Clousd Mcp

io.github.Clousd-Android由 GitHub 账号 Clousd-Android 发布v1.6.0更新于 Nov 4, 2026

Real Android phones in the cloud for AI agents: give phones a task in plain words and get the answer

已验证Streamable HTTP可网页运行效率与工作流云端与基础设施浏览器自动化

概览

AI 生成的概览

Clousd 为助手提供云端真实 Android 手机:用自然语言给手机下任务,或逐步操作屏幕。

可以做什么
  • 查看账户套餐、余额、即将续费和警告
  • 列出、启动、停止和重启账户的 Android 手机
  • 用自然语言给手机下任务,跟踪运行、回答问题或取消
  • 查看手机屏幕,并直接点击、输入或滑动
  • 管理账户 Library 中的文件和自己上传的 APK,并发送或安装到手机
  • 列出已保存的登录信息,并安排定时或按需执行的任务
适用场景
适合需要助手在真实 Android 应用内操作时使用,例如填写表单、查看应用状态、在多台手机上执行同一任务;也适合把重复的手机流程按计划自动执行,而不必使用本地模拟器。
运行要求
可使用托管端点并通过 OAuth 登录,或本地安装 clousd-mcp(Python)并以 CLOUSD_API_KEY 环境变量启动;其他客户端在 Authorization 请求头中发送密钥。需要 Clousd 账户、手机和余额,并能访问 Clousd API。
安装前请注意
密钥(CLOUSD_API_KEY 或 Authorization 请求头)是账户的实时凭证,具有只读或读写权限及指定手机范围。手机按用量计费,运行和定时任务都会产生费用。工具会启动、停止、重启和重置手机,上传文件和 APK、安装应用、发送文件并创建定时任务;重置会清除应用和数据。已保存的登录值在手机上输入时不显示,但账户中存有这些信息。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Clousd Mcp,将其添加到工作区。
  2. 连接 所需的 凭据。它们会被私密存储,绝不会在公开页面上展示。
  3. 为需要使用其工具的对话启用该服务。

Web executable,通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体,或 SourceWeft 对话中。智能体会读取该服务器的安装指南,先向你展示发布者、端点和登录方式,经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/io.github.Clousd-Android%2Fclousd-mcp/install.md,并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前,先告诉我它的发布者、端点和登录方式,等我确认。不要把 API 密钥写进命令或我会提交的文件;问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。 客户端会在首次连接时完成 OAuth 登录。

{
  "mcpServers": {
    "clousd": {
      "type": "http",
      "url": "https://api.clousd.com/mcp"
    }
  }
}

README

Clousd: real Android phones in the cloud, for AI agents

[tests] [PyPI clousd] [PyPI clousd-mcp] [MCP registry] [license]

Every agent framework can tap a phone. Almost none has a phone worth tapping.

Clousd is the phone. Each device is a full Android system on the Android version you pick (12 to 17), with Google Play inside and its own network exit. Give it a task in plain words and Clousd's phone agent does the steps and answers; or drive the screen yourself, step by step. Apps see a phone, not an emulator, so your agent sees the screens a person sees.

This repository is the public side of it: the OpenAPI description, the Python client, the MCP server and the skill for coding agents.

Two ways to use a phone

Give it the goal. POST /v1/runs {"task": "...", "device": "..."}, c.run(...) in Python, the run tool over MCP. Clousd's agent looks at the screen, taps and types until the goal is reached, and finishes with an answer - plus a replay of every step, and JSON in your schema when you ask for it. One call gives the same task to up to 32 phones. Saved logins are typed on the phone without your agent or the model seeing them.

Drive it yourself. observe returns a fresh screenshot with a number on every element; act taps an element, types, swipes, opens apps and links, and waits until the screen settles. Snapshots reset a phone to a saved state in seconds.

Quick start

MCP

Hosted, nothing to install: https://api.clousd.com/mcp. Apps with connectors (Claude.ai, ChatGPT…) sign in with OAuth - just add the URL. Other clients send the key in the header:

json
{  "mcpServers": {    "clousd": { "url": "https://api.clousd.com/mcp", "headers": { "Authorization": "Bearer cl_live_..." } }  }}

Local: pip install clousd-mcp, then {"command": "clousd-mcp", "env": {"CLOUSD_API_KEY": "cl_live_..."}}.

Tools: account, devices, run, run_status, answer, observe, act, credentials, schedules (mcp/).

Coding agents learn the workflow from the skill:

bash
npx skills add https://github.com/Clousd-Android/clousd-sdk --skill clousd

Python

bash
pip install clousdexport CLOUSD_API_KEY=cl_live_...
python
from clousd import Clousd
c = Clousd()run = c.run("Open Settings and tell me the Android version", device="dev_04f2", wait=True)print(run.status, run.answer)                     # completed  Android 15
phone = c.device("dev_04f2")                      # or step by stepphone.open_app("com.android.chrome")phone.tap_text("Search")phone.type("weather in Berlin")phone.key("enter")phone.wait_text("Berlin", timeout=30)

More in python/.

Plain HTTPS

bash
curl -H "Authorization: Bearer cl_live_..." https://api.clousd.com/v1/devicescurl -H "Authorization: Bearer cl_live_..." -H "Content-Type: application/json" \     -d '{"task":"Open Settings and tell me the Android version","device":"dev_04f2"}' https://api.clousd.com/v1/runscurl -N -H "Authorization: Bearer cl_live_..." https://api.clousd.com/v1/runs/r_8f2c/events      # live steps (SSE)curl -H "Authorization: Bearer cl_live_..." "https://api.clousd.com/v1/devices/dev_04f2/observe?w=540"

API in one table

RoutePurpose
POST /runsa task in plain words on device or devices[] (up to 32): {task, device, credentials:[{package,name}], options:{output_schema, mode, max_steps, timeout}}; Idempotency-Key header
GET /runs, GET /runs/{id}runs of the key; status, answer, structured_output, the card it waits on
GET /runs/{id}/eventslive events (Server-Sent Events): status, step, plan, card, answer, end
GET /runs/{id}/trajectory, /steps/{n}.jpgevery step with its screen
POST /runs/{id}/answer, /message, /cancelanswer a question or permission, talk to the running agent, stop it
GET/POST /schedules, POST /schedules/{id}/toggle, /deletetasks at a time on days of the week
GET/POST /credentials, POST /credentials/{id}/fields, DELETE …saved logins; values are never returned
GET/POST /webhooks, …/{id}/rotate, /toggle, /test, GET …/deliveriesevents to your server, signed with your secret
GET /android, GET /devices, GET /devices/{name}, POST /devices {android}, start/stop/restart/resetphones of the key and their lifecycle; a phone is created by Android version; reset wipes apps and data
DELETE /devices/{name} (?forever=1), GET /trash, POST /trash/{name}/restoredelete moves the phone to Trash for 14 days (stopped, not billed, restorable); forever=1 wipes it at once
GET /devices/{name}/observe, POST /devices/{name}/act, /screenshot, /screenrecordthe screen, step by step
GET/POST /devices/{name}/network, GET/POST /proxiesthe phone's network exit
GET/POST/DELETE /devices/{name}/adbADB over a relay: one-time code, IP allow-list, expiry
GET /balanceplan, phones, money left and warnings to show the user

Keys have a scope (read or read & write) and an explicit list of phones; keys travel only in the Authorization header. Limits: 120 requests a minute per key and device, 600 a minute per key across devices, 60 a minute for everything else; 429 comes with Retry-After. Errors are JSON: {"error": "short_code", "message": "what happened"}. Full description: openapi.yaml; reference: clousd.com/docs/reference.

What is in this repository

PartWhat it is
openapi.yamlOpenAPI 3.1 description of the public API v1 (https://api.clousd.com/v1)
python/clousd: Python client (runs, schedules, logins, webhooks, devices, observe/act, reset), clousd.agent (a step-by-step loop on your own model), clousd.bench, the clousd CLI
mcp/clousd-mcp: the MCP server, local or hosted
skills/clousdskill for coding agents: how to work with Clousd phones
docs/agent-loop.mddriving a phone yourself: the observe → act → verify loop
server.json, smithery.yamlregistry metadata

Upgrading from 1.5

1.6 adds your apps. An app is an APK you upload once and install by id on any of your phones; there is no catalog of other people's apps (Google Play apps are installed in the Play Store on the phone). GET /v1/apps lists your APKs (id, name, size, package, version), POST /v1/apps uploads one as a file up to 300 MB, POST /v1/apps/{id}/install {devices} installs it, DELETE /v1/apps/{id} removes it from storage. POST /v1/media also takes a multipart file now (photo 25 MB, video 200 MB, document 50 MB, apk 300 MB). Every APK is checked on upload and before each install - a real Android app, 64-bit ARM native code when it has any, not a root or system-modification tool - and refused with 422 apk_rejected and the reason otherwise. SDK: c.apps(), c.upload_app(path), c.install_app(id, devices), c.delete_app(id); c.upload_media(path) sends the file as a file. MCP: the apps tool.

Upgrading from 1.4

1.5 creates phones by Android version only. POST /v1/devices {android: "14"} (versions: GET /v1/android); the phone model is picked by the service. GET /v1/models answers 410, and devices no longer carry a model field. A phone not renamed is called Phone #N. An old body with profile still works and keeps only that model's Android version. Python: c.create_device(android="14"), c.android_versions(); c.models() is deprecated.

Upgrading from 1.2

1.3 adds the Library: GET /v1/media lists the account's files (photos, videos, documents, apks) with ids and folders, POST /v1/media uploads one, POST /v1/media/send copies files to phones. Pass ids to POST /v1/runs as files and they are placed on the phone before the first step (photos and videos in the gallery, documents in Downloads, apks installed). SDK: c.run(task, files=[...]), c.media(), c.upload_media(), c.send_media(); MCP: the library tool and files on run. Creating a device no longer takes plan: phones run pay as you go, a plan's minutes are used first.

Upgrading from 1.1

1.2 removes saved device states. /v1/devices/{name}/snapshots, /profiles and /worlds answer 410; to start from a clean phone call POST /v1/devices/{name}/reset (apps and data are wiped, the Android version stays the same). Python: device.reset() replaces snapshots(), save_snapshot(), restore() and clone().

Upgrading from 0.3

1.0 gives phones goals instead of scripts. /v1/recipes, /v1/groups, /v1/accounts, /v1/texts and the device recipe / record routes answer 410 with where to go instead: tasks are POST /v1/runs, logins are /v1/credentials. MCP: run / run_status / answer / credentials / schedules replace fleet, recipe, resources, network, snapshots and inspect. Keys in the URL are no longer accepted.

Building from this repository

bash
pip install -e pythonpip install -e mcppython -m unittest discover -s python/tests     # offline: a fake server and a scripted model, no key needed

Links

License

MIT. See LICENSE.

来源:README.md,提交 35ef0b4

工具

11

account

Account

The account: plan and phones, money left, what is about to renew, and warnings. Check it before starting phones or long work, and tell the user every warning in plain words with the top_up link.

无需输入参数。

devices

Devices

Phones this key can use. op=list: name, label (the name people see in the dashboard, e.g. "Phone #3"), number, state (running / stopped / starting), Android version and network exit. Every tool takes the name, the label or the number. op=start / op=stop / op=restart with device: start a stopped phone (one to three minutes until it has booted), stop one (it keeps its apps and data) or restart it. op=reset with device: back to a fresh state - apps and data wiped, same Android version; works whether the phone runs or is stopped. Ask the user before reset. op=rename with device and label: give a phone a name (empty = back to "Phone #N"). op=delete with device: the phone moves to Trash for 14 days - stopped, not billed, apps and data kept - and op=restore brings it back (stopped). forever=true deletes it with all its data at once and cannot be undone: only when the user says so. op=trash: what is in Trash (name, label, android, deleted_at, purge_at, reason). op=restore with device: back from Trash (the name or label as shown by op=trash).

输入参数oplabeldeviceforever

run

Run a task

Give phones a task. Clousd's agent looks at the screen, taps and types until the goal is reached, and finishes with an answer. Several devices = the same task on each, all at once (one run per phone under one parent). Returns the run with its status; if it is still going, follow it with run_status. A run in mode=ask may stop with waiting_for: reply with answer. Logins: never ask the user for passwords - pass credentials; their values are typed on the phone without you or the model seeing them. A task that needs a file (post this photo, fill this form from this PDF, install this app): upload it once with the library tool and pass its id in files - do not push files by hand each time.

输入参数modetask必填filesdevices必填max_stepscredentialswait_secondsoutput_schema

library

Library

The user's Library on Clousd: every file they want on their phones lives here once (photos, videos, documents, apps) and is used by id. op=list shows it; op=upload adds a local file and returns its id; op=send copies files to phones right away (photos and videos into the gallery as if taken there, documents into Downloads, apks installed). For a task that needs a file, prefer run(files=...) - the file is placed on the phone before the first step.

输入参数opidskindpathdevices

apps

Apps

The user's own apps on Clousd: APKs they uploaded, kept in their private storage. Upload an APK once (op=upload), then install it by id on any of their phones (op=install), as often as needed. Clousd has no catalog of other people's apps - for apps from Google Play, open the Play Store on the phone. Every APK is checked on upload (a real Android app, 64-bit ARM code, no root or system tools) and refused with the reason otherwise. To have an app on the phone before a task, pass its id in run(files=...).

输入参数opapppathdevices

run_status

Run status

Follow a run. op=status: status (queued, running, paused = waits for an answer, completed, failed, cancelled), the answer or error, the last steps, per-device results for a run on several phones, and a replay link for the user. op=list: recent runs of this key. op=cancel: stop it. op=message: tell the running agent something (it reads it before its next step).

输入参数opruntextstepswait_seconds

answer

Answer a run

Reply to a run that waits for you: answer its question with text, or allow / decline the step it asks permission for. If the question is something only the user knows, ask the user first. The run goes on at once.

输入参数run必填textallow

observe

Observe the screen

Look at the phone yourself: a fresh screenshot, the app on screen and (with_text) every labelled or tappable element with a number. Act on an element by its number (act element=N); coordinates, when needed, are in this image's pixels.

输入参数device必填numberswith_text

act

Act on the phone

Do one thing on the phone yourself, then wait until the screen settles. Prefer element numbers from observe over coordinates; tap_text taps the element whose text contains `text`. Typing goes into the focused field (tap it first) and works in any language. "the screen did not change" means the action probably missed - observe again. Read without touching the screen: installed (apps), notifications (the shade), clipboard, app_info (version, installer, running), crashes (recent app crashes and ANRs), health (running, booted, network exit, last check). For a whole goal, prefer run: it does all the steps for you.

输入参数xymsx2y2textaction必填device必填elementtimeout

credentials

Saved logins

Saved logins of the account: package, name and which fields are set (username, email, password, phone, totp...). Values are never shown. Pass one to run as "package/name"; the agent types it on the phone without seeing it. The user adds logins in the dashboard (Tasks → Sign-in) - do not ask them for passwords in the chat.

输入参数package

schedules

Schedules

Tasks that run by themselves. op=list: id, name, task, phones, at, days, tz, enabled and the runs it started last time. op=add: run `task` on `devices` at `at` on `days`, or every hour (every=hour), or only on demand (trigger=api); each firing starts runs like the run tool. op=run: start it now. op=toggle: pause or resume. op=delete: remove it.

输入参数atoptzdaysnametaskeverydevicestriggerschedulecredentials

版本历史

9
  1. v1.6.0最新Nov 4, 2026
  2. v1.5.0Nov 3, 2026
  3. v1.4.0Nov 1, 2026
  4. v1.3.0Oct 31, 2026
  5. v1.2.0Oct 22, 2026
  6. v1.1.0Oct 21, 2026
  7. v0.3.2Oct 13, 2026
  8. v0.2.0Oct 10, 2026
  9. v0.1.2Oct 9, 2026

评分

已满足适用检查项的 100%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 已列出工具15 分
  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分