Azure DevOps Remote MCP Server Reaches GA, Without Support for Claude, ChatGPT, or Cursor

I’ll turn the supplied summary into a bilingual JSON article, keeping the examples clearly marked as adaptable rather than source claims.{"title_zh":"Azure DevOps 远程 MCP Server GA：云端入口已就绪，主流 AI 客户端还在等 Entra","body_zh":"# Azure DevOps 远程 MCP Server GA：云端入口已就绪，主流 AI 客户端还在等 Entra\n\nMicrosoft 将 Azure DevOps Remote MCP Server 推到 GA，关键信号很明确：Azure DevOps 正在把 work items、repos、pipelines 这些工程上下文，以托管 MCP 端点的形式开放出来。团队不需要先部署一台本地 MCP Server，理论上就能让支持 MCP 的工具读取项目状态、代码仓库和流水线信息。\n\n但这次 GA 也有一个现实边界：Claude Desktop、Claude Code、ChatGPT 和 Cursor 目前还不能直接连上。问题不在 MCP 概念本身，而在身份认证链路上：Entra 尚不支持这些客户端所依赖的 dynamic client registration 和 Client ID Metadata Documents。\n\n## GA 真正改变了什么\n\n过去很多 MCP 接入是“本地适配器”模式：开发者在机器上装一个 server，由它代理 IDE、CLI、数据库、工单系统或内部 API。Azure DevOps Remote MCP Server 走的是另一条路：服务端由 Microsoft 托管，入口直接面向 Azure DevOps 资源。\n\n这对工程组织有三点价值：\n\n- 部署面缩小：不用在每位开发者机器上安装和维护 Azure DevOps MCP Server。\n- 上下文更集中：work items、repos、pipelines 都属于真实研发流，不是额外维护的一份知识库。\n- 治理更自然：权限最终会落回 Azure DevOps 和 Entra 的身份边界，而不是分散在本地脚本里。\n\n不过，“GA”不等于“所有 AI 客户端马上可用”。Remote MCP 要进入日常工作流，还需要客户端、OAuth 元数据、租户策略和企业权限模型一起打通。\n\n## 为什么 Claude、ChatGPT、Cursor 还不能直接连接\n\nMCP 客户端连接远程服务器时，通常不能只靠一个 URL。它还需要知道如何完成授权、如何识别客户端、如何拿 token、token 又该发给哪个资源。对桌面工具和通用 AI 客户端来说，dynamic client registration 可以让客户端在运行时注册自己；Client ID Metadata Documents 则帮助授权服务器理解客户端身份和元数据。\n\n来源摘要里提到的限制是：Entra 目前缺少这两项能力支持，因此 Claude Desktop、Claude Code、ChatGPT 和 Cursor 这类客户端无法按它们期望的方式完成连接。\n\n这意味着团队短期内要避免两个误判：\n\n- 不要把 GA 理解成所有 MCP 客户端即插即用。托管端点可用，不代表客户端认证流程已经兼容。\n- 不要用个人 token 草率绕过企业身份边界。Azure DevOps 上下文往往包含源码、工单、发布流水线和审计信息，接入方式应当先过安全评审。\n\n## 可以先做的工程准备\n\n即使主流客户端暂时连不上，平台团队仍然可以提前准备接入清单：哪些项目允许被 MCP 读取，哪些 pipeline 日志包含敏感信息，哪些 work item 字段可能暴露客户或安全数据。\n\n还可以先把认证能力检查自动化。下面的脚本不是官方接入流程，而是一个可改造的预检工具：它检查 Entra OpenID Connect 元数据里是否暴露 registration_endpoint，并探测远程 MCP URL 在未带 token 时返回的状态。你需要替换 TENANT_ID 和 ADO_MCP_URL。\n\nbash\n#!/usr/bin/env bash\nset -euo pipefail\n\n: \"${TENANT_ID:?Set TENANT_ID to your Entra tenant ID}\"\n: \"${ADO_MCP_URL:?Set ADO_MCP_URL to the Azure DevOps Remote MCP endpoint you want to test}\"\n\nAUTHORITY=\"https://login.microsoftonline.com/${TENANT_ID}/v2.0\"\nOIDC_METADATA=\"${AUTHORITY}/.well-known/openid-configuration\"\n\necho \"== Entra OIDC metadata ==\"\nmetadata=\"$(curl -fsS \"$OIDC_METADATA\")\"\necho \"$metadata\" | jq '{\n issuer,\n authorization_endpoint,\n token_endpoint,\n registration_endpoint,\n jwks_uri\n}'\n\nif echo \"$metadata\" | jq -e '.registration_endpoint? // empty' >/dev/null; then\n echo \"Dynamic client registration appears to be advertised.\"\nelse\n echo \"No registration_endpoint found; clients requiring dynamic registration may not connect directly.\"\nfi\n\necho\n\necho \"== Remote MCP endpoint probe ==\"\ncurl -i -sS \\\n -H 'Accept: application/json, text/event-stream' \\\n \"$ADO_MCP_URL\" \\\n | sed -n '1,20p'\n\n\n运行方式：\n\nbash\nexport TENANT_ID=\"00000000-0000-0000-0000-000000000000\"\nexport ADO_MCP_URL=\"https://your-azure-devops-mcp-endpoint.example/mcp\"\n./check-ado-remote-mcp.sh\n\n\n如果 registration_endpoint 为空，不代表 Azure DevOps Remote MCP Server 不可用；它只说明依赖动态客户端注册的通用客户端还缺少认证闭环。这个检查结果可以放进平台接入文档，避免开发者把问题误判成网络、MCP 协议或 Azure DevOps 权限错误。\n\n## 落地时的取舍\n\n对企业团队来说，现在更像是“平台能力已到，客户端生态未完全对齐”的阶段。可以开始评估 Azure DevOps Remote MCP Server 的资源边界、权限模型和审计要求，但不宜承诺所有开发者马上能在 Claude、ChatGPT 或 Cursor 里直接使用。\n\n一个务实的采用清单是：\n\n- 确认哪些 Azure DevOps organization、project 和 repo 可以被 AI 工具访问。\n- 梳理 work item 字段、pipeline 日志和 repo 内容里的敏感数据。\n- 等待 Entra 对 dynamic client registration 和 Client ID Metadata Documents 的支持进展。\n- 在客户端支持前，避免用临时脚本或长期个人凭据绕开正式身份治理。\n- 把 Remote MCP 当作未来工程上下文入口来设计，而不是当作一次简单插件安装。\n\n这次 GA 的价值不在于“今天所有 AI 工具都能连”，而在于 Azure DevOps 已经把远程 MCP 入口产品化。真正的大规模使用，还要等身份协议和客户端实现补齐最后一段路。","title_en":"Azure DevOps Remote MCP Server Is GA, but Popular AI Clients Still Need Entra Support","body_en":"# Azure DevOps Remote MCP Server Is GA, but Popular AI Clients Still Need Entra Support\n\nMicrosoft has made the Azure DevOps Remote MCP Server generally available. The important shift is that Azure DevOps context can now be exposed through a hosted MCP endpoint covering work items, repos, and pipelines, without every developer first installing a local server.\n\nThere is a practical catch: Claude Desktop, Claude Code, ChatGPT, and Cursor cannot connect yet. According to the source summary, the blocker is Entra support for dynamic client registration and Client ID Metadata Documents, not the general idea of remote MCP itself.\n\n## What GA Changes\n\nMany early MCP integrations have followed a local adapter model: install a server on a developer machine, then let that server bridge the AI client to local tools, databases, APIs, or internal systems. Azure DevOps Remote MCP Server moves that server-side responsibility to a hosted endpoint.\n\nThat matters for engineering teams because:\n\n- Less local setup: developers do not need to install and maintain an Azure DevOps MCP server on each workstation.\n- Richer engineering context: work items, repos, and pipelines are part of the actual delivery workflow, not a separate knowledge base.\n- Cleaner governance path: access can align more naturally with Azure DevOps and Entra identity boundaries.\n\nStill, GA does not mean every MCP client can immediately connect. Remote MCP depends on the client, OAuth metadata, tenant policy, and enterprise access controls lining up correctly.\n\n## Why Claude, ChatGPT, and Cursor Are Blocked\n\nA remote MCP client generally needs more than a URL. It must know how authorization works, how the client identifies itself, how it obtains tokens, and which protected resource those tokens target.\n\nFor general-purpose desktop and AI clients, dynamic client registration allows a client to register itself at runtime. Client ID Metadata Documents help the authorization system understand the client identity and related metadata.\n\nThe source summary states that Entra does not yet support those pieces in the way these tools require. As a result, Claude Desktop, Claude Code, ChatGPT, and Cursor cannot directly connect today.\n\nThere are two common mistakes to avoid:\n\n- Do not equate GA with universal client compatibility. The hosted endpoint may be available while the client authentication flow remains unsupported.\n- Do not bypass identity governance with casual personal tokens. Azure DevOps context can include source code, work items, release details, pipeline logs, and audit-sensitive data.\n\n## A Practical Readiness Check\n\nEven before the most popular clients can connect, platform teams can prepare. Decide which projects are allowed to be exposed, review which pipeline logs contain sensitive output, and identify work item fields that may include customer or security data.\n\nThe script below is not an official connection flow. It is an adaptable readiness check: it inspects Entra OpenID Connect metadata for a registration_endpoint and probes a remote MCP URL to see how it responds without a token. Replace TENANT_ID and ADO_MCP_URL before running it.\n\nbash\n#!/usr/bin/env bash\nset -euo pipefail\n\n: \"${TENANT_ID:?Set TENANT_ID to your Entra tenant ID}\"\n: \"${ADO_MCP_URL:?Set ADO_MCP_URL to the Azure DevOps Remote MCP endpoint you want to test}\"\n\nAUTHORITY=\"https://login.microsoftonline.com/${TENANT_ID}/v2.0\"\nOIDC_METADATA=\"${AUTHORITY}/.well-known/openid-configuration\"\n\necho \"== Entra OIDC metadata ==\"\nmetadata=\"$(curl -fsS \"$OIDC_METADATA\")\"\necho \"$metadata\" | jq '{\n issuer,\n authorization_endpoint,\n token_endpoint,\n registration_endpoint,\n jwks_uri\n}'\n\nif echo \"$metadata\" | jq -e '.registration_endpoint? // empty' >/dev/null; then\n echo \"Dynamic client registration appears to be advertised.\"\nelse\n echo \"No registration_endpoint found; clients requiring dynamic registration may not connect directly.\"\nfi\n\necho\n\necho \"== Remote MCP endpoint probe ==\"\ncurl -i -sS \\\n -H 'Accept: application/json, text/event-stream' \\\n \"$ADO_MCP_URL\" \\\n | sed -n '1,20p'\n\n\nExample run:\n\nbash\nexport TENANT_ID=\"00000000-0000-0000-0000-000000000000\"\nexport ADO_MCP_URL=\"https://your-azure-devops-mcp-endpoint.example/mcp\"\n./check-ado-remote-mcp.sh\n\n\nIf registration_endpoint is missing, that does not mean the Azure DevOps Remote MCP Server is unavailable. It means clients that depend on dynamic registration may not be able to complete the authentication loop directly. That distinction is useful when debugging, because it keeps teams from mislabeling the issue as a network problem, MCP protocol failure, or Azure DevOps permission error.\n\n## Adoption Guidance\n\nFor enterprise teams, this is a platform-readiness moment rather than an instant rollout moment. Azure DevOps now has a productized remote MCP entry point, but the broader client ecosystem still needs the identity layer to catch up.\n\nA pragmatic checklist looks like this:\n\n- Define which Azure DevOps organizations, projects, and repositories AI tools may access.\n- Review sensitive data in work item fields, pipeline logs, and repository content.\n- Track Entra support for dynamic client registration and Client ID Metadata Documents.\n- Avoid temporary shortcuts based on long-lived personal credentials.\n- Treat Remote MCP as a future engineering context layer, not just another plugin install.\n\nThe GA milestone is still meaningful. Microsoft has moved Azure DevOps MCP access into a hosted model. The remaining work is making identity and client behavior mature enough for everyday use in tools like Claude, ChatGPT, and Cursor.","seo_description_en":"Azure DevOps Remote MCP Server is GA with hosted access to work items, repos, and pipelines, but Claude, ChatGPT, and Cursor still cannot connect."}