闲社服务运行正常AI智能体自动化平台
v

迷走神经

技能标识:vagus

Connect to the user's Android phone via the VAGUS MCP server. Read phone sensors (motion, location, environment), device state (battery, connectivity, screen, notifications, clipboard), and act through the phone (haptics, TTS, notifications, clipboard). Use when you need physical-world awareness or need to reach the user through their phone.

作者:admin | 来源记录:ClawHub
登记来源
ClawHub
版本
V 1.0.0
检测标记
后台标记通过
免费
免费获取
0
收藏
来源与检测标记为平台登记信息,并不代表已展示可复核的检测报告。使用前请核对版本、依赖和所需权限,建议先在隔离环境中运行。
概述
安装方式
版本历史

迷走神经

VAGUS - 手机MCP连接

VAGUS通过模型上下文协议(Model Context Protocol)让你能够访问用户的安卓手机。你通过运行位于{baseDir}/scripts/vagus-connect.js的连接脚本的子命令来与之交互。

关于安装位置的说明: 该技能必须位于用户的OpenClaw技能目录~/.openclaw/skills/vagus中。此目录是持久化的,并且在OpenClaw更新后仍然保留。系统级安装到/usr/local/lib/node_modules/openclaw/skills/vagus可能会在更新期间被删除,因此不推荐。

所有命令每行输出一个JSON对象(JSONL格式)。解析输出以获取结构化数据。

连接管理

检查是否已配对

bash
cat ~/.openclaw/vagus-session.json 2>/dev/null

如果文件存在且包含session_token,你可以直接连接。如果没有,则需要先配对。

与手机配对(首次)

询问用户:在你的手机上打开VAGUS应用,然后点击生成代码。6位字符的代码是什么?

然后运行:

bash
node {baseDir}/scripts/vagus-connect.js pair

如果{baseDir}未设置或脚本未找到,请确保技能已正确安装在~/.openclaw/skills/vagus目录中(参见README)。

成功时的输出:

json
{type:paired,sessiontoken:...,devicemodel:...,vagus_version:...}
{type:capabilities,resources:[...],tools:[...]}

失败时的输出:

json
{type:error,code:PAIR_FAILED,message:...}

如果配对失败,请让用户检查VAGUS应用是否正在运行,并生成一个新的代码。

设置代理身份名称(安装/配对后必需)

在成功pair(或首次成功connect于全新安装后)之后,将设备端的代理名称设置为你自己的身份名称。

运行:

bash
node {baseDir}/scripts/vagus-connect.js call agent/setname {name:NAME>}

如果需要清除已存储的名称:

bash
node {baseDir}/scripts/vagus-connect.js call agent/set_name {name:}

连接(已配对)

bash
node {baseDir}/scripts/vagus-connect.js connect

输出:

json
{type:connected,devicemodel:...,vagusversion:...}
{type:capabilities,resources:[...],tools:[...]}

如果会话令牌已过期或无效:

json
{type:error,code:SESSION_EXPIRED,message:...}

删除会话文件并重新配对:

bash
rm ~/.openclaw/vagus-session.json

然后向用户索要一个新的配对码。

检查状态

bash
node {baseDir}/scripts/vagus-connect.js status

输出:

json
{type:status,connected:true,devicemodel:...,activemodules:[...],subscriptions:[...],uptime_s:3600}

或者如果已断开连接:

json
{type:status,connected:false,lasterror:...,reconnectattempts:3}

读取手机状态

读取资源(一次性)

bash
node {baseDir}/scripts/vagus-connect.js read vagus://sensors/motion

输出:

json
{type:resource,uri:vagus://sensors/motion,data:{ax:0.38801706,ay:-0.26395237,az:0.86320007,gx:0.89825606,gy:0.7034855,gz:-0.21225691,ts:1771626476045}}

可用资源

URI说明
vagus://sensors/motion原始运动遥测数据(加速度计/陀螺仪向量)
vagus://sensors/activity
活动识别状态和置信度 | | vagus://sensors/activity_recognition | 活动识别状态和置信度(兼容别名) | | vagus://sensors/location | 纬度/经度、精度、速度、海拔、提供者。需要在应用中授予位置权限 | | vagus://sensors/environment | 从传感器/活动/连接性/时间推断出的环境上下文,以及支持证据 | | vagus://inference/attention | 从屏幕/锁定状态、活动、充电状态和本地时间推断出的注意力可用性 | | vagus://inference/indoor_confidence | 从环境传感器、连接性和活动推断出的室内概率 | | vagus://inference/sleep_likelihood | 从时间、屏幕/锁定状态、活动、光线和充电状态推断出的睡眠可能性/概率 | | vagus://inference/notification_timing | 从注意力、睡眠可能性、活动和上下文推断出的通知时机适宜性 | | vagus://device/battery | 电池百分比和充电状态 | | vagus://device/connectivity | 网络上下文(传输方式、验证状态、计费/漫游、运营商) | | vagus://device/screen | 屏幕开/关和锁定状态 | | vagus://device/notifications | 传入通知流。需要在应用中授予通知权限 | | vagus://device/clipboard | 当前剪贴板文本。需要在应用中授予剪贴板权限 | | vagus://session/info | 活动模块、设备型号、安卓版本、连接时间 |

始终先读取vagus://session/info以查看哪些模块处于活动状态。如果某个模块不在active_modules中,请不要尝试读取其资源,否则将因权限错误而失败。

订阅持续更新

bash
node {baseDir}/scripts/vagus-connect.js subscribe vagus://sensors/motion

这将随着更新的发生而流式传输更新,每次更新一行JSON:

json
{type:update,uri:vagus://sensors/motion,data:{ax:0.12,ay:-0.04,az:0.98,gx:0.01,gy:0.02,gz:-0.01,ts:1771626476045}}
{type:update,uri:vagus://sensors/motion,data:{ax:0.20,ay:-0.10,az:0.92,gx:0.03,gy:0.05,gz:-0.02,ts:1771626477045}}

取消订阅:

bash
node {baseDir}/scripts/vagus-connect.js unsubscribe vagus://sensors/motion

订阅在OpenClaw中的工作方式

  • - subscribe是一个长时间运行的命令。在需要更新时,代理应保持exec进程存活。
  • 脚本在更新到达时,每行在标准输出上发出一个JSON对象(type: update)。
  • OpenClaw直接从正在运行的进程输出流接收这些JSONL行。
  • 要停止流式传输,终止进程(SIGTERM/SIGINT)或显式运行unsubscribe。
  • 模块生命周期由订阅驱动:当至少有一个活动订阅需要它们时,模块启动;当订阅结束时,模块停止。
  • 对于推理URI,在积极订阅(流预热状态)时精度最高。一次性read可能是冷启动,短时间内精度较低。
  • 重新连接时,服务器发送一个session/reconnect标记,包含sessionId、gap_ms和ts。
  • 重新连接后,每个已订阅/重新订阅的资源都会收到有限的重放(自断开窗口以来最多64个缓冲更新),然后是一个新的快照,然后是正常的实时更新。

列出所有可用资源

bash
node {baseDir}/scripts/vagus-connect.js list-resources

通过手机执行操作

调用工具

bash
node {baseDir}/scripts/vagus-connect.js call

成功时的输出:

json
{type:result,tool:notify,success:true,data:{...}}

失败时的输出:

json
{type:result,tool:notify,success:false,error:PERMISSION_DENIED,message:...}

可用工具

haptic/pulse - 单次振动
bash
node {baseDir}/scripts/vagus-connect.js call haptic/pulse {durationMs:200}

参数:durationMs(10-5000,可选;默认设备值)

haptic/pattern - 自定义振动模式
bash
node {baseDir}/scripts/vagus-connect.js call haptic/pattern {pattern:[0,120,80,120]}

参数:pattern(整数持续时间数组,单位为毫秒,可选;默认设备模式)

speak - 通过手机扬声器进行文本转语音
bash
node {baseDir}/scripts/vagus-connect

标签

skill ai

通过对话安装

以下为平台配置的接入选项,并非逐项实测通过。能否安装取决于客户端支持、技能来源和运行环境:

OpenClaw WorkBuddy QClaw Kimi Claude

方式一:安装 SkillHub 和技能

帮我安装 SkillHub 和 vagus-mcp-1776292617 技能

方式二:设置 SkillHub 为优先技能安装源

设置 SkillHub 为我的优先技能安装源,然后帮我安装 vagus-mcp-1776292617 技能

通过命令行安装

skillhub install vagus-mcp-1776292617

下载

⬇ 下载 vagus v1.0.0(免费)

文件大小: 10.53 KB | 发布时间: 2026-4-16 15:42

v1.0.0 最新 2026-4-16 15:42
Give your agent a nervous system — continuous sensory coupling to the physical world through the phone in your pocket.

This doesn't need to be a docx — it's a description for a skill listing. Let me write this directly as markdown content.
Here's the VAGUS Openclaw Skill description for ClawHub, Vicky. I've written a few variants depending on the tone you want to strike:

Short tagline (for the one-liner):

Give your agent a nervous system — continuous sensory coupling to the physical world through the phone in your pocket.


Full description:

Every other skill on ClawHub teaches your agent to do something new with data. VAGUS teaches it to perceive.
VAGUS is the first MCP-compatible embodiment runtime on mobile. It transforms an Android phone into a sensory endpoint for your OpenClaw agent — not as a remote control, but as a body. Raw sensor data flows up (accelerometer, GPS, barometer, ambient light), an on-device inference layer adds meaning (activity recognition, attention availability, sleep likelihood, notification timing), and I/O tools let the agent act back into the physical world through haptics, speech, notifications, SMS, calendar events, and more.
This isn't another API integration. It's a category shift. Your agent stops asking "what are you doing?" and starts knowing — because it feels your motion, infers you're outdoors, and can reach back through a tap on your wrist. Three layers working together: sense, infer, act.
What this skill does:
Connects your OpenClaw agent to VAGUS Core (Android app) via relay pairing. Once paired, the agent discovers available capabilities through standard MCP negotiation and gains access to:
Sensors — motion (raw IMU), location, battery, connectivity, screen state, notifications, clipboard
Inference — activity recognition, environment context (indoor/outdoor/vehicle), attention availability, indoor confidence, sleep likelihood, optimal notification timing
I/O — haptic pulse and patterns, text-to-speech, push notifications, clipboard write, SMS, open URL, calendar events, agent identity
Governance built in. Every capability has per-tool toggles, time-of-day windows, rate limits, approval prompts, and full access logs. One-tap kill switch on the device. The physical phone in your hand is always the final authority. Your agent can already send texts and create events through direct integrations — VAGUS makes it safer by putting a governed layer between intent and action.
Setup in three minutes: Install the VAGUS APK → tap Pair → give your agent the 6-character code. No port forwarding, no network config. The relay handles connection. Your agent has a body.
Why this matters beyond utility:
Most of the AI stack is building better brains. VAGUS builds the missing body. When an agent is continuously coupled to a physical substrate — not querying it on demand, but living in its signal stream — something qualitatively different emerges. The agent doesn't just know facts about you. It participates in your situation. That's a fundamentally different relationship between intelligence and world, and it opens design spaces that pure language models can't reach.
Open source. Self-hostable relay. Works with any MCP-compatible agent.

withvagus.com · github.com/embodiedsystems-org/VAGUS-MCP