Connect an MCP server to Copilot Studio and give your agent real tools

Model Context Protocol (MCP) is the open standard that lets an agent talk to tool servers — and Copilot Studio speaks it natively. Instead of wiring one connector action at a time, you point the agent at a server and every tool it publishes becomes callable, with changes on the server reflected automatically.

What MCP gives an agent

A connected MCP server exposes tools (functions the model can call) and resources (file-like data a tool can return) — Copilot Studio supports both, while MCP's third concept, prompt templates, isn't supported yet. The server owns the names, descriptions, inputs and outputs; when its owner updates or removes a tool, your agent sees the change without re-publishing. One server can carry a whole family of capabilities, which is exactly why it beats configuring actions one by one.

Two prerequisites set the frame: your agent needs generative orchestration turned on, and the server must use the Streamable HTTP transport — Copilot Studio dropped the deprecated SSE transport after August 2025, so older servers need updating.

Connect with the onboarding wizard

1. On the agent's Tools page, select Add a tool > New tool > Model Context Protocol.
2. Fill in Server name, Server description and the HTTPS Server URL. Write the description for the orchestrator, not for people — it's how the agent decides when to call this server.
3. Pick the authentication type and select Create, then create the connection and add the server to the agent.

Prebuilt Microsoft MCP connectors appear in the same Add a tool dialog ready to add, and an existing server can alternatively be wrapped as a custom connector through Power Apps when you need connector-level control.

Pick the authentication

• None — for servers that are open inside your network boundary.
• API key — the simple option: sent as a header or query parameter you name.
• OAuth 2.0 — per-user sign-in with consent, in three flavours: dynamic discovery (the server supports discovery plus dynamic client registration — the least configuration), dynamic (registration without discovery; you provide the endpoints), and manual (client ID, secret, authorisation, token and refresh URLs, scopes). Manual setups surface a callback URL to register with your identity provider.

Trim the toolbox, then test

Adding a server enables all its tools (Allow all is on). On the server's settings page, turn that off and enable only what the agent needs — new tools the server adds later then arrive disabled, which is the safer default. In the test chat, ask something that requires a tool and check the agent actually called it; if it didn't, the server description or the agent instructions are usually the fix, while the tool names and descriptions themselves belong to the server's author.

Pause animation

1. 1Add tool, Model Context Protocol
2. 2Describe the server for the orchestrator
3. 3Authenticate
4. 4Disable the tools you don't want
5. 5Test a tool-requiring question

Governance comes free but real: MCP connectivity rides Power Platform connectors, so your data policies (DLP) govern MCP servers the same way they govern connectors — and a connected non-Microsoft server is your responsibility to vet. Agents on the GitHub Copilot harness add one more limit: a cap on concurrent MCP servers per conversation, so keep each agent's server list small and current.

Sources

• Extend your agent with Model Context Protocol — Microsoft Learn
• Connect your agent to an existing MCP server — Microsoft Learn
• Add tools and resources from an MCP server — Microsoft Learn
• Add an MCP server to your agent (GitHub Copilot harness, preview) — Microsoft Learn