see exactly what each agent will receive.

The project view shows installed plugins, configured providers, and your full capability inventory. The bar across the top tracks token savings from on-demand tool exposure: the agent sees only the tools it's actively using, and pulls any of the rest in by name when it needs them.

Scrolling down the same page brings up sub-agents, rules, project options, and credentials. Every entry carries an INLINE / GITHUB / REMOTE badge so you can see at a glance where each one came from.

The Registries tab pulls skills and plugins from external catalogs. Need a private one? Run capa registry add owner/repo@my-adapter (or use the Manage registries page). capa fetches the adapter from GitHub, GitLab, or an HTTPS URL, validates it, and it shows up here too.

What it does

  • Skills from inline content, raw URLs, GitHub, GitLab, local paths, or a configured registry.
  • 35+ supported agents: Cursor, Claude Code, Codex, Windsurf, GitHub Copilot, Cline, Continue, Goose, Gemini CLI, Roo Code, Qwen Code, and more.
  • One MCP server per agent. Tools load on demand, so the context window stays small. In our benchmarks that runs 19-40% cheaper than exposing every tool up front, with no drop in quality (150 trials on claude-opus-4-8).
  • Any CLI command can be wrapped as an MCP tool the agent (and capa sh) can call.
  • Rules go to each provider's native location: Cursor .cursor/rules/, Windsurf .windsurf/rules/, Copilot's instructions file, or a managed marker block in AGENTS.md / CLAUDE.md for providers without a rules directory. Glob scoping works.
  • Lifecycle hooks for providers that support them (Claude Code, Cursor, Codex, Gemini CLI). Declare canonical events like beforeShell or afterFileEdit once and capa translates them into each provider's hook config (.claude/settings.json, .cursor/hooks.json, .codex/config.toml, .gemini/settings.json), tagging each one with capa:<id> so user-authored entries are never touched. Providers without hook support emit a warning and skip.
  • Sub-agents get their own filtered MCP endpoint that exposes only the tools the specialist actually needs.
  • Skills and plugins are browsable from capa add and the web UI. Add a registry with capa registry add owner/repo@my-adapter (GitHub/GitLab/HTTPS sources, slug auto-derived, adapter validated before install) and it shows up too.
  • capabilities.lock records resolved commit SHAs. A SHA-keyed content cache makes repeat installs near instant.
  • Credentials are encrypted at rest, edited in the web UI, or read from .env.
  • Blocked-phrase enforcement, tool output sanitisation, and CLI prerequisite checks all run before any install touches your filesystem.
  • Private GitHub and GitLab repos work over OAuth.
  • The bundled capabilities-manager skill teaches the agent how to read and edit its own capabilities.yaml.

Installation

macOS and Linux:

curl -LsSf https://capa.infragate.ai/install.sh | sh

Windows:

powershell -ExecutionPolicy ByPass -c "irm https://capa.infragate.ai/install.ps1 | iex"

Quick start

1. Initialize your project

cd your-project
capa init

This drops a capabilities.yaml next to your code.

2. Add a skill

Pull a skill from any public GitHub repo:

capa add anthropics/skills@frontend-design

3. Edit capabilities.yaml

providers:
  - cursor
  - claude-code

skills:
  - id: frontend-design
    type: github
    def:
      repo: anthropics/skills@frontend-design

  - id: web-researcher
    type: inline
    def:
      description: Search the web for fresh information
      requires:
        - "@brave.search"
      content: |
        ---
        name: web-researcher
        description: Use when you need current information from the web.
        ---
        Use `brave.search` and always cite a link.

rules:
  - id: commit-style
    type: inline
    description: Conventional Commits
    alwaysApply: true
    content: |
      Always use Conventional Commits (feat/fix/chore/docs/refactor).
      Subject ≤ 72 chars, imperative mood, no trailing period.

  - id: typescript-conventions
    type: local
    description: Team TypeScript conventions
    appliesTo:
      - "**/*.ts"
      - "**/*.tsx"
    # Path is resolved relative to this capabilities file.
    path: rules/typescript.md

servers:
  - id: brave
    type: mcp
    def:
      cmd: npx -y @modelcontextprotocol/server-brave-search
      env:
        BRAVE_API_KEY: ${BraveApiKey}

tools:
  - id: search
    type: mcp
    description: Search the web with Brave Search
    def:
      server: "@brave"
      tool: brave_web_search

4. Install

capa install

capa install resolves SHAs and downloads anything that isn't already in the cache. It then writes the per-provider files (.cursor/rules/, .claude/agents/, AGENTS.md, and so on) and registers one MCP endpoint with each agent on your list. The resolved SHAs land in capabilities.lock so the next clone gets the same bytes.

5. Turn any MCP server into a CLI tool

capa sh                                  # list every configured tool
capa sh brave                            # list brave subcommands
capa sh brave search --query "…"         # run a tool directly

Every tool you define is also a CLI command under capa sh. MCP tools live at capa sh <server> <tool>. Shell tools live at the top level (or under whatever group you assigned).

Server

CAPA runs a local HTTP server (default 127.0.0.1:5912) for the web UI, REST API, and MCP endpoints.

When the server binds to a non-loopback address (anything other than 127.0.0.1, localhost, or ::1), all /api/* routes and MCP endpoints require authentication. Loopback-only bindings skip auth entirely.

Set a fixed token with the CAPA_AUTH_TOKEN environment variable. If unset on a non-loopback host, CAPA generates a random token at startup, saves it to ~/.capa/auth.token (mode 0600), and prints it once to stderr.

Clients must send the token via header (query-string tokens are rejected):

  • Authorization: Bearer <token> (preferred)
  • X-Capa-Auth-Token: <token> (fallback)

Documentation

Guides, the full schema reference, and the registry catalog:

https://capa.infragate.ai

License

MIT

Ver o original no GitHub

Bom para

Relacionados