claude-code-up — Claude Code Up

The fastest way to bootstrap a Claude Code project with the right skills, agents, MCPs, and hooks — pre-curated, project-scope only, fully JSON-driven.

npx claude-code-up

That's it. One command. Interactive prompts. Your project is ready to ship — with battle-tested agents, the right MCPs, runtime hooks, a working statusline, and the scaffolding for your stack.

Démo

What it does (in 60 seconds)

Most "Claude Code best practice" docs are 30-page Notion pages and a Twitter thread. claude-code-up turns them into one prompt. It asks 6 questions, then it:

1. Bootstraps your framework — npx create-expo-app@latest / create-next-app / create-vue / 7 more, picked based on what you answered.tt
2. Curates the agents and skills that match — 131 items from wshobson/agents + obra/superpowers + Matt Pocock skills, filtered to ~16 defaults + the conditionals that fit your stack. The other ~50 noisy ones are skipped.
3. Writes .mcp.json with the MCP servers your project actually uses — GitHub, Context7, Postgres, Supabase, Stripe, RevenueCat, Better Auth, Playwright. Credentials prompted (masked), .env written, gitignored. Then each MCP is smoke-tested end-to-end with a real JSON-RPC initialize handshake.
4. Installs the harness — Karpathy's CLAUDE.md, optional Superpowers/Pocock skills, claude-mem (persistent memory across sessions, documented in STACK.md), agent-teams flag (research preview), pretty 256-color statusline, agentshield audit hook, project-scope settings.json.
5. Creates the GitHub repo (or clones an existing one), commits, pushes.
6. Runs claude /init headless to verify everything is wired correctly. Then offers claude doctor for the official diagnostic.

Everything is project-scope — nothing ever touches ~/.claude/.

Quick start

# Just run it
npx claude-code-up

# Or with a target directory
npx claude-code-up my-new-app

# Or force-overwrite a non-empty dir (with confirmation)
npx claude-code-up ../legacy-project --force

You'll go through a funnel of prompts:

1. Project name, target directory — sensible defaults (sibling dir, not a subdirectory).
2. Git? GitHub? — gh is detected and reused. claude-code-up checks if the repo exists on GitHub; if not, creates it empty so the remote is ready. If it exists, offers to clone it as the starting point.
3. Funnel (the entonnoir) — Category (Web / Mobile / Backend / Fullstack / Data / CLI / Generic) → Language (TS / Py / Go / Rust / Dart / Swift / …) → Framework (Next.js / Expo / FastAPI / Django / Rails / 30+ more) → Database (Postgres / Supabase / Mongo / Vector / …) → ORM (Prisma / Drizzle / SQLAlchemy / …).
4. Scaffolder — if the framework has one (Expo, Next.js, Vite, Vue, SvelteKit, Remix, Astro, NestJS, FastAPI), claude-code-up offers to run npx create-X@latest . for you. Output is streamed live.
5. Stack components — tri-state multi-select: ● fully on, ◐ partial (orange — press D to drill into specifics), ○ off. The defaults are computed from a JSON-driven decision tree based on your funnel answers.
6. MCPs — preselected by applyWhen rules. Credentials prompted via masked input, written to .env, validated by a real MCP initialize + tools/call handshake.

Then a clear plan recap (with ◐ orange for partial selections and ⚠ warnings for missing env vars), confirm, and claude-code-up runs:

✓ Scaffold complete                        npx create-expo-app@latest .
✓ Claude Code 2.1.142 detected
✓ .claude/ skeleton created
✓ Karpathy CLAUDE.md fetched
✓ CLAUDE.md written
✓ .gitignore preserved (scaffolder generated)
✓ Stack applied — installed: 2 files, 14 agents, 7 skills, 4 scripts, 1 hooks
✓ MCP initialize handshake: 5 ok · 0 failed · 2 skipped
✓ git repository initialized
✓ Pushed initial commit to steph-frtech/my-app
✓ /init completed
✓ claude doctor: clean

Why claude-code-up

| Without claude-code-up | With claude-code-up | |---|---| | Manually curate which of wshobson's 100 agents you actually need | Decision tree filters them by your funnel choices | | Cross-reference Superpowers + Pocock for duplicates (tdd exists in both) | The 49 conflicting items are pre-classified as skip | | Discover skillListingBudgetFraction after seeing "192 skill descriptions dropped" | Auto-set to 0.05 when wshobson is installed | | Set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in 3 locations | One checkbox, written to .claude/settings.json | | Read the MCP docs to figure out which env vars Stripe needs | Masked prompt + .env write + handshake verification | | Forget to test the MCP servers before launching claude | Each MCP is smoke-tested with initialize + tools/call | | Edit settings.json by hand for the statusLine command | Bundled colorful statusline with tokens/limits/diff/cost |

What gets installed

| Source | What | When | |---|---|---| | wshobson/agents (100 agents) | Code-reviewer, security-auditor, debugger, test-automator (defaults). Plus conditionals like typescript-pro (TS), fastapi-pro (FastAPI), mobile-developer (Mobile), database-architect (Backend+DB), mobile-security-coder (Mobile) — picked by your funnel | 4 defaults + 8–15 conditionals typical | | obra/superpowers (14 skills) | TDD, brainstorming, writing-plans, using-git-worktrees, code-review, subagent-driven-development, systematic-debugging (defaults) | 7 defaults | | Matt Pocock skills (21 skills, via mp-skills mirror) | caveman (-75% tokens), qa, zoom-out, improve-codebase-architecture (defaults) | 4 defaults | | MCPs (8 servers) | GitHub, Context7, Playwright, Postgres, Supabase, Stripe, RevenueCat, Better Auth — preselected based on funnel | 2–7 typical | | Bundled skills | octo-issue-tracker, octo-scenario-tester (scenarios-before-code TDD discipline) | Opt-in | | Bundled subagents | code-reviewer, security-auditor, test-writer (short, focused) | Opt-in default | | claude-mem | Persistent memory across Claude Code sessions (compresses transcript context and re-injects on next launch). User-scope: claude-code-up documents npx claude-mem install in STACK.md rather than auto-running it (user-scope install). | Opt-in default | | Project files | CLAUDE.md (Karpathy), .claude/statusline.sh (executable, 256-color), .claude/scripts/audit.sh (agentshield), .claude/settings.json, .claude/stack.json (inventory), STACK.md (human-readable report) | Auto |

All metadata lives in catalog/*.json. Adding a new tool = editing one JSON file, no TypeScript needed.

The decision tree (JSON-driven)

Every catalog item carries a tier and an optional applyWhen rule:

{
  "id": "mobile-security-coder",
  "kind": "agent",
  "container": "frontend-mobile-security",
  "tier": "conditional",
  "description": "XSS, CSP, secure storage for mobile",
  "applyWhen": { "categories": ["mobile"] }
}

Semantics:

• tier: "default" → always installed, no applyWhen.
• tier: "skip" → never offered, has a reason.
• tier: "conditional" with rules → installed when the funnel matches.
• tier: "conditional" with applyWhen: {} → opt-in only.

Inside applyWhen:

• Multiple sections (categories, languages, frameworks, databases, orms) are AND'd.
• Multiple values in a section are OR'd.
• Omitted section = wildcard.

So { "categories": ["backend"], "databases": ["postgres", "supabase"] } means "Backend AND (Postgres OR Supabase)".

Each item's activation rule is shown live in the prompt as a label:

◆  Pick wshobson items (default + matching conditional pre-checked):
│
│ ❯ ◼ code-reviewer [General] — security/prod review
│   ◼ security-auditor [General] — OWASP, vulnerabilities
│   ◼ debugger [General] — Error resolution
│   ◼ test-automator [General] — unit + integration + e2e
│   ◼ python-pro [PY] — Python expert
│   ◼ fastapi-pro [fastapi] — Async Python backend
│   ◼ database-architect [Backend + Fullstack + Data] — Schema from scratch
│   ◼ backend-security-coder [Backend + Fullstack] — API sec
│   ◯ rust-pro [RUST] — CLI perf
│   ◯ flutter-expert [flutter] — Flutter / Dart
│   ...
└  space toggle · d details · enter confirm · esc cancel

The [General] / [PY] / [Backend + Fullstack + Data] chips tell you exactly why each item is pre-checked.

Architecture

claude-code-up/
├── catalog/                          ← JSON, the single source of truth
│   ├── project-types.json            ← funnel options (categories/langs/frameworks/dbs/orms)
│   ├── stack.json                    ← stack components
│   ├── mcps.json                     ← MCP servers + applyWhen rules
│   ├── subagents.json                ← bundled subagent content (.md inline)
│   ├── skills.json                   ← bundled skills (octo-*)
│   ├── scaffolders.json              ← framework→scaffolder mapping (Expo, Next, Vite, …)
│   └── items/
│       ├── wshobson.json             ← 100+ agents, each tiered + conditioned
│       ├── superpowers.json          ← 14 skills
│       └── pocock.json               ← 21 skills
├── src/
│   ├── cli.ts                        ← entry point
│   ├── commands/init.ts              ← orchestrator
│   ├── prompts/                      ← @clack/prompts + custom tri-state prompt
│   │   ├── tristate-multiselect.ts   ← ●/◐/○ with D-key drill-down (built on @clack/core)
│   │   ├── funnel.ts                 ← cascading category→lang→framework→db→orm
│   │   ├── scaffolder.ts             ← propose npx create-X@latest
│   │   ├── stack.ts                  ← uses decision tree resolver
│   │   └── mcp.ts                    ← applyWhen-driven preselection
│   ├── stack/
│   │   ├── decision-tree.ts          ← matchesApplyWhen + resolveForSource
│   │   ├── wshobson.ts               ← clone + file-level agent extraction
│   │   ├── superpowers.ts            ← clone + filtered skill flatten
│   │   ├── pocock.ts                 ← clone + filtered skill flatten
│   │   └── karpathy.ts               ← fetch CLAUDE.md with bundled fallback
│   ├── generators/
│   │   ├── stack.ts                  ← orchestrates per-component install + report
│   │   ├── mcp.ts                    ← writes .mcp.json + .env + MCP handshake test
│   │   ├── github.ts                 ← create-new / clone-existing flow
│   │   └── ...
│   ├── lib/gh.ts                     ← gh CLI wrapper (repo exists, create, clone, login)
│   └── catalog.gen.ts                ← AUTO-GENERATED from catalog/*.json by tsup pre-build
└── scripts/gen-templates.mjs         ← JSON → typed TS gen + schema validation

Mechanism in TS, policy in JSON. The TS code is generic — match logic, install dispatch, prompt rendering. Adding tools means JSON edits + npm run gen.

MCP handshake (real e2e verification)

When you add an MCP, claude-code-up doesn't just write .mcp.json and hope. It actually:

1. Spawns the MCP server with the env vars you provided.
2. Sends a JSON-RPC initialize request.
3. Waits for the result envelope.
4. Sends notifications/initialized.
5. If the catalog has a test field for that MCP, runs tools/call <test.tool> with test.args, validates the response (expectField / expectPattern).
6. Kills the subprocess, reports ok / failed: <stage>: <reason> / skipped: missing <env vars>.

So instead of seeing the MCP fail silently when claude launches, you see it during claude-code-up:

MCP initialize handshake: 3 ok · 1 failed · 1 skipped
    ✓ github
    ✓ context7
    ✓ stripe
    ✗ postgres — tools/call query: connect ECONNREFUSED 127.0.0.1:5432
    ○ supabase — Missing env vars: SUPABASE_SERVICE_ROLE_KEY

Requirements

• Node ≥ 18
• Git
• gh CLI (for GitHub flow — optional but recommended)
• Claude Code ≥ 2.1.32 (auto-installed/upgraded by claude-code-up if missing — stable channel by default)

Roadmap

• [x] ~~Publish to npm so npx claude-code-up works without a GitHub clone~~ — shipped in 0.1.1
• [ ] Live D-key drill within the prompt (currently submits-and-redraws — works, but a single-frame interaction would be slicker)
• [ ] User-local overrides (~/.claude-code-up/items-override.json) to promote/demote items per personal preference
• [ ] More scaffolders: T3, Astro, Tauri, Solid Start, Remix Vite
• [ ] More MCPs: Slack, Notion, Linear, Sentry — with funnel-driven preselection
• [ ] Translate the bundled octo-scenario-tester skill content to English
• [ ] claude-code-up add command — apply claude-code-up to an existing project mid-way
• [ ] claude-code-up doctor — diagnose an existing claude-code-up-bootstrapped project

Contributing

95% of contributions = a single JSON edit. No TypeScript required.

| You want to add | Edit | Time | |---|---|---| | A new MCP server (Stripe, Linear, Sentry…) | catalog/mcps.json | ~5 min | | A new agent / skill / command from a source | catalog/items/<source>.json | ~3 min | | A new bundled skill with inline .md content | catalog/skills.json | ~5 min | | A new framework scaffolder (T3, Tauri…) | catalog/scaffolders.json | ~3 min | | A new command bundle (lib installs for a stack) | catalog/command-bundles.json | ~5 min |

The full guide with templates, the applyWhen semantics, validation, and the PR checklist is in CONTRIBUTING.md.

When you click New Issue, pick the right 🆕 template — it pre-fills the JSON shape we expect, mergeable in minutes.

git clone https://github.com/steph-frtech/claude-code-up
cd claude-code-up
npm install
npm run dev   # iterate
npm run typecheck && npm run build

License

MIT — see LICENSE.

Acknowledgements

• Anthropic for Claude Code and the MCP protocol
• wshobson for the 100-agent marketplace
• Jesse Vincent (obra) for the Superpowers harness
• Matt Pocock for the process skills (via ismaelJimenez/mp-skills mirror)
• Andrej Karpathy for the coding pitfalls observations that inspired the default CLAUDE.md
• iamhenry, Stripe, nahmanmate for the community MCPs

Built with @clack/prompts, tsup, execa.