Back to Product Wiki

Back to MCP Server overview

Overview

This help article will show you how to utilize iTwin IoT's MCP server with Microsoft Copilot.  

 

Setting up the MCP Server for Microsoft Copilot

As of today, we do not recommend the Federated connector offered by Microsoft; that connector only offers read-only tools.

The best way to set up the MCP server connection for Microsoft Copilot is via a Custom Connector. 

For reference, please read through Microsoft's documentation on configuring MCP Connect clients: https://learn.microsoft.com/en-us/power-pages/configure/mcp-connect-clients

 

Bentley IMS doesn't support OAuth 2.0 Dynamic Client Registration (DCR), so Copilot Studio's Dynamic discovery / Dynamic OAuth options don't work with our MCP server. Use Manual OAuth 2.0 instead — each client self-registers their own confidential IMS client, no shared secret needed.

Creating/editing an agent in Copilot Studio requires a Copilot Studio license (or consumed Copilot Studio credits/capacity) for the user or tenant. Once the agent is set up with the MCP tool, it must be published before it's actually usable (e.g. surfaced in Microsoft 365 Copilot / Teams / a shared channel) — testing it in the Copilot Studio authoring canvas alone isn't enough for end users to reach it.

So the flow is always:

- create (or open) an agent first

- add the MCP server as a tool to that agent

- If you don't already have an agent, you must create one (even an empty/minimal one) before you can do the MCP setup below.

Set up IMS authentication

- Register a Web Application type client at developer.bentley.com  (or IMS Wizard /wizard/api).

- Redirect URI: https://global.consent.azure-apim.net/redirect (fixed Power Platform / Copilot Studio callback — same for every registration, known upfront, no need to come back and update it after creating the tool).

- Scopes: itwin-platform — the only scope offered for Web Application registration on developer.bentley.com.

- Copy the new Client ID + Client secret and save it for setting up Microsoft Copilot

- Send us the Client ID, then we request the IMS team to adjust the scopes (to itwiniot-ai:mcp) for you.

Add the MCP server as a tool in Copilot Studio

In an existing (or newly created) Copilot Studio  agent: Tools → Add a tool → New tool → Model Context Protocol, then OAuth 2.0 → Manual:

Field	Value
Server URL	https://itwiniotai-mcp.bentley.com/mcp
Authorization URL	https://ims.bentley.com/connect/authorize
Token URL template	https://ims.bentley.com/connect/token
Refresh URL	Same as Token URL
Client ID / Client secret	from step 1.4 above
Scopes	itwiniot-ai:mcp offline_access itwiniot-ai:mcp-tool-search

 

Using the MCP Server in Microsoft Copilot

Coming soon.

 

Data handling and privacy

• What we collect. When Microsoft Copilot calls a tool, the connector logs the authenticated user identifier (your IMS sub claim), the tool name, the duration of the call, and the outcome (success/error code). We do not log the contents of conversation messages or the raw parameter values you pass to tools. The data is kept for only 30 days and is used solely for monitoring, troubleshooting, and abuse prevention.
• What we forward. Each tool call is authorised with your IMS access token. The iTwin IoT backend independently re-validates the token and enforces RBAC on every call; the connector cannot return data that your account is not entitled to.
• Storage. Audit logs are stored in Bentley's Azure tenant under the same governance as iTwin IoT itself.
• Third-party sharing. None beyond the iTwin IoT backend that you already use.
• Retention. Audit logs follow the default Bentley operational retention for monitoring data.
• Full privacy policy: see Bentley's global data privacy notice at https://www.bentley.com/legal/privacy-policy/.

 

Limits and quotas

To protect the iTwin IoT backend, the connector applies per-user sliding-window rate limits:

• 100 tool calls / 60 seconds per user (read-mostly tools).
• 30 tool calls / 60 seconds per user for destructive tools (delete_*, cancel_*, mutating ML-job operations).

When you hit the limit, the connector returns a structured rate_limited error and the agent will wait and retry, or surface the limit to you to decide.