# OpenControl — Universal Control Plane for AI Agents

> SST team's universal control-plane MCP server exposing AWS / Cloudflare / Vercel infra to AI agents in natural language. Audit-logged, scoped.

## Install

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

## Quick Use

1. `npx opencontrol@latest init` to scaffold opencontrol.config.ts
2. Configure providers (aws/cloudflare/vercel) and audit settings
3. Add to Claude Desktop mcpServers config and prompt away

---

## Intro

OpenControl is a universal control-plane MCP server from Dax Raad's team — exposes AWS, Cloudflare, Vercel, and SST infrastructure to AI agents (Claude Code, Cursor, custom agents) via natural language. Every action is audit-logged, scope-limited, and reviewable before execution. Best for: SRE-with-agents workflows, on-call triage, infra cleanup, anywhere you want Claude to actually take action on real infra but with guard rails. Works with: any MCP host. Setup time: 15 minutes.

---

### Install

```bash
npx opencontrol@latest init
# Answers a few questions, generates opencontrol.config.ts
```

### Sample config

```typescript
// opencontrol.config.ts
import { defineConfig } from "opencontrol";

export default defineConfig({
  audit: {
    storage: "s3://my-audit-bucket/opencontrol/",
    require_approval: ["delete", "scale_down", "destroy"],
  },
  providers: {
    aws: {
      regions: ["us-east-1", "eu-west-1"],
      services: ["s3", "lambda", "dynamodb", "cloudfront"],
      role_arn: "arn:aws:iam::123456789:role/opencontrol",
    },
    cloudflare: {
      account_id: "...",
      api_token_env: "CF_API_TOKEN",
    },
    vercel: {
      team_id: "...",
      token_env: "VERCEL_TOKEN",
    },
  },
});
```

### Claude Desktop config

```json
{
  "mcpServers": {
    "opencontrol": {
      "command": "npx",
      "args": ["-y", "opencontrol", "mcp"],
      "env": {
        "OPENCONTROL_CONFIG": "./opencontrol.config.ts"
      }
    }
  }
}
```

### Sample prompts that work

- "Find Lambda functions that haven't been invoked in 90 days and tag them as candidates for deletion."
- "Which Cloudflare workers have error rate above 1% in the last hour?"
- "Roll back the Vercel deployment of `tokrepo-web` to the version before the May 8 deploy."
- "Show me S3 buckets without lifecycle rules and estimated monthly cost."

### Action approval workflow

Actions matching `require_approval` patterns generate a pending audit entry. Approve via:

```bash
opencontrol approvals list           # see pending
opencontrol approvals approve <id>   # green-light
opencontrol approvals reject <id>    # block
```

Configure Slack / email notifications for new approval requests via webhook.

---

### FAQ

**Q: How is this different from cloud MCP servers like AWS MCP?**
A: Cloud MCP servers expose a single cloud's API. OpenControl unifies multiple clouds + adds audit logging + approval workflows. For one-cloud teams, AWS MCP is fine; for multi-cloud or compliance-strict shops, OpenControl is the upgrade.

**Q: What stops an agent from doing real damage?**
A: Three layers: (1) `require_approval` patterns force human green-light on destructive verbs; (2) IAM role scoping limits what credentials can do; (3) every action is S3-logged for forensics. Treat OpenControl as the seatbelt, not the steering wheel — still review what the agent proposes.

**Q: Is it open source?**
A: Yes — MIT licensed. Hosted version planned by Toolbeam for teams who don't want to self-host the control plane. The MCP server itself is fully open-source today.

---

## Source & Thanks

> Built by [Toolbeam](https://github.com/toolbeam) (Dax Raad's team). Licensed under MIT.
>
> [toolbeam/opencontrol](https://github.com/toolbeam/opencontrol) — ⭐ 1,200+

---

<!-- ZH -->

## 快速使用

1. `npx opencontrol@latest init` 生成 opencontrol.config.ts
2. 配 providers（aws/cloudflare/vercel）和 audit 设置
3. 加到 Claude Desktop mcpServers 配置就能用

---

## 简介

OpenControl 是 Dax Raad 团队做的通用控制面 MCP server —— 把 AWS / Cloudflare / Vercel / SST 基建通过自然语言暴露给 AI agent（Claude Code / Cursor / 自定义 agent）。每个动作都审计、范围限定、执行前可复核。适合 SRE 配 agent 的工作流、on-call 分流、基建清理、任何想让 Claude 实际操作真基建但要护栏的场景。任何 MCP 宿主可用。装机时间 15 分钟。

---

### 安装

```bash
npx opencontrol@latest init
# 回答几个问题，生成 opencontrol.config.ts
```

### 配置示例

```typescript
// opencontrol.config.ts
import { defineConfig } from "opencontrol";

export default defineConfig({
  audit: {
    storage: "s3://my-audit-bucket/opencontrol/",
    require_approval: ["delete", "scale_down", "destroy"],
  },
  providers: {
    aws: {
      regions: ["us-east-1", "eu-west-1"],
      services: ["s3", "lambda", "dynamodb", "cloudfront"],
      role_arn: "arn:aws:iam::123456789:role/opencontrol",
    },
    cloudflare: {
      account_id: "...",
      api_token_env: "CF_API_TOKEN",
    },
    vercel: {
      team_id: "...",
      token_env: "VERCEL_TOKEN",
    },
  },
});
```

### Claude Desktop 配置

```json
{
  "mcpServers": {
    "opencontrol": {
      "command": "npx",
      "args": ["-y", "opencontrol", "mcp"],
      "env": {
        "OPENCONTROL_CONFIG": "./opencontrol.config.ts"
      }
    }
  }
}
```

### 实战 prompt 例子

- "找出 90 天没被调用的 Lambda 函数，给它们打标候选删除。"
- "过去 1 小时错误率 >1% 的 Cloudflare worker 是哪些？"
- "把 `tokrepo-web` 的 Vercel 部署回滚到 5 月 8 日发布前那版。"
- "列没有生命周期规则的 S3 bucket + 预估月成本。"

### 动作审批流

匹配 `require_approval` 模式的动作生成待审批 audit 条目。审批方式：

```bash
opencontrol approvals list           # 看待审批
opencontrol approvals approve <id>   # 放行
opencontrol approvals reject <id>    # 阻止
```

通过 webhook 配 Slack / 邮件通知新审批请求。

---

### FAQ

**Q: 跟 AWS MCP 这种云专属 MCP server 啥区别？**
A: 云 MCP 暴露单家云的 API。OpenControl 统一多云 + 加审计 + 加审批流。单云团队 AWS MCP 够用；多云或合规严的店 OpenControl 是升级。

**Q: 怎么防止 agent 干坏事？**
A: 三层：(1) `require_approval` 模式强制人工放行破坏性动词；(2) IAM 角色范围限制凭证能做啥；(3) 每个动作 S3 日志便于取证。OpenControl 是安全带不是方向盘 —— 还是要审 agent 的提案。

**Q: 开源吗？**
A: MIT 开源。Toolbeam 计划出托管版给不想自托管控制面的团队。MCP server 本身今天就是完全开源。

---

## 来源与感谢

> Built by [Toolbeam](https://github.com/toolbeam) (Dax Raad's team). Licensed under MIT.
>
> [toolbeam/opencontrol](https://github.com/toolbeam/opencontrol) — ⭐ 1,200+


---
Source: https://tokrepo.com/en/workflows/opencontrol-universal-control-plane-for-ai-agents
Author: SST