Local Power Automate MCP Server

A self-hosted Model Context Protocol (MCP) server that lets an AI assistant (GitHub Copilot CLI, etc.) inspect and manage your Microsoft Power Automate cloud flows — and the Dataverse records behind them — directly from your own machine, using your own identity.

• ✅ Runs locally as a stdio subprocess of the MCP host (e.g. Copilot CLI).

• ✅ Authenticates with your az login session (Azure CLI credential), falling back to an interactive browser sign-in.

• ✅ No third-party SaaS — no flow data ever leaves your device. The process talks straight to https://api.flow.microsoft.com and your environment's Dataverse Web API over HTTPS.

• ✅ Safe by default — every write to Dataverse requires an explicit confirm=True; without it you get a dry-run preview and nothing changes.

It was built as a privacy-safe alternative to hosted offerings that route your tenant's flow data through an external vendor's cloud.

Table of contents

• How it works

• Prerequisites

• Setup

• Integrate with your MCP host — Copilot CLI, VS Code, Claude

• Verify it works

• Tool reference — what each of the 11 tools is for

• Safety model

• Configuration (env vars)

• Troubleshooting

• Uninstall

• License

How it works

The server exposes a set of tools to your MCP host. When the assistant calls a tool, the server:

1. Acquires a bearer token for the right API using your local Azure CLI login (tokens are cached per scope and refreshed automatically).

2. Calls one of two Microsoft APIs:

   • the Power Automate REST API (https://api.flow.microsoft.com) for listing environments, flows, runs, and turning flows on/off; or

   • the environment's Dataverse Web API (https://<org>.crm.dynamics.com/api/data/v9.2) for reading/updating environment variables and raw flow definitions.

3. Returns compact JSON (truncated if it would overflow the model's context).

Nothing runs at import time, so the server starts instantly and only prompts for auth on the first tool call.

Prerequisites

• Python 3.10+ on your PATH (python --version).

• Azure CLI installed and signed in — install, then run az login once.

• Python packages (installed in Setup): mcp, azure-identity, requests.

• An MCP host that launches stdio servers (e.g. GitHub Copilot CLI).

Setup

1. Get the code

git clone https://github.com/prasadgd9022/power-automate-mcp.git
cd power-automate-mcp

2. Install dependencies

python -m pip install -r requirements.txt

3. Sign in to Azure

az login

The server requests a token for https://service.flow.microsoft.com/ (Flow API) and, for the Dataverse skills, an org-scoped token for your environment — both minted from this same session.

4. Register the server with your MCP host

Pick your host below and follow the matching steps in Integrate with your MCP host. In every case you point the host at the absolute path to server.py and run it over stdio.

There is also a SETUP.md written for an AI agent to perform this registration for you automatically in the Copilot CLI.

Integrate with your MCP host

This server is a standard stdio MCP server, so any MCP-capable client can run it. The launch command is always the same:

python <ABSOLUTE_PATH_TO>/power-automate-mcp/server.py --transport stdio

Replace <ABSOLUTE_PATH_TO> with the folder where you cloned the repo, and use whichever interpreter is on your PATH (python, py -3, or python3). On Windows, escape backslashes in JSON — e.g. "C:\\Users\\you\\power-automate-mcp\\server.py". In every client, if the config file already exists, merge the power-automate entry into the existing servers object rather than overwriting the file.

GitHub Copilot CLI

Config file:

• Windows: %USERPROFILE%\.copilot\mcp-config.json

• macOS / Linux: $HOME/.copilot/mcp-config.json

{
  "mcpServers": {
    "power-automate": {
      "type": "stdio",
      "tools": ["*"],
      "command": "python",
      "args": ["<ABSOLUTE_PATH_TO>/power-automate-mcp/server.py", "--transport", "stdio"]
    }
  }
}

Then restart the Copilot CLI so it picks up the new server. Verify with /mcp inside the CLI, which should list power-automate and its 11 tools.

VS Code (Copilot Chat — Agent mode)

VS Code discovers MCP servers from a workspace file at .vscode/mcp.json (or your user settings.json under "mcp"). Note the key is servers, not mcpServers:

// .vscode/mcp.json
{
  "servers": {
    "power-automate": {
      "type": "stdio",
      "command": "python",
      "args": ["<ABSOLUTE_PATH_TO>/power-automate-mcp/server.py", "--transport", "stdio"]
    }
  }
}

Save the file, then click Start on the server shown above the block (or run MCP: List Servers from the Command Palette). Open Copilot Chat, switch to Agent mode, and the 11 tools appear in the tools picker. Requires a recent VS Code with MCP support enabled.

Claude Desktop

Edit Claude Desktop's config (Settings → Developer → Edit Config, or open it directly):

• Windows: %APPDATA%\Claude\claude_desktop_config.json

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "power-automate": {
      "command": "python",
      "args": ["<ABSOLUTE_PATH_TO>/power-automate-mcp/server.py", "--transport", "stdio"]
    }
  }
}

Fully quit and reopen Claude Desktop (not just close the window). The server appears under the tools/plugin (🔌) icon in the chat box. If it fails to load, use an absolute interpreter path (e.g. C:\\Python312\\python.exe or the output of which python3) as command.

Claude Code (CLI)

Register it with one command (run from anywhere):

claude mcp add power-automate -- python <ABSOLUTE_PATH_TO>/power-automate-mcp/server.py --transport stdio

Check it registered with claude mcp list, then use /mcp inside a claude session to confirm the tools are available.

Verify it works

From the repo folder:

python probe.py

Expected output (auth prompt may appear on first run):

TOOLS: ['check_auth', 'list_environments', 'list_flows', 'get_flow', 'get_flow_runs', 'get_flow_run_details', 'set_flow_state', 'get_environment_variable', 'update_environment_variable', 'get_flow_definition', 'update_flow_definition']
CHECK_AUTH_OK: { ... "authenticated": true, "environmentCount": N ... }

If TOOLS prints but auth fails, the server is registered correctly and the issue is just sign-in — run az login again.

Tool reference

11 tools. 8 are read-only; 3 write. All environment arguments are optional — omit them to use your default environment.

Read-only tools

Write tools

Example prompts

Once registered, you can ask your assistant things like:

• "List my Power Automate environments."

• "Show the last 10 runs for flow <flow-id> and tell me why the failures failed."

• "Read the environment variable contoso_ApiBaseUrl in my dev environment."

• "Repoint contoso_ApiBaseUrl to https://new-endpoint/... — preview first, then apply." (dry-run, then confirm=True)

• "Disable flow <flow-id>."

Safety model

The three write tools are designed to be hard to misuse:

• update_environment_variable and update_flow_definition require confirm=True. Called with confirm=False (the default) they return a dry-run preview (old value → new value, or a validity check) and change nothing.

• update_environment_variable only touches the per-environment current value. The shared default value — and therefore other environments such as Production — are left untouched. This makes it the preferred way to repoint a URL or connection string for a single environment.

• update_flow_definition is the advanced escape hatch. It fully overwrites the flow's clientdata (definition + connection references), so the string you pass must be complete and valid. Recommended flow:

  1. get_flow_definition and keep the returned clientdata,

  2. make your minimal edit to that JSON,

  3. call with confirm=False to preview / validate JSON,

  4. call with confirm=True to apply.

• set_flow_state simply starts/stops a flow (no dry-run — enabling/disabling is easily reversible).

Configuration (env vars)

All optional — the defaults work for standard commercial Microsoft 365 tenants.

Troubleshooting

• Could not acquire a token / auth errors — run az login, then retry. On Windows you can warm the cache with az account get-access-token --scope https://service.flow.microsoft.com/.default.

• list_flows returns 0 flows — the Flow REST API only lists flows you own or are shared with. Flows owned by a service account or another user won't appear; reach them directly by id with get_flow / get_flow_definition.

• Dataverse tool says the environment has no linked instance — the write skills require a Dataverse-backed environment (most standard environments are).

• 403 Forbidden from Dataverse — your account lacks permission on that environment's Dataverse org; this is expected for environments you can't administer.

Uninstall

Remove the power-automate block from your MCP host's config (~/.copilot/mcp-config.json for Copilot CLI), restart the host, and delete this folder.

License

MIT.