# Coze Loop — Agent Prompt, Eval, and Observability Hub

> Coze Loop unifies prompt iteration, evaluation, and trace observability, helping agent teams debug workflows without jumping across separate tools.

## Install

Save as a script file and run:

## Quick Use

1. Clone and enter the repo:
   ```bash
   git clone https://github.com/coze-dev/coze-loop.git
   cd coze-loop
   ```
2. Configure `release/deployment/docker-compose/conf/model_config.yaml`, then boot locally:
   ```bash
   make compose-up
   ```
3. Verify:
   - Open `http://localhost:8082` and confirm prompt playground, evaluation, and trace views load.

## Intro

Coze Loop unifies prompt iteration, evaluation, and trace observability, helping agent teams debug workflows without jumping across separate tools.

- **Best for:** teams that want prompt debugging, evals, and traces in a single operator console
- **Works with:** Docker Compose, Helm/Kubernetes, model config files, browser-based ops UI
- **Setup time:** 30-60 minutes

## Practical Notes

- Quant: the README exposes two deployment paths out of the box: Docker Compose for local trials and Helm for Kubernetes.
- Quant: the default local entry lands on `http://localhost:8082`, which makes environment verification straightforward.

## Why it matters

Coze Loop is useful when agent teams are already iterating prompts and evaluations, but the evidence is scattered across notebooks, ad-hoc dashboards, and model vendor consoles.

- Prompt development, evaluation, and observability are described as connected modules, which matches how real agent incidents are usually debugged.
- The repo documents both Compose and Helm flows, so it can graduate from a local lab to a shared cluster without switching products.
- The README includes explicit security warnings for public deployments, a good sign that the project understands exposure risks.

## Rollout pattern

- Start in Docker Compose with one model configuration and one evaluation set before you touch Helm.
- Promote only after you can tie a bad answer back to a prompt diff or trace record inside the same system.
- Treat internet exposure as a separate security review because the maintainers call out SSRF and privilege risks directly.

## Watchouts

This is not a no-config SaaS toy: you need to own model credentials, deployment topology, and public-network hardening before treating it as a shared operations plane.

### FAQ

**Q: Can I try it without Kubernetes?**
A: Yes. The README puts Docker Compose first for local deployment and Helm second for cluster rollout.

**Q: What problem does it solve best?**
A: It centralizes prompt iteration, evaluation data, and traces so debugging is less fragmented.

**Q: What is the main risk?**
A: Publishing it on a public network without hardening; the maintainers explicitly warn about security review first.

## Source & Thanks

> Source: https://github.com/coze-dev/coze-loop
> License: Apache-2.0
> GitHub stars: 5,452 · forks: 755

---

<!-- ZH -->

## 快速使用

1. 克隆并进入仓库：
   ```bash
   git clone https://github.com/coze-dev/coze-loop.git
   cd coze-loop
   ```
2. 配置 `release/deployment/docker-compose/conf/model_config.yaml`，然后本地启动：
   ```bash
   make compose-up
   ```
3. 验证：
   - 打开 `http://localhost:8082`，确认 Prompt Playground、评测与 Trace 页面都能访问。

## 简介

Coze Loop 把提示词调试、评测实验与全链路观测放进同一个开源平台，适合希望在同一控制台里同时定位 Agent 工作流问题、比较实验结果并长期保留执行证据、排障线索和复盘沉淀资料的团队。

- **适合谁：** 希望把提示词调试、评测和执行链路观测合并到一个控制台里的团队
- **可搭配：** Docker Compose、Helm/Kubernetes、模型配置文件与浏览器运维界面
- **准备时间：** 30-60 分钟

## 实战建议

- 量化信息：README 直接给了两种部署路径，分别是本地 Docker Compose 和面向 Kubernetes 的 Helm。
- 量化信息：本地默认入口是 `http://localhost:8082`，环境验证路径清晰。 

## 为什么值得收录

如果你的团队已经在做提示词迭代和 Agent 评测，但证据散落在多个控制台和脚本里，Coze Loop 就会显得很有价值。

- 它把 Prompt 开发、评测与观测写成互相关联的模块，更贴近真实 Agent 故障定位流程。
- 同时提供 Compose 与 Helm 路径，意味着它既能本地试用，也能逐步迁移到共享集群。
- README 对公网部署风险写得很直接，说明项目对安全暴露面有清醒认识。

## 落地路径

- 先在 Docker Compose 中固定 1 套模型配置与 1 套评测集，不要一开始就上 Helm。
- 只有当你能在同一个系统里把坏结果追溯到 prompt diff 或 trace 记录时，再考虑推广。
- 公网暴露必须当成单独的安全审查事项，因为维护者明确提到了 SSRF 与权限风险。 

## 注意事项

它不是“零配置即用”的玩具控制台，模型密钥、部署拓扑和公网加固都需要你自己负责。

### FAQ

**没有 Kubernetes 也能试吗？**
答：可以。README 先给了 Docker Compose，本地跑通后再考虑 Helm。

**它最适合解决什么问题？**
答：把提示词迭代、评测数据和执行链路放回同一个系统里，减少排障碎片化。

**最大的风险是什么？**
答：未经加固就直接公网部署；维护者已经明确提醒先做安全评估。

## 来源与感谢

> Source: https://github.com/coze-dev/coze-loop
> License: Apache-2.0
> GitHub stars: 5,452 · forks: 755


---
Source: https://tokrepo.com/en/workflows/coze-loop-agent-prompt-eval-and-observability-hub
Author: Agent Toolkit