ONE REPO. EVERY AI TOOL.

Pull once.
Power every tool.

Your AI coding skills, rules, and workflows, in sync across your entire stack.

Get started View on GitHub

npm install -g @raditia/craftkit

CraftKit syncs one repository of skills, rules, commands, and agents across Claude Code, Cursor, Gemini CLI, and Codex CLI. — One shared toolkit for Claude Code, Cursor, Gemini CLI, and Codex CLI.

FROM IDEA TO SHIP

A shared workflow, from idea to ship.

STEP 1

/define

Turn an idea into a clear plan. Move from an interview to a spec, test cases, and an implementation plan before writing code.

Explore the workflow

Example request

“Help me plan a todo app with sign-in, a database, and tests before we start building.”

Clarify the goal and write a spec

Define test cases and an ordered implementation plan

Start building with shared context

FOLLOW ONE REQUEST

From “build this” to a checked result.

One request. Four stages.

YOUR REQUEST

“Add a page where I can see my saved items. Show a loading state while they’re being fetched, and a helpful message when I haven’t saved anything yet.”

Context and implementation lead to fast checks. A failed check blocks review. When checks pass, specialists review alongside test authoring. The main agent combines findings, fixes errors and runs tests before reporting.

STAGE 03

Parallel reviews. One coordinated result.

After selecting specialists, the main agent writes tests while subagents review without editing. It then combines the findings, makes fixes, and runs the tests.

ARCHITECTURE AND WORKFLOWS

Diagram gallery

Explore all 16 diagrams. Open a row to see the full figure; on small screens, swipe across it to see every step.

Derive the change. Read the intent.

Derived Git context informs scaffold, review and tests; feature intent informs review, evaluation and documentation. — Open full size View source

Two providers. One evidence-based review.

Cross-review requires approved Claude and Codex access, runs one independent round and one peer critique, validates responses and lets the host adjudicate. — Open full size View source

Five roles in a frontend feature.

An Entry provides boundaries and context, a View renders Presenter output, a Presenter owns state and effects, a Model owns types and pure logic, and a Resource owns strings. — Open full size View source

Use the workflow the host can run.

Build, review and ship choose parallel workflows when spawning is available and sequential twins otherwise, reporting the validation tradeoff. — Open full size View source

Runtime map: the original proposal.

The historical MCP-context plan connects a Claude-only Gateway to orchestrators, skills, agents, host MCP servers and repository feature state. — Open full size View source

Plan sets the window. Access sets the model.

The Claude routing hook falls back to aliases when account data cannot be read and otherwise selects entitled models within the plan window. — Open full size View source

Build: implement, review, then test.

The main agent routes and checks the project, runs read-only specialist reviews alongside test work, then combines results before reporting. — Open full size View source

Review: independent checks, one verdict.

The main agent routes and checks the project, runs read-only specialist reviews alongside test work, then combines results before reporting. — Open full size View source

Ship: verify before declaring readiness.

The main agent routes and checks the project, runs read-only specialist reviews alongside test work, then combines results before reporting. — Open full size View source

Approve the plan before building.

Define progresses through interview, spec, approved test cases and task planning, with user checkpoints before handing off to build. — Open full size View source

The nearest project marker wins.

The routing hook walks up from cwd, repeats on missing markers and injects the platform selected at the nearest matching level. — Open full size View source

Install once. Work inside the host.

The distributor installs hooks and workflows; orchestrators coordinate workers, host MCP access and per-feature repository state. — Open full size View source

One source. Four tools.

CraftKit distributes rules, skills, commands, agents and injected partials through sync.sh to four independent tool destinations. — Open full size View source

Agreement raises confidence. Evidence decides.

Independent findings are deduplicated, contradictions are settled by evidence and missing reviewers prevent a ready verdict. — Open full size View source

One file, one owner.

An explicit Claude agent-team build checks its runtime, scaffolds and assigns ownership, then stages implementation, review, tests and integration checks. — Open full size View source

Each tool has its own model selector.

Claude uses account entitlements, Gemini uses CLI aliases, Codex inherits its configured default and reasoning effort, and Cursor remains picker-controlled. — Open full size View source

YOUR PLAYBOOK, EVERYWHERE

Switch tools.
Keep your standards.

Keep the source in one place. Let CraftKit handle the tool-specific formats.

One source of truth

Manage skills, rules, commands, and agents in a single repository. Update the source, then sync it across your tools.

Built for your platform

Workflows detect web, React Native, Android, and iOS projects and route work to the relevant specialists.

Stay in sync as you work

Git installs use post-merge and post-rewrite hooks. Pull an update to add, refresh, or remove installed skills.

MAKE IT YOUR TOOLKIT

One install.
A shared way to build.

Bring CraftKit to the AI tools you already use.

Read the installation guide

Install with npm

Install globally, with version pinning and rollback.

npm install -g @raditia/craftkit

Requires bash 3.2+ and curl, plus npm. In Codex, review and trust the hooks with /hooks after installation.