Codex

Equip Codex with raw iPhone inspection and native event synthesis through a standard MCP interface.通过标准 MCP 协议为 Codex 挂载 iPhone 读屏与原生事件合成能力。

MCP clients

Poki MCP
Device control

iPhone

Stateless tool surface无状态工具层

Codex handles all planning and reasoning while Poki MCP acts as a deterministic hardware bridge without intermediate agent state.由 Codex 完全主导推理与规划,Poki MCP 仅提供确定性的 iPhone 读写能力,不夹带额外代理状态。

Structured screen context结构化无障碍树读屏

Query the iOS accessibility hierarchy on demand instead of parsing screenshots, keeping token usage predictable during long sessions.按需拉取 iOS 界面节点层级与属性,免去高频截图解析的上下文开销。

Native touch and keyboard synthesis原生触摸与键盘输入

Dispatch hardware-level touch coordinates, gestures, and string inputs directly from Codex tool execution.从 Codex 直接下发坐标点击、手势滑动与文本输入,底层以系统原生事件驱动设备。

Device Control: Poki MCP设备控制层(Poki MCP)

Poki MCP is a standard Model Context Protocol server. It exposes screen reading, touch, text input, batch execution, and screen recording tools to MCP clients like Claude Code and Codex. Screen structure is read from the system accessibility tree, and touch inputs are delivered to the physical device as native events without jailbreaking or external hardware.Poki MCP 是标准的 Model Context Protocol 服务,向 Claude Code、Codex 等 MCP 客户端提供读屏、触摸、输入、批量执行和录屏工具。屏幕结构取自系统无障碍树,触摸通过原生事件发送到真实设备,全程无需越狱或外接硬件。

How do I connect my own agent setup?如何接入现有的工作流?

You can run Poki Agent directly with your OpenAI account, or connect MCP clients like Claude Code and Codex straight to the Poki MCP server.可运行内置的 Poki Agent 并配置 OpenAI 账号;使用 Claude Code 或 Codex 时,将 Poki MCP 注册为本地 MCP 服务即可调用设备控制工具。

Read the docs阅读文档