Claude Local MCP
rjalabanzas/claude-local-mcp
Instale apenas os MCP servers que sua máquina precisa, com segurança
Ver no GitHubBootstrapper para instalar seletivamente MCP servers locais em Macs, permitindo escolher quais servidores cada máquina precisa sem instalar tudo. Gerencia autenticação por-servidor de forma segura, sem armazenar secrets no repositório.
Instalação
npx convez-ai rjalabanzas/claude-local-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
- Configurar múltiplos MCP servers (Teamwork, Outlook, Telegram, WhatsApp, Read.ai) em um único Mac
- Gerenciar autenticação isolada por servidor sem exposição de secrets
- Adicionar novos MCP servers dinamicamente sem editar scripts centrais
Veredito

Solução prática para gerenciamento local de múltiplos MCP servers em macOS com foco em segurança (sem secrets commitados) e flexibilidade (instalação seletiva). Bem estruturada para desenvolvedores que trabalham com vários servidores de integração.
README
A tradução para o português deste README ainda está na fila.
claude-local-mcp
One repo to set up my local Claude/Cowork MCP servers on any Mac. Clone it, run one script, pick the servers that machine needs, do the per-server logins.
Quick start
git clone <this-repo-url> ~/ClaudeOS/Code\ Repos/claude-local-mcp
cd ~/ClaudeOS/Code\ Repos/claude-local-mcp
./bootstrap.sh --list # see what's available
./bootstrap.sh teamwork readai # install just these
./bootstrap.sh # interactive pick-list
./bootstrap.sh --all # everything
You install only what a machine needs — it's not all-or-nothing.
⚠️ Quit Cowork/Claude (Cmd+Q) before running — the app rewrites its config file on quit, so edits made while it's open are lost. The installer refuses to run while it's open.
Servers
| Server | What | Built how | Per-machine auth |
|---|---|---|---|
teamwork | Projects/Desk/Spaces (tw-mcp, patched for due-date/overdue) | from official upstream + our patch, provenance-verified | paste bearer token |
ms365 | Outlook/Calendar/SharePoint/Teams/etc. | config only (npx) | device-code login (keychain) |
readai | Read.ai meetings/transcripts | vendored (own code) | uv run servers/readai/auth.py |
telegram | Telegram | vendored (chigwell + my patch) | phone login on first run |
outlook | Outlook (community server, calendar fix) | vendored + npm build | OAuth (uses ms365 app) |
whatsapp | vendored + go build bridge | QR scan on first run |
How it works
- Each server is a self-contained folder:
servers/<name>/manifest.sh— description, prerequisites, auth notesetup.sh— build/vendor that one serverconfig.json— its Claude config entry (with${REPO}/${BIN:…}/${SECRET:…}placeholders)
bootstrap.shauto-discoversservers/*/, builds the selected ones, theninstall_config.pymerges their entries into Claude's config (backup first, prompts for any secrets via hidden input).
Adding a new MCP later
Drop a new servers/<name>/ folder with manifest.sh, setup.sh, and
config.json. No edits to bootstrap.sh — it's picked up automatically.
Security
- No secrets in the repo.
.gitignoreblocks tokens,.env, session files,*.db, and the Claude config. Secrets are prompted at install time and written only into Claude's config. teamworkbuilds from the official Teamwork GitHub and verifies the clone's commit + the SDK module hash against a known-good reference before building — it won't build on a mismatch. Other servers are vendored from the exact source I run (third-party ones keep their upstreamLICENSE+ attribution).- Go/npm builds run no untrusted lifecycle scripts beyond the projects' own.