# Cloudflare Agents — Stateful Agents on Durable Objects

> Cloudflare Agents provides stateful execution environments for agent workloads on Durable Objects, with scheduling, realtime, MCP, and Workers deployment.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

1. Install the verified package or repo entrypoint:

```bash
npm create cloudflare@latest -- --template cloudflare/agents-starter
```

2. Run the first local check:

```bash
npm run dev
```

3. Add a repeatable verification command:

```bash
npx wrangler deploy --dry-run
```

---

## Intro

Cloudflare Agents is a verified GitHub-backed tool for modern AI and developer workflows, sourced from `cloudflare/agents` with 4,898 stars and a MIT license snapshot. Best for: apps that need one persistent agent per user, room, workflow, or session, deployed globally without managing servers. Works with: Cloudflare Workers, Durable Objects, Wrangler, realtime channels, scheduled tasks, MCP tools. Setup time: 15 minutes. Use it when you need a concrete, repeatable path rather than another one-off shell snippet.

---

## Operating Pattern

### Fit check

| Question | Practical answer |
|---|---|
| What do you install? | `agents` from `cloudflare/agents` |
| What is the first command? | `npm run dev` |
| What proves it works? | `npx wrangler deploy --dry-run` |
| How long should a pilot take? | 15 minutes for a small repo or sandbox |

### Adoption loop

1. Run the tool on a disposable branch or sandbox project.
2. Capture before/after output so reviewers can see the exact effect.
3. Add the smallest CI or local check that prevents regressions.
4. Document owner, upgrade command, and rollback command in the repo.

### Recommended use

Model each durable conversation or job as an agent with its own state. Keep external side effects in explicit tools so retries and wakeups stay predictable.

### Guardrails

Design storage boundaries before scale tests. Millions of idle agents are viable only if per-agent state stays small and lifecycle cleanup exists.

### Rollout checklist

- Pin the package or release version before using it in CI.
- Keep credentials in environment variables or the platform secret store.
- Add one owner who is responsible for upgrades and breaking-change triage.
- Re-check the GitHub repo before writing docs that mention APIs or install paths.


---

### FAQ

**Q: Is this production-ready?**
A: The repo exists at `https://github.com/cloudflare/agents` and has 4,898 GitHub stars. Treat the first rollout as a controlled pilot until your team has tested install, rollback, and CI behavior.

**Q: Why use it instead of a generic script?**
A: The value is repeatability: a named package, a documented command, a source repo, and a small verification path that can be reviewed by teammates.

**Q: What should I measure first?**
A: Measure setup time against the 15 minutes target, count how many files or tasks it changes, and record whether the CI command catches the same issue locally.

---

## Source & Thanks

> Built from [cloudflare/agents](https://github.com/cloudflare/agents). License: MIT.
>
> GitHub stars verified from `api.github.com/repos/cloudflare/agents`: 4,898.

---

<!-- ZH -->

## 快速使用

1. 安装已验证的包或仓库入口：

```bash
npm create cloudflare@latest -- --template cloudflare/agents-starter
```

2. 跑第一次本地检查：

```bash
npm run dev
```

3. 加一个可重复验证命令：

```bash
npx wrangler deploy --dry-run
```

---

## 简介

Cloudflare Agents 是一个已通过 GitHub 仓库验证的现代 AI / 开发者工具，来源 `cloudflare/agents`，当前星标快照 4,898，许可证 MIT。适合：需要按用户、房间、工作流或 session 保留一个持久 agent，并希望全球部署但不想管服务器的应用。兼容：Cloudflare Workers、Durable Objects、Wrangler、实时通道、定时任务、MCP 工具。装机时间：15 minutes。当你需要的是可重复落地路径，而不是一次性 shell 片段时，用它。

---

## 操作模式

### 适配检查

| 问题 | 实用答案 |
|---|---|
| 安装什么？ | 来自 `cloudflare/agents` 的 `agents` |
| 第一条命令？ | `npm run dev` |
| 如何证明可用？ | `npx wrangler deploy --dry-run` |
| 小范围试点多久？ | 小仓库或 sandbox 约 15 minutes |

### 接入循环

1. 先在一次性分支或 sandbox 项目里跑。
2. 记录 before / after 输出，让 reviewer 看见实际影响。
3. 加最小的 CI 或本地检查，防止同类问题回归。
4. 在仓库文档里写清 owner、升级命令和回滚命令。

### 推荐用法

把每个持久对话或任务建模成一个拥有自身状态的 agent。外部副作用放进显式工具，确保重试和唤醒行为可预测。

### 风险边界

扩容测试前先设计存储边界。只有单 agent 状态足够小且有生命周期清理时，百万级 idle agents 才可行。

### 推广检查表

- 进入 CI 前固定包版本或 release 版本。
- 凭据放进环境变量或平台 secret store。
- 指定一个 owner 负责升级和 breaking change 分流。
- 文档里写 API 或安装路径前，重新核验 GitHub 仓库。


---

### FAQ

**Q: 能直接上生产吗？**
A: 仓库已验证存在：`https://github.com/cloudflare/agents`，GitHub 星标 4,898。第一次上线仍建议控制试点，先验证安装、回滚和 CI 行为。

**Q: 为什么不用普通脚本？**
A: 价值在可重复：有命名包、文档化命令、源仓库和可被团队 review 的最小验证路径。

**Q: 第一步应该量什么？**
A: 先量装机时间是否接近 15 minutes，它改了多少文件或任务，以及 CI 命令是否能在本地抓到同类问题。

---

## 来源与感谢

> 来源：[cloudflare/agents](https://github.com/cloudflare/agents)。许可证：MIT。
>
> GitHub stars 已通过 `api.github.com/repos/cloudflare/agents` 验证：4,898。


---
Source: https://tokrepo.com/en/workflows/cloudflare-agents-stateful-agents-on-durable-objects
Author: Cloudflare