Claude Code setup

A complete, working Claude Code configuration, exported from a daily-driver machine: 61 agents, 154 skills, 29 plugins, 9 MCP servers, hooks, rules, slash commands, a file-based memory system with scheduled maintenance, and hardened settings.

Clone it, run the installer, restart Claude Code.

No secrets, API keys, hostnames or personal data are in this repo. Every credential is a <YOUR_...> placeholder.

Quick start

git clone https://github.com/duart-555/claude-code-setup.git && cd claude-code-setup
./install.sh --dry-run        # show what it will do
./install.sh                  # files + marketplaces + plugins + keyless MCP servers
./tools/prereqs.sh            # CLIs: bun, gbrain, ollama(+embeddings), skillspector, caliper, ...
                              # (bootstraps its own toolchain; tested from a bare Ubuntu 24.04; --core for the minimum)
./tools/install-schedules.sh        # optional, macOS: memory backup/archive/health jobs
./tools/install-crawl4ai.sh         # optional: Crawl4AI in Docker, auto-restarting (for the c4ai-sse MCP)

Needs jq, node 20+, git and the claude CLI. The installer backs up what it replaces to ~/.claude/backup-before-setup-<timestamp>/, never overwrites an existing settings.json or CLAUDE.md (it writes settings.json.from-setup and leaves CLAUDE.md.example for you to merge).

Flags: --dry-run, --no-plugins, --no-mcp.

What is in here

Path	Contents
agents/	61 subagents: planner, architect, code / security / database / language reviewers, build resolvers, TDD guide, and the gsd-* phase agents
skills/	154 skills: 48 engineering (ECC), 71 gsd-*, 13 design-taste, 6 web-quality, 3 Caliper, last30days, ui-ux-pro-max, stop-slop, anti-slop, 5 hand-authored (brainstorming, systematic-debugging, verification-before-completion, receiving-code-review, memory-audit), and 4 without a confirmed upstream license (see SKILLS.md)
gsd-core/	Runtime for the gsd-* skills and agents (workflows, templates, references, gsd-tools). Installed to ~/.claude/gsd-core; the skills call it by that path
hooks/, scripts/	Hook implementations and libraries: dash-path guard, memory auto-capture, recall-on-failure, session tracking, quality gates, memory generator/backup/archive with tests
rules/	Always-on rules: coding style, testing, security, git workflow, review standards
commands/	23 slash commands
settings.json	Hooks, enabled plugins, 11 marketplaces, skill overrides, deny rules. No personal allow-list
plugins.txt	29 plugins: caveman, chisle, claude-mem, context-mode, reticle, superpowers, vercel, stripe, ui-ux-pro-max, marketing-skills, and the Knowledge-Work set (productivity, product-management, engineering, data, design, marketing, sales, legal, finance, operations, support, enterprise-search, small-business, product-tracking)
mcp-servers.json	context7, playwright, gbrain, greptile, mobbin, magic, skillspector, gitmcp, c4ai-sse
CLAUDE.md.example	Portable global instructions (no personal content): memory layers, 95% rule, effort triage, GSD vs superpowers, agent/MCP triage, memory contract
launchd/	macOS job templates: weekly memory backup + archive, daily health report, nightly gbrain consolidation
evals/	Caliper spec for the chisle skill
tools/	prereqs.sh (bun, gbrain, ollama + embeddings, skillspector, caliper, code-review-graph, age, ripgrep, yt-dlp, opencli, mcporter, vercel, agent-reach CLIs), install-schedules.sh, install-crawl4ai.sh

The memory system

scripts/memory-index.mjs implements a small durable-memory contract: one fact per file, a generated index, verify_by dates, archiving and search. Hooks capture memory at session end and after failures; launchd/ jobs archive and back it up. 122 tests:

node --test scripts/memory-index.test.mjs scripts/dash-path-guard.test.mjs

Read the layer model and rules in CLAUDE.md.example.

Credentials you must supply

Keyless MCP servers are registered automatically. These are skipped until you edit mcp-servers.json and run claude mcp add-json <name> '<json>' -s user:

• magic (21st.dev): API_KEY
• c4ai-sse (Crawl4AI): Authorization header, plus the Crawl4AI container (see below)
• mobbin, greptile: registered automatically; sign in on first use via /mcp

Memory backup (optional) needs ~/.claude/.env.memory-backup; start from env.memory-backup.example. It encrypts with age before upload and refuses to run unencrypted.

Always pass -s user to claude mcp add; without it a server is tied to the directory it was added from.

Crawl4AI in Docker, running automatically

The c4ai-sse MCP server talks to a local Crawl4AI container. In this setup it runs permanently in Docker: the container is created with --restart unless-stopped, Docker Desktop starts at login, so Crawl4AI is simply always there and no one has to start it.

./tools/install-crawl4ai.sh

It pulls unclecode/crawl4ai, runs it on 127.0.0.1:11235 with 1 GB of shared memory, a generated API token and the restart policy above, then prints the claude mcp add-json line to register it. Re-running it is safe. Needs Docker running. Crawl4AI is a fallback target, not a default: the instructions in CLAUDE.md.example use plain WebFetch unless a task needs a whole-site crawl, a JS-rendered page, structured extraction, or a screenshot/PDF, and fall back to WebFetch when the container is unreachable.

Small settings that matter

These live outside the files above, so set them yourself:

• claude-mem (~/.claude-mem/settings.json): CLAUDE_MEM_PROVIDER=claude, CLAUDE_MEM_CLAUDE_AUTH_METHOD=subscription (summaries go through your existing Claude login, not a separate API key), CLAUDE_MEM_RUNTIME=worker, CLAUDE_MEM_MODE=code.
• gbrain (~/.gbrain/config.json): engine: pglite, embedding_model: ollama:mxbai-embed-large, embedding_dimensions: 1024. Pull the model with ollama pull mxbai-embed-large (tools/prereqs.sh does it). Mixing 768-dim and 1024-dim embedding models fails with a dimension mismatch, so pick one.
• Claude in Chrome: install the extension, then alias claude='claude --chrome' in your shell profile.
• Auto mode is not enabled by this repo (defaultMode is unset). If you turn it on, write your own autoMode.environment describing your repos and domains; the author's is project-specific and deliberately not shared.
• Permissions: only a small generic allow-list is shipped. Your own allow rules will grow with use.

Review before trusting

• settings.json sets no defaultMode: permissions are Claude Code's defaults plus a deny list for .env files. Add your own allow rules.
• Hooks run local scripts from ~/.claude. They are short; read them, as with any hook.
• skillOverrides turns off the C++ and Django skill packs. Flip them if you use those stacks.
• 29 plugins and 154 skills is a lot of always-loaded description text. Trim plugins.txt and skills/ to what you use; skills/context-budget audits the cost.
• Four skills have no confirmed upstream license (apple-design-hig, apple-design, agent-reach, humanizer). See SKILLS.md before redistributing.

Not included, on purpose

• Personal memory corpora, session history, logs, caches, and any credential.
• Runtime-generated skills (learned, synced).

Licensing

Original files: MIT (LICENSE). Vendored third-party skills, agents, hooks and rules keep their upstream MIT licenses; sources and licenses are listed in THIRD_PARTY_LICENSES.md and NOTICE.