Two hard gates
No code ships without an approved plan and a passing review. Every change must clear both stops — no bypasses, no self-approval.
Enforced discipline for AI coding agents.
Hard gates, TDD, security scanning, and human approval — so your AI agent ships production-quality code, not untested prototypes.
The harness is the product.
Models provide intelligence. MeowKit provides the constraints, safety gates, and repeatable workflows that production engineering demands.
TDD opt-in
Strict test-first discipline when enabled, fast spikes when you need speed. The harness adapts without relaxing quality standards.
4-layer security
Prompt injection defense across input, instruction, context, and output layers. Untrusted content is data — never instructions.
The 7-phase workflow.
Every task follows the same enforced sequence — orient, plan, test, build, review, ship, reflect. Two hard gates block shipping unreviewed or untested code.
Orient
Detect task domain, classify complexity, assign model tier and agents.
Plan
GATEScope-adaptive plan with acceptance criteria. No code until the plan is approved.
Test
Write failing tests first when TDD is enabled. Correctness proof before implementation.
Build
Implement against the approved plan and passing tests. File ownership enforced.
Review
GATEAdversarial structural audit across 5 dimensions. Security scan for BLOCK patterns.
Ship
PR creation, conventional commit, deploy pipeline. Only after Gate 2 clears.
Reflect
Capture lessons, update memory files, run retrospective. Knowledge persists.
Before
# No plan, no gates, no tests
claude "add user auth to the API"
# AI ships code directly:
✗ No approved spec
✗ No failing tests first
✗ No security scan
✗ No review gate
→ untested code in productionWith MeowKit
# Enforced 7-phase workflow
npx meowkit "add user auth"
# Harness enforces:
✓ Plan approved at Gate 1
✓ Tests written first (TDD)
✓ Security scan — no BLOCKs
✓ Review passed at Gate 2
→ production-quality PREverything the harness enforces for you.
- 17 specialist agents
- Dedicated agents for planning, security, review, testing, documentation, and more — each with scoped file ownership and model-tier routing.
- 77 domain skills
- From database migrations to multimodal AI, frontend design to CTF research — skills activate only when the task demands them.
- Cross-session memory
- Lessons, fixes, review patterns, and architecture decisions persist across sessions. The harness learns from every run.
- Scale-adaptive intelligence
- Domain complexity classification routes every task into TRIVIAL, STANDARD, or COMPLEX — picking the right model and scaffolding density automatically.
- Party mode
- Multi-agent deliberation for architectural decisions. Multiple agents argue different positions before a decision is made.
- Adversarial review
- Gate 2 runs parallel reviewers across correctness, security, design, scope, and craft. Any FAIL blocks the ship.
- Zero external dependencies
- Pure prompt engineering — no SDK required. Works offline, works in any Claude Code environment, no vendor lock-in.
- TDD pipeline
- Opt-in test-first enforcement with RED → GREEN → REFACTOR gates. Self-healing loop with 3-attempt cap and human escalation.
By the numbers.
77skills
17agents
21commands
7modes
19rules
27hooks
zero external dependencies · MIT license · works offline · SSR compatible
One command to enforce discipline.
Quick start
$ npx mewkit init # scaffold the harness into your project
$ npx mewkit setup # choose your workflow modes and agents
$ /mk:cook "add feature X" # the 7-phase pipeline enforces the restReady to ship production-quality AI code?
The harness is free, open-source, and works inside Claude Code today. No sign-up. No external dependencies. Just discipline.
Install MeowKit