10.2 Configure MCP Tools
Key Takeaways
- MCP exposes server-published tools (callable functions) and resources (file-like context); Copilot Studio currently supports MCP tools and resources, not server-side prompt templates as a first-class separate surface.
- Generative orchestration must be enabled to use MCP with an agent.
- Add MCP from Tools → Add a tool → Model Context Protocol, authorize the connection, then Add and configure; the server appears as one agent tool with Tools and Resources lists.
- Tool catalogs are dynamic: updates on the MCP server reflect in Copilot Studio; you can turn off Allow all and disable individual tools; new tools stay off when Allow all is off.
- MCP connectivity rides Power Platform connectors, so DLP/data policies that regulate connectors also regulate MCP server and tool access.
Why MCP tools matter on AB-620
Configure MCP tools sits next to computer use, custom connectors, and REST APIs in the Add tools to agents cluster. Exam items often ask how MCP differs from a one-off REST import, what the agent actually receives from a server, and how governance still applies.
What Model Context Protocol is
Model Context Protocol (MCP) is a standardized way to connect AI models and agents to external knowledge servers and systems. When you connect an MCP server to Copilot Studio, the server can publish:
| MCP concept | Meaning | Copilot Studio support |
|---|---|---|
| Tools | Functions a language model can call to perform an action (with names, descriptions, inputs, outputs) | Supported |
| Resources | File-like data the agent can read for more context (API payloads, file contents) | Supported |
| Prompts | Predefined prompt templates for specific tasks | Not the primary Copilot Studio MCP surface today—focus study on tools and resources |
Each tool or resource published by a connected server becomes available for the agent. The server supplies name, description, inputs, and outputs. When the server updates or removes items, Copilot Studio dynamically reflects those changes so makers are not manually re-importing a static OpenAPI file for every catalog edit.
A single MCP server can host many tools; the agent can access each enabled tool through the server connection. For non-Microsoft or external MCP servers, your organization remains responsible for the tools and data the agent can reach.
Hard prerequisite: turn on generative orchestration. MCP is not designed as a classic-only Action node substitute without orchestration.
How MCP fits the broader tool map
| Approach | Catalog style | Best when |
|---|---|---|
| MCP | Live server catalog of tools/resources | Partner or internal platform already speaks MCP; you want dynamic discovery |
| REST API tool | OpenAPI uploaded into Copilot Studio | You own a REST surface and want per-operation tools without standing up MCP |
| Custom connector | OpenAPI wrapped as a Power Platform connector | Reuse across Power Automate, Power Apps, and agents with DLP/connector lifecycle |
| Computer use | Natural-language GUI plan | No API/MCP endpoint; human UI only |
Extend an agent with MCP (end-to-end path)
Microsoft’s integration path has four conceptual steps:
- Connect to an existing MCP server (MCP onboarding wizard / existing server flow).
- Create an MCP server if you do not already have one (custom development or Microsoft samples).
- Add MCP server tools and resources to the agent so runtime can call them.
- Optionally publish/certify an MCP connector so it can be used more broadly across tenants (connector certification path).
Add an MCP server as an agent tool (click path)
Whether you use a prebuilt Microsoft MCP connector or a server you connected earlier, the agent-side path is the same:
- Open the agent Tools page.
- Select Add a tool.
- Select Model Context Protocol — available MCP connectors list appears.
- Select the MCP connector.
- Authorize the connection (enter required credentials or complete the identity flow).
- Select Add and configure.
The agent adds the MCP server as a tool and opens the MCP settings page. On Tools, the server shows as a single tool entry that expands into the server’s capability lists.
View tools and resources after connection
Open the MCP server from the agent’s Tools list. The settings page includes:
- Details — same style of name/description and agent-level options as other tools (when to allow dynamic use, ask before running, authentication mode where applicable).
- Tools — names and descriptions of MCP tools the server offers.
- Resources — sampling of resource names and descriptions.
Important resource rule for exams: for a Copilot Studio agent to use a resource, the MCP server owner must configure that resource as an output of one of the MCP tools. Resources are not a free-floating second knowledge search by default; they ride tool outputs in the reference implementation model.
Customize tool selection (Allow all)
By default, when you add an MCP server, all tools are on and the Allow all toggle is on.
To restrict:
- Turn Allow all off — per-tool toggles become available.
- Disable tools the agent should never call (for example delete or admin operations).
When Allow all is off, new tools that appear later on the MCP server are off by default. That is a security and change-management feature: catalog growth does not silently expand agent privileges.
Tool configuration differences vs ordinary tools
For most connector/REST tools, the configuration page emphasizes Details, Inputs, and Completion. For MCP servers, Tools and Resources replace the usual Inputs/Completion layout because inputs live on each server-defined tool. Still treat the server-level Description as orchestration fuel: poor descriptions cause wrong tool selection or under-use.
Security boundaries and data policies
Access to MCP servers in Copilot Studio relies on Power Platform connectors for connectivity. Therefore:
- If a data policy (DLP) regulates Power Platform connectors, it also regulates access to the MCP server and its tools for your agent.
- Plan MCP the same way you plan premium connectors: environment policies, connector classification, and least-privilege connections.
- External MCP servers expand your supply chain—review tool names, descriptions, and side effects before enabling Allow all in production.
Authentication is connection-specific. Complete the authorize step when adding the connector; do not assume anonymous MCP is always available (some public servers such as documentation MCP endpoints may differ, but enterprise servers usually require auth).
Scenario: Contoso engineering MCP for work items
Contoso Engineering publishes an internal MCP server with tools: list_work_items, get_work_item, add_comment, and delete_work_item.
- Maker enables generative orchestration on the support agent.
- Tools → Add a tool → Model Context Protocol → select Contoso Work Items MCP → authorize with app registration credentials.
- Add and configure; open Tools list and confirm all four tools appear.
- Turn Allow all off; leave list/get/comment on; turn delete_work_item off so the conversational agent cannot destroy records.
- Confirm environment DLP allows the MCP connector classification in this environment.
- Test prompts: “Show open bugs for Project Atlas” should select list/get tools; “Delete all bugs” should not invoke delete because the tool is disabled.
- When Engineering adds
reassign_work_itemon the server, it appears in the catalog but stays off until a maker explicitly enables it.
Troubleshooting mindset for the exam
| Symptom | Likely check |
|---|---|
| MCP option missing or tools never selected | Generative orchestration off |
| Connector missing from list | Server not onboarded / not shared to environment; auth incomplete |
| Unexpected new capabilities in production | Allow all left on while server catalog grew |
| Policy error at runtime | DLP blocks the underlying connector |
| Resource content never used | Resource not exposed as tool output on the server |
Exam checklist
- Define MCP tools vs resources in one sentence each.
- State generative orchestration as mandatory.
- Recite Add tool → Model Context Protocol → authorize → Add and configure.
- Explain dynamic catalog updates and selective tool toggles with Allow all behavior.
- Tie MCP access to Power Platform connector DLP.
- Contrast MCP with REST upload and computer use so scenario questions do not force a single pattern.
What must be true before a Copilot Studio agent can use tools from a connected MCP server under generative tool selection?
A maker adds an MCP server and leaves Allow all enabled. Two weeks later the server owner publishes a new destructive admin tool. What is the default effect on the agent?
How do Power Platform data policies relate to MCP tools in Copilot Studio?