MeowKit

Introduction

What MeowKit is, what problem it solves, and how it differs from raw Claude Code.

MeowKit is an AI agent toolkit for Claude Code that gives your coding assistant enforced discipline — hard gates, TDD, security scanning, and human approval — so it ships production-quality code instead of untested prototypes.

See what's new

Changelog — latest features, improvements, and fixes.

What problem does MeowKit solve?

AI coding tools are powerful but undirected. Without structure, they skip tests, ignore security, and ship untested code. A single "implement this feature" prompt can produce code that compiles but has no tests, no review, and secrets hardcoded in source.

MeowKit fixes this by installing a .claude/ directory that Claude Code reads automatically. It contains specialist agents, domain skills, lifecycle hooks, security rules, and a memory system that together enforce a structured development workflow.

How it works

When you start a Claude Code session in a MeowKit project, the toolkit automatically:

  1. Routes your task to the right agent based on complexity (orchestrator)
  2. Creates a plan before any code is written (planner, Gate 1)
  3. Writes tests — failing tests first if --tdd / MEOWKIT_TDD=1 is set (strict TDD mode); otherwise tests are recommended but not gated (tester invoked on-request)
  4. Implements with discipline — builds until tests pass (developer)
  5. Reviews across 5 dimensions — architecture, types, tests, security, performance (reviewer, Gate 2)
  6. Ships safely — conventional commits, PR, CI verification, rollback docs (shipper)
  7. Learns from the session — captures patterns and lessons for next time (analyst)

No step can be skipped. Two hard gates (plan approval + review approval) require explicit human sign-off.

How it differs from raw Claude Code

ConcernRaw Claude CodeWith MeowKit
PlanningStarts coding immediatelyCreates and gets approval for a plan first
TestingTests optional, often skippedTDD opt-in via --tdd — strict failing-test-first when enabled, recommended otherwise
SecurityRelies on model knowledge4-layer defense + security agent + post-write hooks
ReviewAsk "review this" and hope3 parallel adversarial reviewers + triage step
Shipping"git add -A && git push"Conventional commits, PR, CI verification, rollback docs
MemoryForgets everything between sessionsPersists lessons, patterns, and costs
Model selectionSame model for everythingDomain-adaptive routing — fintech forces COMPLEX tier
Architecture decisionsAsk and hope for the bestParty Mode: 2-4 agents deliberate, forced synthesis
Parallel workSingle-threadedWorktree-isolated parallel agents, max 3, with integration gate

Architecture at a glance

.claude/
├── agents/          Specialist agents for each phase
├── skills/          Domain skills with mk: namespace (step-file decomposition)
├── hooks/           Lifecycle hooks (security scan, TDD gate)
├── rules/           Enforcement rules (security, injection, TDD, parallel execution)
├── scripts/         Python validators (stdlib only)
├── memory/          Cross-session persistence
└── settings.json    Hook registrations + permissions

CLAUDE.md            Entry point — Claude reads this at session start
docs/
└── project-context.md   Agent "constitution" — loaded by ALL agents at session start

Next steps

On this page