Azure DevOps remote server now available
The Azure DevOps Remote MCP Server is now available to all developers, providing a hosted endpoint that links AI assistants to work items, pull requests, repositories, and pipelines without local installations. The service currently excludes third-party tools like Claude Desktop, Claude Code, ChatGPT, and Cursor due to limitations in Microsoft’s authentication system.
How it works—and current limitations
The server operates at https://mcp.dev.azure.com/{organization} and uses streamable HTTP for communication. Connecting a supported client requires adding one line to an mcp.json configuration file:
- “url”: “https://mcp.dev.azure.com/{organization}”
Authentication occurs through Microsoft Entra, which creates a bottleneck. Dan Hellem, product manager for Azure Boards, Repos, and Wiki, explained that tools like Claude and Cursor lack support for dynamic OAuth client registration or Client ID Metadata Documents in Entra. Microsoft is collaborating with the Entra team to add these features, but users of those tools must still rely on the local MCP server for now.
The release timing complicates matters further. The MCP 2026-07-28 specification, published a week before the general availability announcement, revised the preferred authentication methods. It now prioritizes pre-registered clients and Client ID Metadata Documents, while marking Dynamic Client Registration as deprecated and set for removal after summer 2027. Entra does not currently support either of the specification’s preferred methods, and one is already scheduled for deprecation.
Another restriction limits access to organizations with an Entra tenant. Those using standalone Microsoft accounts cannot use the remote server.
Current user base
Supported clients remain confined to Microsoft’s ecosystem. Visual Studio Code with GitHub Copilot, Microsoft Foundry, and Copilot Studio connect without additional setup, as do Visual Studio, the GitHub Copilot CLI, and the GitHub Copilot app. Teams using other AI assistants must continue running the local MCP server, though Microsoft has aligned the local toolset with the remote version and will maintain parity during Entra updates.
Related: Is Four Years of College Becoming Obsolete in the Age of AI and Startup Culture?
The divide between first-party and third-party access carries practical implications. Farhan Shahnewaz, an AI and cloud solutions engineer at Microsoft, highlighted how the system simplifies build triage. Developers can ask an AI assistant which pipeline stage failed and why, with the tool retrieving context and logs from the same system the team already uses. A hosted server removes the need for individual instances per developer or team, along with the associated credential management. Teams unable to access the remote server must manage that overhead themselves.
Shahnewaz also described the authentication model as a security advantage. Unlike local setups where personal access tokens might remain in config files, Entra ensures AI assistants inherit only the developer’s exact permissions. This addresses a key concern for security teams, as the hosted server enforces it by default.
The discrepancy between the MCP specification and Entra’s capabilities reflects a broader issue. While the protocol standardizes how clients discover and call tools, the underlying identity layer remains proprietary. A vendor can define an interoperable interface, but identity providers may still refuse to authenticate third-party clients. Microsoft has not provided a timeline for resolving this.
For now, the remote server offers a zero-install option for Microsoft’s tools, while others must maintain their own infrastructure. The company’s commitment to the local server suggests this is not intentional exclusion, but the result is clear: teams using Claude or Cursor face more friction than those relying on Copilot’s efficiency.
This gap may narrow as Entra evolves, but for the moment, the hosted endpoint remains a Microsoft-centric solution.
