LoopKit 🔄

The Open-Source Standard for Autonomous AI Agent Loops Compose. Verify. Deploy.

---

Why LoopKit?

Every team building AI agents is reinventing the same wheel — designing loops, wiring up Maker/Checker verification, managing state across sessions, handling failure recovery. There is no standard.

LoopKit is that standard. Inspired by what React did for UI components and what Docker did for containers, LoopKit gives you:

• Six composable primitives — Trigger, Pipeline, Skill, Connector, SubAgent, Memory
• Built-in Maker/Checker — Verification is in the architecture, not an afterthought
• Loop Readiness Score (LRS) — Know if your loop is production-ready before you run it
• Dual runtime — TypeScript CLI + Python execution engine

Demo

Try it yourself:

# Install Python runtime
cd runtime && pip install -e .

# Run a code review loop (dry run — no LLM cost)
loopkit-runtime --dry-run run pr-review

============================================================
  Running Loop: pr-review
============================================================
Config: loops\pr-review.yaml
Pipeline: 1 step(s)
Verification: Maker=['Bug Review', 'Security Review',
              'Quality Review', 'Performance Review']
              | Checker=Adversarial Verifier

── Pipeline Execution ────────────────────────────────
  Running 4 agents in parallel...
  -> Agent: Bug Review...
  -> Agent: Security Review...
  -> Agent: Quality Review...
  -> Agent: Performance Review...
    + Completed (424ms, 0 up/0 down tokens, $0.0000)

── Maker/Checker Verification ────────────────────────
  Round 1: [ok] PASS (score: 0.95)

── Results ───────────────────────────────────────────
[ok] Loop 'pr-review' completed successfully
  Duration: 424ms
  Steps: 4
  Est. Cost: $0.0000

Quick Start

# 1. Install
npm install -g loopkit

# 2. Initialize a project
loopkit init

# 3. Validate
loopkit validate

# 4. Run a loop (requires ANTHROPIC_API_KEY or OPENAI_API_KEY)
loopkit-runtime run pr-review

Example: PR Review Loop

# loops/pr-review.yaml
name: pr-review
trigger:
  type: webhook
  event: pull_request

pipeline:
  - parallel:
      - prompt: "Review for correctness bugs"
        label: Bug Review
      - prompt: "Review for security vulnerabilities"
        label: Security Review
      - prompt: "Review for code quality"
        label: Quality Review

verify:
  maker: [Bug Review, Security Review, Quality Review]
  checker: Adversarial Verifier
  maxRounds: 3
  autoRetry: true

budget:
  maxTokens: 500000
  maxDurationMinutes: 30

memory:
  store: filesystem
  path: .loopkit/state

Architecture

┌── User ──────────────────────────────────┐
│  loopkit.yaml        loops/              │  ← Declarative YAML
└────────┬──────────────────────┬───────────┘
         │                      │
    ┌────▼──────────────┐ ┌────▼──────────┐
    │  TypeScript CLI   │ │ Python Runtime │   ← Dual runtime
    │  (init/validate   │ │ (execution     │
    │   /run/status)    │ │  engine)       │
    └───────────────────┘ └────┬───────────┘
                               │
                    ┌──────────▼──────────┐
                    │   Loop Engine       │
                    │                     │
                    │  ┌───────────────┐  │
                    │  │ Pipeline      │  │
                    │  │ ┌───┐ ┌───┐   │  │
                    │  │ │ A │ │ B │…  │  │  ← Parallel agents
                    │  │ └─┬─┘ └─┬─┘   │  │
                    │  └───┼─────┼─────┘  │
                    │      │     │        │
                    │  ┌───▼─────▼─────┐  │
                    │  │  Verifier     │  │  ← Maker/Checker
                    │  │  (Checker)    │  │
                    │  └───────┬───────┘  │
                    └──────────┼──────────┘
                               │
                    ┌──────────▼──────────┐
                    │    State Store      │  ← Filesystem persistence
                    │  .loopkit/state/    │
                    └─────────────────────┘

The Six Primitives

Component	Role	Examples
Trigger	What starts the loop	cron, webhook, manual, event
Pipeline	The execution flow	parallel, sequential, conditional
Skill	Reusable domain knowledge	SKILL.md, npm packages
Connector	External system bridge	MCP servers, APIs, databases
SubAgent	Maker/Checker separation	planner, executor, verifier
Memory	Cross-session persistence	filesystem, database, memory

Loop Readiness Score (LRS)

Each loop receives a 0-100 score across 7 dimensions:

loopkit validate
  
  pr-review    — LRS: 85/100 (Grade B)
    [ok] Maker/Checker separation enforced
    [ok] Budget controls configured (3 dimensions)
    [!] State persistence recommended

  daily-triage — LRS: 72/100 (Grade C)
    [ok] Max iterations set
    [!] Add budget controls for production
    [!] Verification missing

Roadmap

• Phase 1 (current): Schema + CLI + LRS + Python Runtime
• Phase 2 (W3-4): Loop Registry, GitHub Action, MCP integration
• Phase 3 (W5-8): Team collaboration, observability dashboard, audit logs
• Phase 4 (Q3): Enterprise SSO, compliance reporting, on-premise deployment

Why This Matters

"I stopped manually prompting Claude. I run a bunch of Loops to prompt it and let it decide what to do next. My job has become writing Loops." — Boris Cherny, Claude Code lead, Anthropic

Loop Engineering is the fourth paradigm shift in AI engineering:

Era	Focus
Prompt Engineering (2022)	Write better prompts
Context Engineering (2023)	Give better context
Tool Engineering (2024)	Build better tools
Loop Engineering (2025+)	Design better loops

License

MIT — see LICENSE