Available on Plus in the beta: the hosted ChangePilot MCP server connects compatible AI assistants to your authorised data without installing a local server. It supports reading data and supported actions on your watches, bookmarks and personal labels.
Connect using your ChangePilot account
Open your AI client's remote MCP or connector settings.
Add
https://changepilot.azure-api.net/mcp.Complete the ChangePilot sign-in and authorisation when prompted.
Check the available tools and try a small read request before asking it to change your items.
In Claude Code, add the HTTP server with claude mcp add --transport http changepilot https://changepilot.azure-api.net/mcp, then use /mcp to authenticate. For other clients, follow the current setup instructions on the MCP Server page.
What you can ask
Find and explain Message Center, Microsoft 365 Roadmap and Azure Updates items, including history and date changes.
Read your watchlist and recent activity, bookmarks, personal labels and saved Views.
Add or remove watches and bookmarks.
Create, edit, delete, apply and remove personal labels.
For example: “Find recent high-admin-impact Teams changes, include IDs and links, then show me which ones I already watch.” Review the result before authorising bulk changes. The available tool list is the authority for supported actions; do not assume View editing or organisation administration is supported because Views can be read.
Scope and client support
Requests run with your account's access. Roadmap and Azure are global datasets; Message Center follows your authorised feed. MCP writes are recorded against your account. A compatible client must support remote MCP and OAuth.
Copilot Studio requires a registered OAuth client ID and secret; contact ChangePilot support for setup. Do not assume the URL alone completes a Copilot connection.
Rate limits and retries
The hosted MCP endpoint currently allows 30 requests per minute per caller credential (checked on 26 October 2026). Requests using the same credential share the allowance. Connection, tool-discovery and other protocol requests count as well as tool calls.
If a request returns HTTP 429 (Too Many Requests), pause and follow the Retry-After header. Your AI client may show this as a failed tool call. Wait before trying again, reduce bursts and avoid asking an agent to repeat a failing call continuously.
We may rate-limit API and MCP use and may change the limits with reasonable notice. See API and MCP rate limits for the current published limits, REST API allowances and guidance for the separate local MCP server.
Disconnect
Remove the connector in the AI client. If access may be compromised, contact support to revoke its authorisation. The separate local server uses an API key and has a smaller, read-only tool set.