MCP Power Automate

Local MCP server, Chromium extension, and optional Codex skill for operating Microsoft Power Automate flows from a local browser-backed session.

No Microsoft Entra ID app registration, admin consent, or custom enterprise app setup is required. This MCP uses your existing logged-in browser session instead, which is often much easier to adopt inside locked-down companies.

Built for MCP users who want to do real Power Automate work without getting stuck in enterprise auth setup, tab confusion, or blind edits.

This project lets an agent:

• list flows from the current environment, including owned and shared-visible flows
• select an explicit target flow and an explicit work tab instead of following whichever tab last emitted capture data
• read, validate, update, create, clone, test, inspect runs, and revert flows
• use the browser only for auth and visual context while the MCP owns the workflow logic

The server is MCP-client agnostic.

• Use the MCP from Codex, Claude clients, Gemini clients, OpenCode, or any other MCP client that supports local stdio servers.
• The bundled power-automate-mcp skill in this repository is Codex-specific helper guidance, not a requirement for the MCP itself.

Why MCP users install this

• No Microsoft Entra ID app registration just to get started
• Explicit target flow locking instead of following whatever tab happens to be active
• Multiple Power Automate tabs can stay open while one explicit work tab drives the effective browser-backed session
• Browser-backed auth that matches how teams already work in Power Automate
• Structured review diff plus one-step rollback after saves
• Local-first architecture that keeps the browser for context while the MCP handles the workflow logic

Why this is useful

Install in 60 seconds

Install the optional Codex skill from GitHub:

npx skills add kaael1/mcp-power-automate --skill power-automate-mcp

Register the MCP locally:

codex mcp add power-automate-local -- npx -y @kaael1/mcp-power-automate

Load the extension:

1. Open chrome://extensions or edge://extensions
2. Enable Developer Mode
3. If you are working from a clone, run bun install and bun run build
4. Click Load unpacked
5. Select the dist/extension folder

Open Power Automate, refresh any flow page, and ask your MCP client to:

1. list_flows
2. select_flow
3. get_flow

In practice, most users start with three wins:

1. inspect a real flow without hand-copying JSON
2. make a focused change and validate it
3. review or revert the last save with confidence

If this saves you time, star the repo:

• https://github.com/kaael1/mcp-power-automate

Public package and registry links

• GitHub: https://github.com/kaael1/mcp-power-automate
• npm: https://www.npmjs.com/package/@kaael1/mcp-power-automate
• Official MCP Registry: https://registry.modelcontextprotocol.io/v0/servers?search=io.github.kaael1/mcp-power-automate
• skills.sh discovery: https://skills.sh/

Fast pitch

If you already use MCP clients and want something genuinely useful for enterprise workflows, this repo gives you a local Power Automate operator with browser-backed auth, explicit targeting, safer edits, and real rollback context.

How it works

The project has three pieces:

• server/ TypeScript source for the local MCP server plus an HTTP bridge on 127.0.0.1:17373
• extension/ TypeScript source for the Chromium extension that captures browser auth, tab flow context, token audit, and snapshots
• dist/ Compiled runtime artifacts used for local loading and npm publishing
• skills/power-automate-mcp/ Optional Codex-specific skill instructions that teach the agent to work safely with explicit flow targeting

Why this is local

This project is intentionally local-first.

• Your browser session remains the auth source
• You do not need to provision a Microsoft Entra ID app just to use the MCP
• The extension captures the environment and tab context you already trust
• The MCP runs on your machine and talks to Power Automate with your browser-backed session
• The selected target flow is explicit, so the agent does not blindly follow whichever tab is active

This gives you a good tradeoff:

• much more agent autonomy than manual clicking
• much less infrastructure and auth complexity than a remote SaaS MCP

Current capabilities

• Read the selected target flow
• List flows from the current environment
• Merge owned, shared-user, and portal-shared flows into one catalog
• Lock the MCP onto an explicit active target flow
• Preview changes before saving them
• Validate flows with legacy validation endpoints
• Update a flow and keep one-step rollback history
• Create a blank request or recurrence flow
• Clone an existing flow
• List runs, inspect action-level results, and wait for completion
• Get a callback URL and invoke manual/request flows

Who this is for

• MCP power users looking for high-leverage local tooling
• Power Automate makers who want AI help without rebuilding auth from scratch
• Teams who need safer automation edits than “let the agent click around and hope”

Install options

1. Skill install from GitHub

npx skills add kaael1/mcp-power-automate --skill power-automate-mcp

You can verify discovery with:

npx skills add kaael1/mcp-power-automate --list

2. MCP install from npm

Use this in Codex-compatible config:

{
  "mcpServers": {
    "power-automate-local": {
      "command": "npx",
      "args": ["-y", "@kaael1/mcp-power-automate"]
    }
  }
}

3. MCP install from local clone

bun install
bun run build
codex mcp add power-automate-local -- node /path/to/mcp-power-automate/dist/server/index.js

Development checks

bun run typecheck
bun run lint
bun run test
bun run build
bun run check

Typical workflow

1. Start the local bridge and MCP.
2. Open any flow in the target Power Automate environment.
3. Let the extension capture auth and environment context.
4. Ask Codex to list_flows.
5. If you have multiple Power Automate tabs open, use list_captured_tabs and select_work_tab to choose the work tab first.
6. Ask Codex to select_flow for the flow you actually want.
7. Continue with get_context, get_flow, preview_flow_update, validate_flow, apply_flow_update, list_runs, get_run, wait_for_run, or invoke_trigger.

The popup and side panel help you:

• use the popup as a quick launcher for the selected target, the top alert, and the safest next action
• use the side panel as the main workspace with Today, Flows, Review, and System sections
• review the current browser flow and keep the assistant synced with it
• refresh run status
• review the last update with a structured visual diff
• revert the last saved change
• browse pinned, recent, and current-environment flows without leaving the workspace

list_flows returns an accessScope hint per flow:

• owned
• shared-user
• portal-shared

Available MCP tools

• get_context
• list_captured_tabs
• select_work_tab
• get_status
• get_health
• list_flows
• refresh_flows
• select_flow
• select_tab_flow
• set_active_flow
• set_active_flow_from_tab
• get_active_flow
• get_flow
• preview_flow_update
• apply_flow_update
• update_flow
• create_flow
• clone_flow
• validate_flow
• get_last_update
• revert_last_update
• list_runs
• get_latest_run
• get_run
• get_run_actions
• wait_for_run
• get_last_run
• get_trigger_callback_url
• invoke_trigger

Recommended v1 workflow:

• get_context
• list_captured_tabs
• select_work_tab
• list_flows
• select_flow
• get_flow
• preview_flow_update
• validate_flow
• apply_flow_update

Legacy aliases such as get_status, set_active_flow, and update_flow remain available for compatibility.

Available resources

• power-automate://status
• power-automate://context
• power-automate://last-run
• power-automate://active-flow

Security model

This MCP is not a remote cloud service.

• It depends on a logged-in Chromium session
• It uses browser-backed tokens captured locally by the extension
• It is best suited for trusted local use
• It is strong enough for supervised real work, but still benefits from change discipline on production flows

Current limitations

• The browser still provides auth and current-tab context
• Rollback is currently one step only
• The review UI is optimized for confidence and inspection, not in-place diff editing
• The current architecture is local-first, not remote SaaS-style
• Critical production use is still best done with supervision

Publishing

This repo is already prepared for:

• GitHub discovery
• skills.sh / Vercel-style skill installation
• npm distribution
• Official MCP Registry publishing

Relevant files:

• package.json
• server.json
• PUBLISHING.md
• CHANGELOG.md

Search and install phrases

People should be able to find this repo when searching for:

• Power Automate MCP
• Microsoft Power Automate MCP server
• local MCP for Power Automate
• browser-backed auth MCP
• Codex skill for Power Automate

Client compatibility

The MCP server itself is designed around standard local stdio transport, so it is not tied to Codex.

• Codex: supported directly, and the bundled skill is written for Codex.
• Claude clients with MCP support: should work with the MCP server directly.
• Gemini clients with MCP support: should work with the MCP server directly.
• Other MCP clients such as OpenCode: should work as long as they support local stdio MCP servers.

What is client-specific:

• the skills/power-automate-mcp skill is Codex-oriented guidance
• the browser extension UX is generic, but any prompt helpers or slash-command conventions depend on the client

The safe rule of thumb is:

• if the client can launch a local stdio MCP server, it can use this MCP server
• if the client understands Codex skills, it can also use the bundled skill directly

Repo guidance

See AGENTS.md for local repository guidance for agents and contributors.

License

MIT. See LICENSE.