# Claude Kit > A beginner-friendly Claude Code starter kit: roles & safety, skills, commands, and a clean filesystem/memory layout (agent-atlas, on one machine). - **Type:** MCP server - **Install:** `agentstack add mcp-zzallirog-claude-kit` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [zzallirog](https://agentstack.voostack.com/s/zzallirog) - **Installs:** 0 - **Category:** [AI & ML](https://agentstack.voostack.com/c/ai-and-ml) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [zzallirog](https://github.com/zzallirog) - **Source:** https://github.com/zzallirog/claude-kit ## Install ```sh agentstack add mcp-zzallirog-claude-kit ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About # claude-kit [](https://github.com/zzallirog/claude-kit/actions/workflows/ci.yml) > A ready-to-run **Claude Code starter kit** — roles & safety, skills, commands, and a > clean filesystem/memory layout. What took one person three months, packaged so a > newcomer gets going in five minutes. Built for a specific beginner (a psychology student on Linux); the shape is general — the persona is just the worked example. The whole thesis fits in one picture — where the AI walks freely, and where it has to stop and ask you: ``` ┌─────────────────────────────┬─────────────────────────────────────┐ │ USERSPACE (your home ~/) │ SYSTEM (root) │ │ free movement │ under grant │ ├─────────────────────────────┼─────────────────────────────────────┤ │ read / write files │ updates, package installation │ │ ~/projects, ~/kb, notes │ firewall, close/open ports │ │ diagnostics (look only) │ restart system services │ │ log search, laptop map │ drivers, time, network, bootloader │ │ → AI acts WITHOUT ceremony │ → every time: GRANT + your password│ └─────────────────────────────┴─────────────────────────────────────┘ ``` No gray middle: either it's clearly your home and the AI just works, or it's the system and you're clearly needed. That line is the kit. ## Quick start ```bash git clone https://github.com/zzallirog/claude-kit && cd claude-kit claude ``` Then tell Claude one phrase: > read BOOTSTRAP.md and begin Claude runs the onboarding itself: probes your machine, asks a short intake, offers skills **one at a time** (installs nothing without your "yes"), sets up commands, and lays down a clean layout. Beginner-facing walkthrough for the human: **[START-HERE.md](START-HERE.md)**. ## What you get - **Skills** — `stats-lab` · `data-wrangle` · `research-helper` · `linux-tutor` · `system-tune` · `man`. Copied into `~/.claude/skills/` only on consent. - **Commands** — `update` (system update + security check in one) · `fix` (self-healing: disk/DNS/Wi-Fi/packages, grows its own known-issues registry) · `where` (map of the laptop) · `recall` (search notes + logs) · `kit-help` (cheatsheet) · `kit-guide` (a browser dossier, Super+\\). - **Safety by construction** — root is never standing: every system action passes a four-beat **grant ritual** (request → your yes → *your* password → auto-revoke). A guard hook (`hooks/deny-destructive.sh`) cuts destructive commands from *any* AI action, not just the kit's. - **A clean layout** — a typed filesystem + memory kept [clean for Claude](conventions/layout.md): a map, not a dump — load only what's needed. The concept is [agent-atlas](https://github.com/zzallirog/agent-atlas) folded down to a single machine. ## The guide `kit-guide` (Super+\\) opens a browser dossier — *Codex of the Machine* — that lays the whole kit out as an illuminated manual: who's who, what's free, what's guarded, every command one click from your clipboard. It's an admin page at heart — but a designed one. *The cast and the two realms · the grant rite and the charters. The same substance as `THE-PROTOCOL.md` and `ROOT-POLICY.md` — built to be read for pleasure, not just reference.* ## The protocol (who's in charge) **You decide · Opus thinks · Sonnet does — and the key is yours.** Home (`~/`) is free; the system (root) moves only through a one-time grant that *you* confirm with your password. The AI doesn't know it and doesn't type it. The grant handshake — how temporary root is taken (4 beats) ``` ┌─ GRANT REQUEST ─ root ──────────────────────────── │ requested by : AI operative │ operation : close unneeded port 5355 │ command : sudo ufw deny 5355 │ risk : 🟡 medium │ authorized ONLY by the owner — without your password the ritual won't pass. └─────────────────────────────────────────────────── Owner, grant access? [y/N]: y ● grant issued. handing to the system — enter the password (this is the confirmation): ✓ CONFIRMED — done. temporary root revoked. all clear. ``` 1. **REQUEST** — the AI shows the exact command, the reason, the risk. Nothing hidden. 2. **GRANT** — you say yes/no. No → not performed, period. 3. **PASSWORD** — the system asks for *your* password. The AI never sees or types it. 4. **CONFIRMED → ALL CLEAR** — the command ran, temporary root immediately revoked. Full model: **[THE-PROTOCOL.md](THE-PROTOCOL.md)**. The honest boundary — what the kit does and doesn't guarantee The ritual is the discipline of the *kit's own commands*: inside `update`/`fix` root goes only that way (the single path is `run_root()` in `bin/kit-lib.sh`). But the kit **cannot by itself forbid** the AI from calling `sudo` around it — that's Claude Code's permissions, not the kit's. So the real guarantee is two layers on top of the ritual: 1. your allowlist has no "allow any `sudo` without asking" (`config/claude-allowlist.json`); 2. the **guard hook** cuts destructive actions from *any* AI command. Full rationale: **[ROOT-POLICY.md](ROOT-POLICY.md)** · **[conventions/security-root.md](conventions/security-root.md)**. ## Map **[AGENTS.md](AGENTS.md)** is the navigator — a route into the kit for both Claude and you, so nobody has to read the whole thing to find one door. Repo layout ``` BOOTSTRAP.md the onboarding script Claude runs START-HERE.md the entry point for the human THE-PROTOCOL.md roles (owner / architect / operative) + temporary root ROOT-POLICY.md the full root policy CONVENTIONS.md working habits (fish · projects · kb · where/recall) conventions/ layout.md (clean-for-Claude) · security-root.md skills/ stats-lab · data-wrangle · research-helper · linux-tutor · system-tune · man bin/ kit-install · update · fix · where · recall · kit-guide · kit-keybind hooks/ deny-destructive.sh (guard hook) kb/ known-issues.sh (self-healing fix) · README.md tweaks/ CATALOG.md (system tweaks) config/ claude-allowlist · fastfetch · layout.example guide/ kit-guide.html (the dossier, Super+\) ``` ## Honest scope - **Linux, one machine.** Built and exercised on a single beginner's Linux laptop; the layout is [agent-atlas](https://github.com/zzallirog/agent-atlas) collapsed to one host. macOS/Windows and multi-machine setups are out of scope for now. - **One worked persona.** Tested against the shape it was built for (a psychology student). The structure generalizes; the intake questions and skill picks are that example. - **Designed for Opus.** It runs on Sonnet — a bit more simply — and Claude says so at start if Opus isn't on. - **CI is a shell lint, not a behavior suite.** Every shell file is shellchecked at error severity on push; the guard hook and the grant path — the load-bearing safety — are reasoned in the docs, not yet exercised by an end-to-end test. ## Lineage The layout is [agent-atlas](https://github.com/zzallirog/agent-atlas) folded down to a single machine; this kit is one working implementation of it. ## Changelog Version history in **[CHANGELOG.md](CHANGELOG.md)**. --- 🤝 **Handshake** — the kit's own rule, turned on itself: *trust the trace, not the claim.* Don't take this README's word that your copy is whole. Clone it and run `git rev-parse HEAD` — it must equal the commit hash GitHub shows at the top of this page. A repo can't print its own commit hash inside itself, so the anchor lives where you can't fake it: the commit, not the copy. ## License MIT. ## Source & license This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [zzallirog](https://github.com/zzallirog) - **Source:** [zzallirog/claude-kit](https://github.com/zzallirog/claude-kit) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/mcp-zzallirog-claude-kit - Seller: https://agentstack.voostack.com/s/zzallirog - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.