Claude Code Toolkit

Install

install

$ git clone https://github.com/applied-artificial-intelligence/claude-code-toolkit.git && cp claude-code-toolkit/commands/setup-toolkit.md ~/.claude/commands/

Update

update

$ git -C claude-code-toolkit pull

Pull the clone, then re-run /setup-toolkit in a project to refresh plugin/marketplace config. Plugins resolve from the local clone path, so updates land on pull.

When to use it

You want a structured explore -> plan -> next -> ship loop with per-task git checkpoints instead of one-shot plan-mode execution.
You keep losing context across session resets and need handoff/continue plus a persistent .claude/memory/ store.
You want Anthropic's documented Claude Code patterns (subagents, progressive-disclosure skills, quality-gate hooks) as ready-to-copy plugins.
You need to search past Claude Code sessions ('how did we implement X last time?') across all projects.

Commands & flags

/setup-toolkit: Bootstrap command (installed globally to ~/.claude/commands/) that detects project type and wires up the local plugin marketplace + .claude/settings.json.
/workflow:explore -> /workflow:plan -> /workflow:next -> /workflow:ship: Stateless explore-plan-next-ship task lifecycle with per-task auto-commits.
/transition:handoff and /transition:continue: Save context to disk before a context reset, then resume cleanly after /clear.
/memory:index and /memory:memory-review: Build and curate persistent project knowledge in .claude/memory/.
python3 scripts/session-db.py search "authentication": Index and query past Claude Code session history across projects (session-db.py index | search | timeline).
./scripts/install-git-safe-commit.sh: Install the git wrapper that blocks --no-verify so quality gates can't be skipped.

How to use it

Clone, install the bootstrap command once, then run setup inside any project:

git clone https://github.com/applied-artificial-intelligence/claude-code-toolkit.git
cd claude-code-toolkit
cp commands/setup-toolkit.md ~/.claude/commands/

In any project, launch Claude Code and run:

/setup-toolkit

It detects project type (Python/JS/web), asks which of the 6 plugins and optional MCP servers you want, and writes .claude/settings.json with a local marketplace pointing at the cloned plugins/ dir. To configure manually, add extraKnownMarketplaces.local (source: {source: "directory", path: ".../plugins"}) plus enabledPlugins for system@local, workflow@local, development@local, transition@local, memory@local, setup@local.

Optional hardening: run ./scripts/install-git-safe-commit.sh (blocks --no-verify) and enable the mdtoken pre-commit hook to keep markdown under context limits. MCP servers (Serena, Context7, Chrome DevTools, FireCrawl) are optional and every command degrades gracefully without them.

Why it earns its place

Distilled from 6+ months of daily Claude Code use and explicitly mapped to Anthropic's own best-practice docs (multi-context-window workflows, the memory tool, progressive-disclosure skills, subagents, quality-gate hooks). You get a coherent, file-based, stateless system (workflow + memory + transition + development + system + setup) rather than a loose pile of commands, plus a cross-project session-history search tool (session-db.py) and the companion mdtoken token-limit linter. MIT-licensed, actively maintained (v1.1.0), and honest about Claude's real limits rather than padding with unvalidated token-reduction claims.

Try it

claude

> /workflow:explore "add JWT user authentication" then /workflow:plan and /workflow:next to execute tasks with auto-commits