YindDao AI Power - RPA MCP Server (Python)

This project is a Python reimplementation of the original aipower-rpa-mcp-server TypeScript project. It keeps the same operating modes and tool names as the original README: Local Mode exposes queryRobotParam, queryApplist, and runApp; Open API Mode exposes uploadFile, queryRobotParam, queryApplist, startJob, queryJob, and queryClientList. It also keeps the same startup semantics: default stdio, and --server for SSE. fileciteturn0file0

Feature matrix

Transport layer

• stdio mode: default startup mode. Suitable for OpenClaw / MCP host local process integration.
• SSE mode: start with --server. Default endpoints:
  • GET /sse
  • POST /messages?sessionId=...

Runtime modes

Mode	Required env	Main use
local	SHADOWBOT_PATH, USER_FOLDER	Run local YindDao apps through the Windows/macOS desktop client
openApi	ACCESS_KEY_ID, ACCESS_KEY_SECRET	Use YindDao enterprise open API to start/query jobs remotely

Installation

python -m venv .venv
source .venv/bin/activate
pip install -e .
cp .env.example .env

Environment variables

Local Mode

RPA_MODEL=local
SHADOWBOT_PATH={your_shadowbot_path}
USER_FOLDER={your_user_folder}
LANGUAGE=zh
SERVER_PORT=3000

Notes:

• SHADOWBOT_PATH should point to the YindDao executable.
• USER_FOLDER should point to the current logged-in ShadowBot/YindDao user directory, not the parent users directory.
• For WSL calling Windows YindDao, use WSL paths such as:
  • SHADOWBOT_PATH=/mnt/c/Program Files/ShadowBot/ShadowBot.exe
  • USER_FOLDER=/mnt/c/Users/<WindowsUser>/AppData/Local/ShadowBot/users/<CurrentUserId>

Open API Mode

RPA_MODEL=openApi
ACCESS_KEY_ID={your_access_key_id}
ACCESS_KEY_SECRET={your_access_key_secret}
LANGUAGE=zh
SERVER_PORT=3000

Startup

stdio

aipower-rpa-mcp-server

SSE

aipower-rpa-mcp-server --server

Tool reference

1. Local Mode tools

queryApplist

Purpose

• Enumerate locally published YindDao apps under USER_FOLDER/apps.

Mode

• local

Input

• This tool keeps the upstream-compatible query shape, but local mode primarily ignores pagination/search fields and scans the local apps directory.

Output

• Returns an array of apps, each typically containing:
  • uuid
  • name
  • description

Behavior

• Only published app packages are returned.
• Typical discovery path is USER_FOLDER/apps/*_Release/xbot_robot/package.json.
• Only entries whose package metadata indicates app type are included.

---

queryRobotParam

Purpose

• Read a local app's input parameter definition.

Mode

• local

Input

• robotUuid: target app UUID.

Output

• Returns the parameter definition list from the app's local metadata / flow definition.

Behavior

• The Python port improves the upstream local implementation by actually locating the target app by robotUuid before reading parameter metadata.
• If the app cannot be found or has no parameter definition, an empty list is returned.

---

runApp

Purpose

• Launch a local YindDao app.

Mode

• local

Input

• appUuid: target app UUID.
• appParams: upstream-compatible field retained for interface compatibility.

Output

• Returns a launch result indicating that the local client launch command has been issued.

Behavior

• Current behavior mirrors the upstream local project: the server launches the desktop client via a shadowbot:Run?robot-uuid=... URI style invocation.
• This is a launch action, not a full job-tracking API.

Current capability boundary

• In local mode, this project can start an app, but it does not provide built-in tools for:
  • querying live execution logs
  • retrieving step-by-step run records
  • waiting for completion and returning a final success/failure result
  • querying a local task id / run history produced by the desktop client
• If you need result backflow in local mode, you should add your own app-side output mechanism, such as writing JSON/log files or posting back to your own service.

2. Open API Mode tools

uploadFile

Purpose

• Upload a file for downstream job usage.

Mode

• openApi

Input

• file: file content
• fileName: target upload filename

Validation / limits

• Max filename length: 100
• Allowed extensions: txt, csv, xlsx
• Upload request body limit: 10 MB

Output

• Returns the Open API upload response.

---

queryRobotParam

Purpose

• Query robot/app parameter definitions from YindDao Open API.

Mode

• openApi

Input

• robotUuid (optional)
• accurateRobotName (optional)

Output

• Returns the parameter definition payload from the API.

---

queryApplist

Purpose

• Query app list from YindDao Open API.

Mode

• openApi

Input

• appId (optional)
• size (optional)
• page (optional)
• ownerUserSearchKey (optional)
• appName (optional)

Output

• Returns the API app list response.

---

startJob

Purpose

• Start a remote job through YindDao Open API.

Mode

• openApi

Input

• robotUuid (required)
• accountName (optional)
• robotClientGroupUuid (optional)
• waitTimeoutSeconds (optional)
• runTimeout (optional)
• params (optional key/value object; transformed to the API parameter array format)

Validation

• robotUuid is required.
• accountName and robotClientGroupUuid are mutually exclusive.
• waitTimeoutSeconds and runTimeout must be in 60 ~ 950400.
• Serialized params length must not exceed 8000.

Output

• Returns the Open API start-job response, typically including job-level identifiers used for later querying.

---

queryJob

Purpose

• Query a started Open API job by jobUuid.

Mode

• openApi

Input

• jobUuid (required)

Output

• Returns job status information, typically including job state and result payload fields returned by the API.

Current capability boundary

• queryJob is suitable for polling job state and reading summary/result fields returned by YindDao Open API.
• It is not a full real-time execution log stream.
• Depending on the upstream API payload, you may get status, remark, timestamps, inputs/outputs, screenshot URL, client info, etc., but not full step-by-step desktop execution logs.

---

queryClientList

Purpose

• Query available robot client machines.

Mode

• openApi

Input

• status (optional)
• key (optional)
• robotClientGroupUuid (optional)
• page (required)
• size (required)

Output

• Returns the client list response from Open API.

Logging / result-query capability summary

Local mode

Supported:

• enumerate local apps
• read local parameter definitions
• launch local apps

Not built in:

• query running status
• query execution logs
• query final result / output from a local launch
• query local execution history as MCP tools

Open API mode

Supported:

• start remote jobs
• query job status / result summary
• query client list
• upload files

Not built in:

• real-time full execution log streaming
• full desktop-level step trace playback

OpenClaw example

WSL + Windows YindDao

openclaw mcp set yingdao-rpa '{
  "type": "stdio",
  "command": "/home/your-user/path/to/.venv/bin/python",
  "args": ["-m", "aipower_rpa_mcp_server"],
  "env": {
    "RPA_MODEL": "local",
    "SHADOWBOT_PATH": "/mnt/c/Program Files/ShadowBot/ShadowBot.exe",
    "USER_FOLDER": "/mnt/c/Users/<WindowsUser>/AppData/Local/ShadowBot/users/<CurrentUserId>",
    "LANGUAGE": "zh"
  }
}'

Then restart the gateway and test:

调用 yingdao-rpa 的 queryApplist，返回当前影刀应用列表

Troubleshooting

stdio server looks "stuck"

This is expected. In stdio mode the process waits for MCP JSON-RPC messages from the host. Do not press Enter in that terminal; sending a blank line will cause JSON parsing errors.

Missing SHADOWBOT_PATH or USER_FOLDER

You started local mode without the required environment variables.

spawn ... ENOENT

The executable path in your MCP config is wrong. Check:

• the Python interpreter path in command
• SHADOWBOT_PATH
• whether the WSL path actually exists

queryApplist returns empty

Check:

• USER_FOLDER points to the current logged-in YindDao user directory
• USER_FOLDER/apps exists
• the target flow has actually been published as an app

Notes

The original minimal README only documented the mode split, tool list, environment variables, startup options, and a brief OpenClaw example. This updated README adds detailed tool-by-tool behavior descriptions, capability boundaries, and log/result-query expectations. fileciteturn0file0