Centaur Loop / 半人马环

English + 中文 | 简体中文 | Website | Technical Design | 项目定位

The open-source workbench for human-governed AI feedback loops.

面向 AI Agent 反馈闭环的人类治理型开源工作台。

Centaur Loop helps teams run AI agents as accountable operating cycles. Agents can plan and execute, but humans keep judgment authority at explicit gates; real-world feedback becomes reviewed memory for the next run.

半人马环关注 AI Agent 在真实业务中的反馈闭环：人类治理、人工卡点、效果反馈、复盘记忆，以及下一轮持续改进。

Related terms / 相关术语：AI Agent 反馈闭环、AI Agent 工作台、人机协作 Agent、Agent 记忆、AI 工作流治理、AI Agent 基础设施、LLMOps。

Plan -> Approve -> Execute -> Review -> Publish -> Feedback -> Reflect -> Remember -> Next Cycle

Cron wakes agents up. Workflows move agents through steps. Centaur Loop helps agents improve after feedback comes back.

Demo

Centaur Loop content growth loop demo

This demo shows the current MVP running a full content growth loop: AI planning, human gates, draft review, manual publish marking, sample feedback, retrospective review, memory confirmation, and a completed cycle with confirmed memory ready for the next run.

Why It Matters

Most agent systems optimize the moment before output: prompting, tool use, scheduling, orchestration. The hard product problem often starts after output leaves the chat window: Was it approved? Was it published? Did it work? What should the agent remember next time?

Centaur Loop makes that whole cycle the product surface: stage state, human gates, feedback capture, retrospective review, memory candidates, and next-cycle suggestions.

中文说明：很多 Agent 系统只处理“生成之前”的问题，Centaur Loop 更关心生成之后的业务闭环：谁来确认，真实效果如何，哪些经验应该进入记忆，下一轮如何变得更好。

What It Is

A chat-first React workbench for driving an AI feedback loop end to end.
A TypeScript state machine for explicit loop stages and human checkpoints.
A local runtime connector layer for OpenAI-compatible models, Ollama, LM Studio, vLLM, and llama.cpp.
A demoable content growth loop that proves planning, draft review, publishing, feedback, review, memory, and improvement.
A design reference for building human-governed AI products.

中文说明：它是一个 AI Agent 反馈闭环工作台，把计划、人工确认、执行、反馈、复盘、记忆和下一轮建议放在同一个可观察界面里。

What It Is Not

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

Existing runtimes execute tasks. Centaur Loop governs the feedback loop around those tasks.

中文说明：执行层可以继续使用 LangGraph、Temporal、Inngest、n8n 或其他 agent runtime；Centaur Loop 负责治理任务周围的人类判断、结果反馈和记忆沉淀。

MVP Experience

The current MVP focuses on one scenario: Content Growth Loop.

当前 MVP 聚焦一个场景：内容增长闭环。

Start with a weekly growth goal.
AI proposes a structured plan.
Human approves or changes the plan.
AI generates reviewable drafts.
Human approves drafts and marks publishing.
Human submits outcome feedback.
AI reviews results and proposes memory candidates.
Human confirms which lessons become memory.
The next cycle starts with prior memory in context.

Core Lifecycle

planning
  -> awaiting_plan_review
  -> generating
  -> awaiting_review
  -> awaiting_publish
  -> awaiting_feedback
  -> reviewing_auto
  -> awaiting_memory
  -> cycle_complete

Architecture

Layer	Role
`src/core/loopEngine.ts`	Explicit state machine that advances cycles 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 drafts and keeps failures inside the cycle record.
`src/core/loopReviewer.ts`	Converts feedback into retrospectives, lessons, and next-cycle suggestions.
`src/protocol/loopChat.ts`	Maps runtime state to chat messages, cards, and user actions.
`src/adapters/*`	Runtime, 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 only calls the local Vite proxy; API keys never 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 adapter examples are shown for LangGraph, Temporal, 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

Then restart the dev server and select the runtime from the floating runtime menu.

Development

npm run typecheck
npm run build

Roadmap

Extract @centaur-loop/core from the demo workbench.
Add durable storage, notifier, model, and memory adapters.
Add regeneration and revision flows for rejected drafts.
Add integration examples for LangGraph, Mastra, Inngest, Temporal, and n8n-style approvals.
Improve durable execution, idempotency, retry behavior, and checkpoint recovery.
Add richer semantic memory retrieval beyond the current local prototype.

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 the feedback layer around agent work concrete, inspectable, and easy to extend.

Contributing

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

License

MIT