Microsoft Power Apps MCP Server: Low-Code AI Agents for the Rest of Your Company

Microsoft's April 2026 Power Apps update shipped three features that quietly change the low-code agentic landscape: deeper Microsoft 365 Copilot integration, a native MCP server inside Power Apps, and a human approval feed that pauses agents for review on sensitive actions. This is Microsoft's answer to "how do business users — not developers — build AI agents safely?"

This guide walks through what the MCP server does, how the approval feed closes the governance gap, and a practical build pattern for delivering an agentic Power App in two afternoons.

Why This Matters

The agentic AI story through 2025 was developer-centric: Claude Code, Cursor, custom Python orchestrations, MCP servers written in TypeScript. The audience was people who could pip install.

Microsoft's bet is that the next 10x of agentic adoption comes from the 20 million people building in Power Apps, Power Automate, and Copilot Studio today. Those builders do not write code. They configure. If agentic patterns can be configured — with the same safety as hand-coded ones — the Power Platform becomes the largest agent-building surface in the world overnight.

That requires two things Microsoft just shipped:

1. A way for Power Apps to act as MCP clients (and be MCP servers themselves)
2. A safety net — human approval on sensitive actions — that does not require the builder to implement approval flows from scratch

The Power Apps MCP Server

Power Apps now exposes every Dataverse entity, every Power Platform connector, and every Power Automate flow as MCP tools. Your AI agent (Claude, Copilot, Gemini, custom) can:

• Read and write Dataverse records through MCP calls
• Invoke any of the 1,100+ Power Platform connectors (Salesforce, ServiceNow, SharePoint, Excel Online, SAP, Oracle, etc.)
• Trigger Power Automate flows and receive their structured output
• Read Canvas app state and write to app variables
• Submit items to the Power Apps human approval feed

Conversely, Power Apps can act as an MCP client, meaning a Power App can call external MCP servers as part of its logic. This two-way integration is the new thing.

Deployment

You enable the MCP server on a Power Apps environment from the admin portal. This provisions:

• A unique MCP endpoint URL for the environment
• Authentication through Microsoft Entra service principals
• Default deny-all policy that you explicitly open up per tool

Connecting Claude Code to your Power Apps MCP server looks like:

{
  "mcpServers": {
    "power-apps": {
      "type": "http",
      "url": "https://contoso.crm.dynamics.com/api/mcp/v1",
      "auth": {
        "type": "oauth2",
        "tenant_id": "${ENTRA_TENANT}",
        "client_id": "${AGENT_CLIENT_ID}",
        "scope": "https://contoso.crm.dynamics.com/.default"
      }
    }
  }
}

Once connected, tools like dataverse_read, dataverse_create, invoke_connector_salesforce, and run_flow_notify_team appear in your agent's toolset.

The Human Approval Feed

This is the underappreciated piece of the April release. The approval feed is a new app inside Teams that collects action items from any agent, app, or flow that wants human approval. The agent marks an action as requiring approval, the approval lands in the assignee's Teams feed with full context, and the agent pauses until the approval comes back.

Structurally:

Agent → requests approval →
  Teams approval feed (with context, requested action, risk level) →
  Assignee clicks approve/deny/modify →
  Response routed back to agent →
  Agent continues (or aborts)

The approval payload can include the proposed email draft, the Salesforce record to be updated, the Power Automate flow to be triggered, and a "why" explanation from the agent. The assignee can approve as-is, modify the draft, or deny with a reason that the agent sees.

For teams that have been stuck on the "how do we let agents take action without letting them take the wrong action" problem, this is the missing primitive.

Build Pattern: Customer Outreach Agent in Two Afternoons

Here is a real pattern that multiple teams have shipped since the feature went live.

Goal: An AI agent that runs nightly, reviews CRM opportunities that have been idle for more than 14 days, drafts a personalized follow-up email for each, and submits the drafts to the account owner for approval.

The smart buy

Why pay $228/year when $69 works?

Lifetime Starter: one payment, no renewals. Covered by 30-day money-back guarantee.

See the math

Afternoon 1: Wire the data and logic.

1. In Power Apps, enable the MCP server on the Sales environment.
2. Grant the agent service principal read/write access to opportunity and contact tables.
3. Create a Power Automate flow draft_followup_email that accepts an opportunity ID and a draft body, and submits them to the Teams approval feed.
4. Confirm the flow works end-to-end by manually triggering it.

Afternoon 2: Wire the agent.

1. Configure your agent (Claude Managed Agents, Copilot Studio, or custom) with the Power Apps MCP server.
2. Write the agent prompt:

You are a sales operations assistant. Every night, review all open
opportunities where last_activity_date is older than 14 days. For each,
read the opportunity context, the primary contact, and the last 5
interactions. Draft a personalized follow-up email that references
specific details from the interactions. Submit each draft to the
opportunity owner via the draft_followup_email flow.

Do not send any email directly. Every outbound message requires owner
approval via the Teams approval feed.

3. Set a cron trigger for 2 AM local time.
4. Deploy and watch the first run in observability.

Day 3 is tuning: which details the agent picks up on, how long the drafts are, how often the owner modifies vs approves as-is. Most teams converge on a stable agent within one week.

Where the Governance Gap Still Sits

Power Apps + MCP + approval feed closes a big governance gap, but three edges still need attention.

1. Least-privilege is manual.

The Power Apps MCP server gives you tool-level allow/deny but not row-level or field-level filtering. If your agent can read opportunity it can read every opportunity. Row-level security is enforceable through Dataverse security roles, but you have to configure them. Teams tend to underconfigure this.

2. Cost observability is not native.

Power Apps does not itself track LLM costs. You need to pair with a governance layer (Unity AI Gateway, LangSmith, or a custom wrapper) if you want per-workflow cost attribution.

3. Cross-environment agents need thought.

An agent that reads from Sales Dataverse, writes to Support Dataverse, and triggers flows in Marketing Dataverse touches three MCP servers. Authentication, audit, and error handling get more complex. Microsoft's guidance is to build single-environment agents where possible and federate through Power Automate flows when you need cross-environment action.

How It Compares to Copilot Studio

Copilot Studio is Microsoft's higher-level agent builder aimed at non-technical users. It sits on top of the same Power Platform primitives. The distinction:

• Copilot Studio: Build a conversational AI agent that handles a narrow set of structured tasks. Best for chatbots, FAQ bots, narrow support workflows.
• Power Apps + MCP: Let any MCP-compatible agent (including external agents built in Claude or Gemini) drive Power Platform actions. Best for autonomous, long-running, multi-system workflows.

You can combine them: Copilot Studio for the user-facing chat interface, Power Apps MCP for the actions the agent takes behind the scenes.

What Happens Next

The Power Apps MCP server is available in commercial Power Apps environments today, with government cloud rollout in Q3. The approval feed is available in Teams Premium environments.

Two second-order effects worth watching:

1. SI/consulting practice boom.

Agentic Power Apps is a greenfield consulting market. Partners with Dynamics 365 expertise are pivoting their practices to agentic deployments. If you work in this space or hire in it, capacity is tight through 2026.

2. Agent sprawl arrives at non-technical orgs.

When business users can build agents in an afternoon, you get dozens of them quickly. Governance primitives (Okta agent identities, Unity AI Gateway, approval feeds) are not optional — they are prerequisites. Organizations that deploy the governance stack first scale agent use five to ten times faster without incident.

Microsoft's agentic strategy is clear: make the Power Platform the default place that non-developers build agents, govern it natively, and let developer-grade agents plug in through MCP. The April 2026 release is the first product where that strategy is fully functional.

AI Magicx integrates with Power Apps and Microsoft 365 for teams building agentic workflows across marketing, sales, and operations. Start your account to try it.