Skip to content

Docs

How to Connect Microsoft Copilot Studio to an MCP Server

Microsoft Copilot Studio agents read MCP (Model Context Protocol) servers as tools. The fastest path is the built-in MCP onboarding wizard. Copilot Studio only supports the Streamable transport: it dropped SSE for MCP after August 2025, so the server needs to speak Streamable HTTP.

Looking for GitHub Copilot in VS Code instead? That uses an mcp.json file, not the agent designer. See Connect GitHub Copilot to an MCP server.

What you need

  • A Copilot Studio license and an agent to attach the server to.
  • An MCP server URL that speaks Streamable HTTP (for example, https://hjarni.com/mcp).

The recommended onboarding wizard runs inside Copilot Studio itself. The alternative Power Apps custom connector path additionally needs maker access to create connectors in your environment.

Option 1: The MCP onboarding wizard (recommended)

Step 1: Open the wizard

  1. Open the Copilot Studio agent that should use the MCP server.
  2. Go to the Tools page and select Add a tool.
  3. Select New tool, then Model Context Protocol. The MCP onboarding wizard appears.

Step 2: Enter the server details

  1. Fill in Server name and Server description. Keep the description clear: the agent orchestrator uses it to decide when to call the server.
  2. Enter the Server URL (for example https://hjarni.com/mcp).
  3. Pick the authentication type: None, API key, or OAuth 2.0. For a server like Hjarni, choose OAuth 2.0 and then Dynamic discovery, which lets Copilot Studio register itself automatically through OAuth dynamic client registration.

Step 3: Create the connection and add it to the agent

  1. Select Create. The Add tool dialog appears.
  2. Select Create a new connection for your MCP server (sign in through the OAuth flow if prompted), or reuse an existing connection.
  3. Select Add to agent to finish.

Option 2: Power Apps custom connector

Prefer to build the connector by hand, or need to hand-tune the schema? Copilot Studio also lets you import an OpenAPI schema file through a Power Apps custom connector: in your agent, select Add a tool, then New tool, then Custom connector, which takes you to Power Apps to Import an OpenAPI file.

The OpenAPI YAML declares the Streamable transport with x-ms-agentic-protocol: mcp-streamable-1.0 on the MCP endpoint. The wizard above is simpler for most servers, so reach for the custom connector only when you need the extra control.

Verifying the connection

Open the test pane on the right side of the agent designer and ask the agent something that needs the new tools. For example, if you connected Hjarni:

"Search Hjarni for our onboarding checklist and list the open steps."

The trace panel shows the tool invocation, the request, and the response. If you see the tool call, the connection is working. If the trace says "Could not reach the MCP server" on the second turn but not the first, that is the streaming-timeout pattern covered below.

Troubleshooting

The agent loses the MCP server after the first call

If the connection works once and then drops on idle, the timeout is in front of your server. Copilot Studio only speaks Streamable HTTP now, so if you carried over a hand-built connector whose OpenAPI schema still declares SSE, recreate it through the MCP onboarding wizard. Then check the Azure Front Door or Application Gateway timeout in front of your server. MCP: SSE timeout or unexpected disconnect has the per-layer numbers.

The OAuth dialog never finishes

The sign-in happens inside the maker portal, so blocking third-party cookies can break it. Try a fresh browser profile. For servers that support it, the Dynamic discovery OAuth option needs the least manual configuration.

"401 Unauthorized" after the first day

The access token expired and was not refreshed. Most servers issue refresh tokens; if yours does not, edit the connection and sign in again. If you used an API key, make sure it has not been rotated.

Try it with Hjarni

Hjarni is a Markdown knowledge base with a built-in MCP server. Connect it to a Copilot Studio agent and the agent can search, read, and write your notes as part of a conversation. The server URL is https://hjarni.com/mcp.

Hjarni speaks Streamable HTTP and OAuth 2.0 with PKCE, so it drops straight into the onboarding wizard and the long-idle SSE drops do not apply. The same notes are read by Claude, ChatGPT, and GitHub Copilot, so the context follows you across clients.

Connect your notes

Write once. You both remember.

Free to start. No credit card required.

Give your AI a memory

Works with Claude and ChatGPT today. Gemini coming soon.