Claude Local Starter
indranildchandra/claude-local-starter
Replica seu setup Claude Code local em qualquer máquina com um script
Ver no GitHubScript de instalação que replica sua configuração completa do Claude Code (plugins, skills, MCP servers, hooks) de uma máquina para outra em segundos. Versioná sua setup local mantendo estado persistente, controle fino de tokens e sem dependência de ambiente na nuvem.
Instalação
npx convez-ai indranildchandra/claude-local-starter1 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
- Sincronizar configuração do Claude Code entre múltiplos computadores
- Manter backup versionado de plugins, skills e MCP servers
- Automatizar setup inicial em máquinas novas preservando fluxo de trabalho
- Gerenciar custo de tokens com dashboard de inspeção de componentes ativos
Veredito

Solução prática para quem trabalha com Claude Code localmente e precisa replicar setup entre máquinas. O script é bem estruturado e documentado, mas requer familiaridade com shell e versionamento Git para máximo aproveitamento.
README
A tradução para o português deste README ainda está na fila.
claude-local-starter
Lift-and-shift your Claude Code setup onto any machine with a single script.
Why this exists
Claude Code's real power isn't the model — it's what you build on top of it. Your CLAUDE.md shapes how it reasons. Your plugins define what agentic workflows it can run. Your skills control what domain knowledge it carries into every session. Your MCP servers determine what external systems it can reach. Your hooks automate the discipline that would otherwise quietly slip.
All of that lives in ~/.claude/ on one machine. It's not backed up. It's not in git. It silently disappears every time you set up a new laptop or need to work from somewhere else.
I kept hitting that wall. So I version-controlled the whole thing.
One bash install.sh on any machine and you're back to full speed: same behaviour, same token budget discipline, same toolchain, same dashboard to inspect what's running. No cloud environment required — and that part matters. Local dev gives you persistent shell state, your own filesystem, and a Claude that remembers your setup because your setup is actually there.
This isn't a generic starter. It's the specific lineup of plugins, skills, and MCP servers I've found most useful day-to-day, packaged into a single installer. Everything is off by default unless it's earned a place in the active session — most skills are context-off, most plugins are disabled, most MCPs are toggled off. You pay for what you actually use. The dashboard shows you exactly what you're loading and what it costs in tokens before you commit to it.
Think of it as the all-stars setup — not everything available, just the things that have proven genuinely useful across real work.
What it installs
| Component | Details |
|---|---|
~/.claude/CLAUDE.md | Global Claude Code behaviour (from claude-md-master/CLAUDE.md) |
~/.claude/settings.json | Plugins, env vars, MCP servers, hooks (deep-merged) |
~/.claude/skills/ | frontend-design, ui-ux-pro-max, shadcn, web-design-guidelines, humanizer, aidlc-tracking, review-council, ppt-creator, last30days, ddg-search |
~/.claude/commands/ | /init-repo, /design-review, /log-context, /create-ppt |
| LSP binaries | typescript-language-server (enabled), pyright (enabled), gopls, rust-analyzer, jdtls |
| Browser automation | @playwright/cli with skills + Chromium |
| MCP servers | context7 (enabled by default), gitnexus, context-mode, claude-mem, filesystem, supabase, vercel |
| Shell functions | see below |
~/.claude-work/ | Private context directory (load on demand with cw) |
Usage
git clone https://github.com/indranildchandra/claude-local-starter
cd claude-local-starter
bash install.sh # safe update — merges, prompts for preferences
bash install.sh --clean-install # full overwrite from repo (backs up first)
bash install.sh --dry-run # preview all actions, no changes written
On update, the installer asks two questions:
- Preserve existing
~/.claude/CLAUDE.md? (default: n — overwrite from master) - Preserve existing MCP / plugin / skill states? (default: y — keep your toggles)
After install, reload your shell and open the dashboard:
source ~/.zshrc
open ~/.claude/claude-local-starter.html # always open this, not the repo file
Dashboard
A single-page HTML dashboard is injected with your live settings.json and CLAUDE.md at install time. Open it to see exactly what's configured, what's enabled, and what each component costs in tokens.
After running install.sh, your personal copy lives at:
~/.claude/claude-local-starter.html
Open it with:
open ~/.claude/claude-local-starter.html
Always open this installed copy — not the claude-local-starter.html file in the repo itself. The repo file is a template; the one in ~/.claude/ is the one with your actual settings.json and CLAUDE.md baked in.
Key files
| File | Purpose |
|---|---|
install.sh | The installer — idempotent, safe to re-run |
scripts/token-audit.sh | Measure token footprint of skills and plugins before enabling them |
scripts/enable-safe-yolo.sh | Add auto-approve permissions to a repo's .claude/settings.json |
scripts/disable-safe-yolo.sh | Remove auto-approve permissions from a repo's .claude/settings.json |
scripts/config/claude-safe-yolo-permissions.txt | Single source of truth for safe-yolo permissions |
claude-md-master/CLAUDE.md | Source of truth for ~/.claude/CLAUDE.md — always overwrites on install |
settings.json | Source of truth for ~/.claude/settings.json |
commands/ | Custom slash commands synced to ~/.claude/commands/ |
skills/ | Custom skills synced to ~/.claude/skills/ |
requirements.txt | Pinned Python deps for bundled skills — edit here, install.sh reads from it |
claude-local-starter.html | Visual dashboard — open the copy at ~/.claude/, not this file |
Shell functions (available after install)
list-skills # show all skills with context-on/off state
enable-skill <name> # let Claude auto-load skill at session start
disable-skill <name> # user-invocable only, zero idle token cost
list-plugins # show all plugins with enabled/disabled state
enable-plugin <name> # enable a plugin (restart Claude Code to apply)
disable-plugin <name> # disable a plugin
list-mcps # show all MCP servers with enabled/disabled state
enable-mcp <name> # enable an MCP server (restart Claude Code to apply)
disable-mcp <name> # disable an MCP server
list-commands # list all slash commands
cw # launch Claude with ~/.claude-work context loaded
# (exit and run plain 'claude' to go without it)
MCP servers
All disabled by default except context7. Enable on demand with enable-mcp <name>.
| Server | Purpose | Source |
|---|---|---|
context7 | Live documentation fetching (enabled by default) | upstash/context7 |
gitnexus | Codebase knowledge graph — run npx gitnexus analyze per repo | abhigyanpatwari/GitNexus |
context-mode | Compresses MCP tool outputs — extends long sessions | mksglu/context-mode |
claude-mem | Session memory (requires claude-mem plugin) | thedotmack/claude-mem |
filesystem | Local file access (${HOME}) | modelcontextprotocol/servers |
supabase | Database queries | supabase-community/supabase-mcp |
vercel | Deploy and logs | vercel-mcp |
Plugins
Installed inside an active Claude Code session via /plugin install. See ~/.claude/plugin_commands.sh after install.
| Plugin | Source |
|---|---|
claude-code-setup | anthropics/claude-plugins-official |
claude-md-management | anthropics/claude-plugins-official |
ralph-loop | anthropics/claude-code |
claude-mem | thedotmack/claude-mem |
superpowers | obra/superpowers |
obsidian-skills | kepano/obsidian-skills |
pr-review-toolkit | anthropics/claude-code |
code-review | anthropics/claude-code |
feature-dev | anthropics/claude-code |
context-mode | mksglu/context-mode |
hookify | anthropics/claude-code |
frontend-design | anthropics/claude-plugins-official |
ui-ux-pro-max | nextlevelbuilder/ui-ux-pro-max-skill |
typescript-lsp | anthropics/claude-plugins-official |
pyright-lsp | anthropics/claude-plugins-official |
gopls-lsp | anthropics/claude-plugins-official |
rust-analyzer-lsp | anthropics/claude-plugins-official |
jdtls-lsp | anthropics/claude-plugins-official |
Skills
Synced to ~/.claude/skills/ on install.
| Skill | Source |
|---|---|
aidlc-tracking | bundled (this repo) |
review-council | bundled (this repo) |
ppt-creator | bundled (this repo) |
ddg-search | bundled (this repo) — DuckDuckGo search fallback, zero API keys |
humanizer | blader/humanizer |
last30days | mvanhorn/last30days-skill — free tier: Reddit, HN, YouTube, GitHub, Polymarket |
playwright-cli | microsoft/playwright-cli |
ui-ux-pro-max | nextlevelbuilder/ui-ux-pro-max-skill |
shadcn | shadcn-ui/ui |
frontend-design | anthropics/skills |
web-design-guidelines | vercel-labs/agent-skills |
docx / pdf / pptx / xlsx | bundled |
Measuring token costs
Before enabling a new skill or plugin permanently, measure its token footprint:
bash scripts/token-audit.sh --save # snapshot current state, enable everything
# Open Claude Code → /reload-plugins → /context (note token counts)
bash scripts/token-audit.sh --restore # restore your original state
# Open Claude Code → /reload-plugins
Plugin installation (manual)
Plugin commands only work inside an active Claude Code session. After install:
cat ~/.claude/plugin_commands.sh
# Open Claude Code and paste the commands
Adding custom skills
Place skill directories under skills/ — each needs a SKILL.md. They sync to ~/.claude/skills/ on install.
Sync behaviour: cp -rn vs cp -rf
The installer uses two copy strategies depending on whether a skill is user-customisable or always kept in sync with the repo:
| Flag | Behaviour | When used |
|---|---|---|
cp -rn (no-overwrite) | Copies files that don't exist at the destination; leaves existing files untouched | All bundled skills by default — preserves local tweaks a user may have made to their installed SKILL.md or supporting files |
cp -rf (force-overwrite) | Overwrites the destination unconditionally on every install run | Skills whose script logic is actively maintained in this repo and where stale code is a real risk (e.g. ddg-search) |
Rule of thumb: use cp -rn for skills users might customise; use cp -rf for bundled scripts where this repo is the source of truth and every install should deliver the latest version.
To opt a new bundled skill into force-updates, add a targeted cp -rf block after the general sync in install.sh (see the ddg-search block as the reference pattern).
Adding custom commands
Place .md files under commands/ — they sync to ~/.claude/commands/ and become available as /command-name inside Claude Code.
Custom-built components
A few things in this repo aren't pulled from anywhere — they're written specifically for this setup and the workflows behind it.
Skills
| Skill | What it does |
|---|---|
aidlc-tracking | Canonical formats for all project tracking files — plan.md, todo.md, tracker.md, lessons.md, changelog.md, design-review.md. Exists so Claude never invents its own structure for these files and every project looks the same. |
review-council | Spins up a multi-persona review council for architecture and design decisions. 20 expert personas, parallel subagent analysis, structured debate, converges on a verdict with your input. Heavy but useful for decisions that actually matter. |
ppt-creator | Full presentation pipeline from brief to .pptx. Research → narrative → design system enforcement → Visual QA before output. Produces conference-ready decks in a consistent personal design system. Triggered via /create-ppt. |
ddg-search | DuckDuckGo web search with a single Python dependency (ddgs, pinned in requirements.txt). Exposes a Python script Claude can call via Bash when native WebSearch/WebFetch isn't available — useful for locally-hosted models or restricted environments. Pairs well with last30days as a supplemental search layer. |
Slash commands
| Command | What it does |
|---|---|
/init-repo | Bootstraps a project from scratch — runs gitnexus analysis, writes a CLAUDE.md tailored to the codebase, and scaffolds all AIDLC tracking files in one shot. Run once per new repo. |
/design-review | Triggers the review-council skill for a full architectural review. Pass a scope (file, module, or decision) or leave blank to review the whole repo. |
/log-context | Writes a detailed session snapshot to tasks/tracker.md before compaction. Preserves enough context that a cold-start in the next session doesn't lose the thread. |
/create-ppt | Triggers the ppt-creator skill. Walks through RESEARCH → ALIGN → STRUCTURE → GENERATE → REVIEW with checkpoints at each stage. Produces a named .pptx in the personal design system with full Visual QA before delivery. |
Contributing
PRs are welcome if you've found something that makes Claude Code meaningfully better.
A few ground rules:
- Don't add things for completeness. Every plugin, skill, and MCP in here is here because it earned its place through actual use. If you're adding something, say why it's better than what's already here or what gap it fills.
- Token cost matters. Anything that loads into context by default needs to justify that cost. Check
scripts/token-audit.shbefore proposing anything context-on by default. - Keep the defaults conservative. New additions should ship disabled. Let users opt in.
- Test the installer. Run
bash install.sh --dry-runfirst, thenbash install.sh --clean-installon a genuinely fresh setup before submitting. A dry run catching no errors is not the same as a clean install working end-to-end.
For custom skills and commands: follow the existing structure. Skills need a SKILL.md, commands need a .md file. Both should have a clear, narrow scope — do one thing well.