centaurloop

Name: centaurloop
Author: finewood2008

by finewood2008

Warning

半人马环 Centaur Loop：AI 员工的最小工作单元框架。把复杂岗位拆解为可由 AI 接管、由人类治理、由反馈和记忆持续进化的循环工作流 / The smallest work unit for building AI employees.

100stars

17forks

TypeScript

Added 8/13/2026

View on GitHub Download ZIP Scan for vulnerabilities

Requires a passing catalog security scan. Resolve the flagged issues and resubmit to enable featuring.

AI Agentsagent-memoryagent-orchestrationagent-runtimeai-agentai-agentsai-governanceai-infrastructureai-startupai-workbenchai-workflowchinesefeedback-loophuman-ai-collaborationhuman-in-the-loopllmopsopen-source-aiopenai-compatiblereacttypescriptworkflow-automation

Installation

# Add to your Claude Code skills
git clone https://github.com/finewood2008/centaurloop

Getting Started

Guides for using ai agents skills like centaurloop.

Security ReportWarning

Last scanned: 8/13/2026

{
  "issues": [
    {
      "type": "npm-audit",
      "message": "@babel/core: @babel/core: Arbitrary File Read via sourceMappingURL Comment",
      "severity": "low"
    },
    {
      "type": "npm-audit",
      "message": "nanoid: nanoid: non-secure generators can loop indefinitely with negative size",
      "severity": "high"
    },
    {
      "type": "npm-audit",
      "message": "postcss: PostCSS: Path Traversal in Previous Source Map Auto-Loading (sourceMappingURL) leads to Arbitrary .map File Disclosure",
      "severity": "high"
    },
    {
      "type": "npm-audit",
      "message": "vite: launch-editor: NTLMv2 hash disclosure via UNC path handling on Windows",
      "severity": "high"
    }
  ],
  "status": "WARNING",
  "scannedAt": "2026-08-13T05:40:57.635Z",
  "npmAuditRan": true,
  "pipAuditRan": true,
  "promptInjectionRan": true
}

README.md

Centaur Loop / 半人马环

English | 简体中文 | Website | Technical Design | 中文定位

Centaur Loop is an open-source framework for building AI employees from governable work loops.

半人马环是 AI 员工的最小工作单元框架。

A digital worker is not one big Agent. It is a set of governable, composable, evolving Centaur Loops.

Each Centaur Loop is the smallest AI-takable unit of work: a repeatable loop with a clear goal, context, AI execution, human gates, real-world feedback, confirmed memory, and next-cycle improvement.

中文简述：AI 员工不是一个大而全的 Agent，而是由多个可治理、可组合、可持续进化的半人马环组成。每个半人马环都是一个可被 AI 接管的最小循环工作流。

Digital Worker
  = Research Loop
  + Draft Loop
  + Review / Publish Loop
  + Feedback Memory Loop
  + more governable loops

The loop is the unit of digital labor.

数字劳动的基本单位不是 prompt、task 或 workflow，而是可治理的工作闭环。

Why This Exists

Most AI systems still optimize the wrong layer:

Prompt tools optimize single outputs.
Workflow engines move steps forward.
Agent frameworks plan and execute tasks.
Cron wakes jobs up.

But an AI employee needs more than execution. It needs bounded responsibility, human judgment, feedback, memory, and the ability to improve across cycles.

Centaur Loop turns work into small, governable loops that can be gradually taken over by AI while keeping humans in control of judgment, risk, publishing, and memory.

中文说明：半人马环关注的不是“让 AI 做一次任务”，而是“如何把岗位拆成可治理、可验证、可逐步交给 AI 的工作单元”。

Core Definition

A Centaur Loop contains:

Part	Meaning
Goal	The business outcome this loop owns.
Context & Input	Data, tools, constraints, history, and memory.
AI Execution	The work that can be delegated to a model, agent, or runtime.
Human Gates	Judgment points for plan approval, quality, risk, compliance, publishing, and memory.
Feedback	Real-world outcomes, metrics, screenshots, notes, and corrections.
Confirmed Memory	Lessons approved by humans and reused in the next cycle.
Next Cycle	The loop runs again with better context.

In short:

Goal -> Context -> AI Execution -> Human Gates -> Feedback -> Memory -> Next Cycle

Product Shape

Layer	Role
Digital Worker	A long-running AI work system composed of multiple Centaur Loops.
Centaur Loop	The smallest AI-takable work unit.
Loop Studio	The UI for designing, driving, and observing loops.
Loop Runtime	The execution layer: model, agent, workflow, or external runtime.
Human Gates	Explicit approval and judgment checkpoints.
Feedback Memory	The layer that turns real outcomes into reusable capability.

Current repo status:

Working local React/Vite workbench.
TypeScript loop state machine.
Chat-first loop protocol.
Runtime connectors for demo and OpenAI-compatible/local model paths.
Built-in starter loops for content growth, customer support, and product iteration.

Demo

Centaur Loop content growth loop demo

The flagship demo is Content Growth Worker / Content Growth Loop. It demonstrates how one digital worker can be decomposed into loops for planning, drafting, human review, publishing, feedback, retrospective review, memory confirmation, and next-cycle improvement.

中文说明：当前旗舰 demo 是内容增长 AI 员工的雏形，用多个半人马环证明“岗位能力如何被拆解、治理、反馈和记忆”。

Starter Digital Workers

Digital Worker	Example Loops
Content Growth Worker	Research loop, draft loop, review/publish loop, performance feedback loop, memory loop.
Customer Support Worker	Ticket triage loop, answer draft loop, escalation gate, satisfaction feedback loop, policy memory loop.
Product Iteration Worker	Feedback mining loop, requirement synthesis loop, prioritization gate, release review loop, outcome memory loop.

What It Is Not

Not a generic workflow canvas.
Not a cron scheduler.
Not a publishing bot.
Not one autonomous super-agent.
Not a replacement for LangGraph, Temporal, Inngest, n8n, Mastra, or agent frameworks.

Those systems can execute work. Centaur Loop defines the governable unit of work that can become part of an AI employee.

Architecture

Module	Role
`src/core/loopEngine.ts`	Explicit state machine that advances loops and stops at human gates.
`src/core/loopPlanner.ts`	Turns goals, memory, business context, and tools into structured plans.
`src/core/loopExecutor.ts`	Generates reviewable outputs and keeps failures inside the loop record.
`src/core/loopReviewer.ts`	Converts feedback into retrospective review, lessons, and next-cycle suggestions.
`src/protocol/loopChat.ts`	Maps runtime state to chat messages, cards, and user actions.
`src/adapters/*`	Runtime, model, tool, feedback, and memory boundaries.
`src/ui/*`	Chat-first workbench, embedded action cards, runtime dropdown, feedback and memory surfaces.

Runtime Connectors

Centaur Loop runs without an API key through the deterministic demo runtime. For real models, the browser calls the local Vite proxy; API keys do not enter the frontend bundle.

Supported runtime paths today:

local-demo: built-in deterministic demo runtime.
openai-compatible-env: any OpenAI-compatible /chat/completions endpoint configured through environment variables.
ollama-local: detected through 127.0.0.1:11434/api/tags and called through /api/chat.
lm-studio-local: detected through 127.0.0.1:1234/v1/models.
vllm-local: detected through 127.0.0.1:8000/v1/models.
llamacpp-local: detected through 127.0.0.1:8080/v1/models.

Planned adapters include LangGraph, Temporal, Inngest, Mastra, and n8n-style approval flows.

Quick Start

npm install
npm run dev

Open the Vite URL printed in your terminal. The app works immediately with the demo runtime.

Real Model Setup

Create .env.local:

cp .env.example .env.local

Configure an OpenAI-compatible endpoint:

CENTAUR_MODEL_BASE_URL=https://api.openai.com/v1
CENTAUR_MODEL_API_KEY=your_key_here
CENTAUR_MODEL_NAME=gpt-4o-mini

Restart the dev server and select the runtime from the floating runtime menu.

Development

npm run typecheck
npm run build

Roadmap

Loop Unit Framework: make Centaur Loop a clear reusable unit for building AI employees.
Core Package: extract @centaur-loop/core from the demo workbench.
Adapter Ecosystem: add storage, notifier, model, memory, LangGraph, Mastra, Inngest, Temporal, and n8n-style adapters.
Digital Worker Studio: compose multiple loops into one role-level worker.
Team Workspace: shared loop history, team review, organizational memory, and loop analytics.
Centaur Loop Cloud: hosted runtime management, durable execution, managed memory, and observability.

Project Status

Centaur Loop is early. The current codebase is a working MVP and product design reference, not a stable library API yet. The goal is to make AI employees concrete: not a vague autonomous agent, but a system of inspectable, governable, improving work loops.

中文说明：当前版本是可运行的开源工作台 MVP；长期方向是用半人马环构建 AI 员工。

Contributing

Focused issues and small PRs are welcome. See CONTRIBUTING.md before opening larger design changes.

License

MIT

Frequently Asked Questions

What is centaurloop?

centaurloop is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by finewood2008. 半人马环 Centaur Loop：AI 员工的最小工作单元框架。把复杂岗位拆解为可由 AI 接管、由人类治理、由反馈和记忆持续进化的循环工作流 / The smallest work unit for building AI employees. It has 100 GitHub stars.

Is centaurloop safe to use?

centaurloop returned warnings in SkillsLLM's automated security scan. It has no critical vulnerabilities, but review the flagged issues in the Security Report section before adding it to your workflow.

How do I install centaurloop?

Clone the repository with "git clone https://github.com/finewood2008/centaurloop" and add it to your Claude Code skills directory (see the Installation section above).

What programming language is centaurloop written in?

centaurloop is primarily written in TypeScript. It is open-source under finewood2008 on GitHub, so you can review or fork the full source.

Are there alternatives to centaurloop?

Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh centaurloop against similar tools.

Agentic AI for Beginners

Build your first AI agent from scratch - tool use, ReAct pattern, memory, deployment

41 minBeginner

Comments (0)

to leave a comment.

No comments yet. Be the first to share your thoughts!

Related Skills

ECC

by affaan-m

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

239,787

36,396

JavaScript

AI Agentsai-agentsanthropic

View details

Compare

superpowers

by obra

An agentic skills framework & software development methodology that works.

234,966

20,863

Shell

AI Agentsaibrainstorming

The agent that grows with you

229,689

45,347

Python

AI Agentsaiai-agent

View details

Compare

n8n

by n8n-io

Fair-code workflow automation platform with native AI capabilities. Combine visual building with custom code, self-host or cloud, 400+ integrations.

200,415

60,102

TypeScript

MCP Serversaiapis

View details

Compare

everything-claude-code

by affaan-m

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

185,940

28,768

JavaScript

AI Agentsai-agentsanthropic

View details

Compare

cc-switch

by farion1231

A cross-platform desktop All-in-One assistant for Claude Code, Codex, OpenCode, OpenClaw, Grok Build & Hermes Agent. Only official website: ccswitch.io

126,799

8,655

Rust

AI Agentsai-toolsclaude-code

View details

Compare

Browse all AI Agents skills

SoloFlow claudekit

Centaur Loop / 半人马环

English | 简体中文 | Website | Technical Design | 中文定位

Centaur Loop is an open-source framework for building AI employees from governable work loops.

半人马环是 AI 员工的最小工作单元框架。

A digital worker is not one big Agent. It is a set of governable, composable, evolving Centaur Loops.

Each Centaur Loop is the smallest AI-takable unit of work: a repeatable loop with a clear goal, context, AI execution, human gates, real-world feedback, confirmed memory, and next-cycle improvement.

Digital Worker
  = Research Loop
  + Draft Loop
  + Review / Publish Loop
  + Feedback Memory Loop
  + more governable loops

The loop is the unit of digital labor.

数字劳动的基本单位不是 prompt、task 或 workflow，而是可治理的工作闭环。

Why This Exists

Most AI systems still optimize the wrong layer:

Prompt tools optimize single outputs.
Workflow engines move steps forward.
Agent frameworks plan and execute tasks.
Cron wakes jobs up.

But an AI employee needs more than execution. It needs bounded responsibility, human judgment, feedback, memory, and the ability to improve across cycles.

Centaur Loop turns work into small, governable loops that can be gradually taken over by AI while keeping humans in control of judgment, risk, publishing, and memory.

中文说明：半人马环关注的不是“让 AI 做一次任务”，而是“如何把岗位拆成可治理、可验证、可逐步交给 AI 的工作单元”。

Core Definition

A Centaur Loop contains:

Part	Meaning
Goal	The business outcome this loop owns.
Context & Input	Data, tools, constraints, history, and memory.
AI Execution	The work that can be delegated to a model, agent, or runtime.
Human Gates	Judgment points for plan approval, quality, risk, compliance, publishing, and memory.
Feedback	Real-world outcomes, metrics, screenshots, notes, and corrections.
Confirmed Memory	Lessons approved by humans and reused in the next cycle.
Next Cycle	The loop runs again with better context.

In short:

Goal -> Context -> AI Execution -> Human Gates -> Feedback -> Memory -> Next Cycle

Product Shape

Layer	Role
Digital Worker	A long-running AI work system composed of multiple Centaur Loops.
Centaur Loop	The smallest AI-takable work unit.
Loop Studio	The UI for designing, driving, and observing loops.
Loop Runtime	The execution layer: model, agent, workflow, or external runtime.
Human Gates	Explicit approval and judgment checkpoints.
Feedback Memory	The layer that turns real outcomes into reusable capability.

Current repo status:

Working local React/Vite workbench.
TypeScript loop state machine.
Chat-first loop protocol.
Runtime connectors for demo and OpenAI-compatible/local model paths.
Built-in starter loops for content growth, customer support, and product iteration.

Demo

Centaur Loop content growth loop demo

中文说明：当前旗舰 demo 是内容增长 AI 员工的雏形，用多个半人马环证明“岗位能力如何被拆解、治理、反馈和记忆”。

Starter Digital Workers

Digital Worker	Example Loops
Content Growth Worker	Research loop, draft loop, review/publish loop, performance feedback loop, memory loop.
Customer Support Worker	Ticket triage loop, answer draft loop, escalation gate, satisfaction feedback loop, policy memory loop.
Product Iteration Worker	Feedback mining loop, requirement synthesis loop, prioritization gate, release review loop, outcome memory loop.

What It Is Not

Not a generic workflow canvas.
Not a cron scheduler.
Not a publishing bot.
Not one autonomous super-agent.
Not a replacement for LangGraph, Temporal, Inngest, n8n, Mastra, or agent frameworks.

Those systems can execute work. Centaur Loop defines the governable unit of work that can become part of an AI employee.

Architecture

Module	Role
`src/core/loopEngine.ts`	Explicit state machine that advances loops and stops at human gates.
`src/core/loopPlanner.ts`	Turns goals, memory, business context, and tools into structured plans.
`src/core/loopExecutor.ts`	Generates reviewable outputs and keeps failures inside the loop record.
`src/core/loopReviewer.ts`	Converts feedback into retrospective review, lessons, and next-cycle suggestions.
`src/protocol/loopChat.ts`	Maps runtime state to chat messages, cards, and user actions.
`src/adapters/*`	Runtime, model, tool, feedback, and memory boundaries.
`src/ui/*`	Chat-first workbench, embedded action cards, runtime dropdown, feedback and memory surfaces.

Runtime Connectors

Centaur Loop runs without an API key through the deterministic demo runtime. For real models, the browser calls the local Vite proxy; API keys do not enter the frontend bundle.

Supported runtime paths today:

local-demo: built-in deterministic demo runtime.
openai-compatible-env: any OpenAI-compatible /chat/completions endpoint configured through environment variables.
ollama-local: detected through 127.0.0.1:11434/api/tags and called through /api/chat.
lm-studio-local: detected through 127.0.0.1:1234/v1/models.
vllm-local: detected through 127.0.0.1:8000/v1/models.
llamacpp-local: detected through 127.0.0.1:8080/v1/models.

Planned adapters include LangGraph, Temporal, Inngest, Mastra, and n8n-style approval flows.

Quick Start

npm install
npm run dev

Open the Vite URL printed in your terminal. The app works immediately with the demo runtime.

Real Model Setup

Create .env.local:

cp .env.example .env.local

Configure an OpenAI-compatible endpoint:

CENTAUR_MODEL_BASE_URL=https://api.openai.com/v1
CENTAUR_MODEL_API_KEY=your_key_here
CENTAUR_MODEL_NAME=gpt-4o-mini

Restart the dev server and select the runtime from the floating runtime menu.

Development

npm run typecheck
npm run build

Roadmap

Loop Unit Framework: make Centaur Loop a clear reusable unit for building AI employees.
Core Package: extract @centaur-loop/core from the demo workbench.
Adapter Ecosystem: add storage, notifier, model, memory, LangGraph, Mastra, Inngest, Temporal, and n8n-style adapters.
Digital Worker Studio: compose multiple loops into one role-level worker.
Team Workspace: shared loop history, team review, organizational memory, and loop analytics.
Centaur Loop Cloud: hosted runtime management, durable execution, managed memory, and observability.

Project Status

中文说明：当前版本是可运行的开源工作台 MVP；长期方向是用半人马环构建 AI 员工。

Contributing

Focused issues and small PRs are welcome. See CONTRIBUTING.md before opening larger design changes.

License

MIT