setup-sso-ui-widget

作者: microsoft

为Microsoft 365 Copilot声明式代理添加Microsoft Entra SSO(单点登录,无OBO),该代理的工具由MCP服务器提供——适用于BOTH widget…

npx skills add https://github.com/microsoft/work-iq --skill setup-sso-ui-widget

Setup SSO for a ui-widget-developer Agent (Minimal-Touch, No OBO)

Why this exists. Copilot MCP-server agents come in two widget-standard shapes, and this skill supports both (defaulting to MCP Apps when the layout is ambiguous):

  • MCP Apps (create-mcp-app / MCP Apps SDK): a runtimes[] plugin manifest such as appPackage/readiness_plugin.json and an Express-based MCP server (main.ts).
  • OAI Apps (ui-widget-developer): appPackage/mcpPlugin.json and a raw-http MCP server under mcp-server/ (no express, no express-jwt).

Both provide a named devtunnel that is already running and env/.env.local. Phase 0 auto-detects the layout (server framework + plugin manifest) and the skill adapts the manifest patch and guard injection accordingly — it never re-scaffolds, so the user's server stays intact.

New to how SSO works here? Read references/sso-explained.md first — it covers what SSO gives you (verified identity, not downstream access), the end-to-end token flow, how claims reach your tools via claimsStore, failure modes, and how to go further with OBO / Microsoft Graph. This SKILL is the procedural runbook; that doc is the mental model.

CRITICAL EXECUTION RULES — READ BEFORE PROCEEDING:

  • Execute every az, devtunnel, atk, npm, and PowerShell command in the TERMINAL yourself. Do NOT tell the user to run them.
  • Do NOT improvise alternate approaches for the Entra/ATK steps — reuse the shared reference files under references/.
  • Execute commands ONE AT A TIME, check output, diagnose failures, retry — never skip.
  • NO SCRATCH FILES — PATTERN-BASED, NOT NAME-BASED: Run commands directly in the terminal and keep all state in shell variables. NEVER create a file whose purpose is to capture, stage, or read back command output — regardless of its name or extension (.txt, .json, .log, .ps1, …). This ban covers redirecting with >, Out-File, Tee-Object, or Set-Content so you can read the result later. Permitted exception: a short-lived temp file used only to pass a request body to az rest --body @file (as the shared reference files do) — written immediately before the call and deleted immediately after with Remove-Item; it never captures or reads back output. Concrete violations seen in testing that are FORBIDDEN: atk provision ... > atk_prov_out.txt, az ad app show ... > appverify.json, plus sso-step*.ps1, sso-*.log, sso-*.txt, sso-state.json, *-precheck.txt, server-sso.*.log, server-pid.txt. The ONLY files this skill writes are the ones explicitly shown in its phases (auth.ts, edits to the plugin manifest (mcpPlugin.json / readiness_plugin.json) / declarativeAgent.json / env/.env.local / m365agents.local.yml / m365agents.yml / the MCP server entry file). Do NOT delete or alter the ui-widget skill's own files (tunnel.log, server.log, pids.txt, etc.).
  • TERMINAL OUTPUT LAGS? DO NOT REDIRECT TO A FILE. If the terminal renders "one step behind", capture the output into a variable in the SAME shell and print it — no file: $out = az ad app show --id $ClientId 2>&1 | Out-String; $out. For atk provision, do NOT scrape stdout at all — read the generated values straight from env/.env.local (Phase 4d). Re-running a read-only query (az ... show) is always safe. Inventing a file to work around lag is never acceptable.
  • TERMINAL RULES: Background/separate terminals get a fresh shell with NO inherited variables. Use literal values (e.g., devtunnel host myapp-tunnel) in those terminals. Never put short timeouts on az commands.

FORMATTING RULES (align with ui-widget-developer):

  • When you need a decision or input from the user, ask it with the AskUserQuestion tool — one structured question at a time — exactly as ui-widget-developer does. Do NOT bury questions in prose.
  • Render every "Tell the user" note as a markdown blockquote (> prefix); do NOT flatten it into a paragraph.

Scope Guardrails

  • SSO only: Entra app registration + ATK OAuth registration + plugin-manifest (mcpPlugin.json / readiness_plugin.json) auth wiring + minimal token validation + sideload.
  • No OBO: do NOT add downstream delegated token exchange / Microsoft Graph calls.
  • Minimal touch: do NOT refactor the widget or rewrite the MCP server to express. Only add a small JWKS guard + a per-request claims store.
  • One tunnel: REUSE the tunnel ui-widget-developer already created. Never create a second tunnel on the same port.

⛔ Phase 0 — Workspace Check (MANDATORY FIRST STEP)

This skill expects a project produced by ui-widget-developer.

▶ Run the Phase 0 step in references/detect-and-inputs.md — it detects the ui-widget layout and STOPS if the project isn't a ui-widget-developer output.


Phase 1 — Prerequisites (EXECUTE)

Windows only — refresh PATH in the current PowerShell session (skip on macOS/Linux):

$env:Path = [System.Environment]::GetEnvironmentVariable("Path","Machine") + ";" + [System.Environment]::GetEnvironmentVariable("Path","User")

Set $SsoScripts to this skill's scripts/ folder (absolute path). Every phase runs its logic via pwsh -NoProfile -File "$SsoScripts/<name>.ps1" — the scripts carry state through env/.env.local, so you rarely pass arguments:

$SsoScripts = "<absolute path to this skill>/scripts"
ToolCheckAuto-install
Azure CLIaz versionwinget install Microsoft.AzureCLI
ATK CLI (>=1.1.8)atk --versionnpm install -g @microsoft/m365agentstoolkit-cli
Dev Tunnel CLIdevtunnel --versionwinget install Microsoft.devtunnel
Node.js (>=20)node --versionwinget install OpenJS.NodeJS.LTS

After installing any tool, refresh PATH with the snippet above. Tag CLI usage once: $env:ATK_CLI_SKILL = "true".


Phase 2 — Gather Inputs + Reuse Existing Tunnel (EXECUTE)

▶ Run the Phase 2 step in references/detect-and-inputs.md — derives the Entra app display name and reuses the EXISTING dev tunnel from env/.env.local (never creates a second one).


Phase 3–5 — Register + Configure the Entra App (EXECUTE)

▶ Execute references/register-app.mdPhase 3 creates the single-tenant Entra app + service principal + redirect URI, Phase 4 injects oauth/register and provisions with --env local (yielding the Auth ID + Application ID URI), and Phase 5 sets the App ID URI, v2 tokens, the access_as_user scope, and pre-authorizes M365 Copilot. Graph User.Read + admin consent stay opt-in (OBO only).


Phase 6–8 — Wire SSO + Inject Guard + Write Env (EXECUTE)

▶ Execute references/wire-and-guard.mdPhase 6 flips mcpPlugin.json runtime auth to OAuthPluginVault (+ conditional starters), Phase 7 adds jose, copies the hardened guard from references/auth.ts, and inserts it into the /mcp handler (+ CORS), and Phase 8 writes TENANT_ID / CLIENT_ID / APP_ID_URI for the server. Includes the Azure/Easy Auth deployment note.


Phase 9–11 — Build, Verify, Clean Up (EXECUTE)

▶ Execute references/build-verify-cleanup.mdPhase 9 builds + re-provisions + atk validate/atk install (sideload), Phase 10 starts the server (with SSO_DEBUG=1) and verifies an unauthenticated /mcp POST returns 401, and Phase 11 cleans up SSO scratch.


🎉 FINAL SUMMARY (render DIRECTLY in your reply — NOT inside a code fence)

Output the following structure as plain markdown in your chat reply. Do NOT wrap it in fences. Fill every <placeholder> with the actual value gathered during the run; mark unknowns N/A. Use bold-label bullets (shown below) — do NOT convert to a markdown table, since some chat surfaces render tables inconsistently.

🎉 ✅ SSO Setup Complete — ui-widget agent

What changed (minimal-touch, no OBO)

  • Registered an Entra app and ATK OAuth (MicrosoftEntra) configuration.
  • Reused the EXISTING dev tunnel — no second tunnel created.
  • Added a JWKS bearer-token guard to the MCP server (new auth.ts + a guard in the /mcp handler).
  • Switched mcpPlugin.json runtime auth from NoneOAuthPluginVault.
  • Wrote TENANT_ID / CLIENT_ID / APP_ID_URI into env/.env.local.
  • Re-provisioned, validated, and sideloaded the agent.

App registration details

  • App display name: <AppDisplayName>
  • Client ID: <ClientId>
  • Object ID: <ObjectId>
  • Tenant ID: <TenantId>
  • Auth configuration ID (SSO): <AuthId>
  • Application ID URI: <AppIdUri>
  • Scope: <AppIdUri>/access_as_user
  • Backend (reused tunnel): <BaseUrl>
  • Tunnel name: <TunnelName>
  • Local port: <Port>

Changed files

  • <AppPackageDir>/mcpPlugin.json — runtime auth
  • <McpServerDir>/src/auth.ts — new
  • <McpServerDir>/src/index.ts — guard + CORS header + success-path log
  • <McpServerDir>/package.jsonjose dep
  • env/.env.localTENANT_ID / CLIENT_ID / APP_ID_URI
  • m365agents.ymloauth/register

Test status

  • Build: <pass/fail>
  • Validation: <pass/fail>
  • Sideload/install: <pass/fail>
  • Local 401 check: <result>

Test in Copilot

  1. Open https://m365.cloud.microsoft/chat (agent may take up to 15 min to appear).
  2. All agents → search <AppDisplayName>.
  3. Ask something that triggers a widget tool.
  4. Accept the one-time consent prompt.
  5. Confirm the widget renders with your signed-in identity (claims like name/oid flow via the SSO token).
  6. In the MCP server terminal, you should see one [auth] Valid SSO token accepted: { sid, aud, tid, iss } line per authenticated call — quick proof SSO is live. (This line only prints when the server runs with SSO_DEBUG=1, as in Phase 10; it is gated off by default.)

Deployment note — Azure (Easy Auth)

auth.ts is for local dev/testing only. When you host the MCP server on Azure App Service, use the platform's built-in Easy Auth instead of the custom guard. Reproduce this entire block in your summary — do NOT paraphrase it away. Configure App Service → Authentication → Microsoft Entra identity provider with:

  • App registration: the Entra app this skill created (<ClientId>) — do not create a new one.
  • Supported account types: Single tenant.
  • Issuer URL: https://login.microsoftonline.com/<TenantId>/v2.0
  • Allowed token audiences: <ClientId> — the bare client-id form (NOT api://…), else 401.
  • Client application requirement: Allow requests from specific client applications → add BOTH <ClientId> and ab3be6b7-f5df-413d-ac2d-abf1e3fd9c0b (M365 Copilot host). Missing the host id → 403 even with a valid token.
  • Unauthenticated action: Return HTTP 401.

📄 Full field-by-field guide (both blade tables + both gotchas): references/easy-auth.md — cite it by name in the summary. Also include:

⚠️ Caveat — if your dev tunnel URL changes later

The OAuth registration's base URL, the Application ID URI, and the Entra app's identifierUris are all tied to the current dev tunnel domain. If that tunnel URL changes, authentication will break. Re-sync with pwsh -NoProfile -File "$SsoScripts/resync-tunnel-url.ps1" (or just re-run /setup-sso-ui-widget) so the OAuth base URL + App ID URI match the new tunnel, then restart the server and re-test.


Notes & Error Handling

Concepts, behavior & failure modes are documented in references/sso-explained.md — including the runtime token flow, the aud/iss validation rules, the claimsStore per-request identity pattern, a symptom→cause table, and how to extend to OBO / Microsoft Graph. Start there when debugging or when you need to understand why a phase does what it does.

  • Two tunnels? This skill reuses the tunnel from env/.env.local. If you ever see a second tunnel, stop it and keep the named one the ui-widget script created.
  • Dev tunnel URL changed → auth suddenly broken. The OAuth registration is created with baseUrl: ${{MCP_SERVER_URL}}, ATK derives the Application ID URI from that tunnel domain, and it's written into the Entra app's identifierUris. So the tunnel domain is baked into the OAuth registration, the App ID URI, and the Entra app. If the tunnel URL changes (tunnel deleted/recreated, anonymous tunnel expired, new name/port, different machine), the token audience stops matching and every authenticated call fails. Recovery: run pwsh -NoProfile -File "$SsoScripts/resync-tunnel-url.ps1" from the project root — it detects the new URL (or pass -NewUrl https://...), updates MCP_SERVER_URL/MCP_SERVER_DOMAIN, and re-runs Phases 4 → 5 → 6 → 9 so the OAuth base URL, App ID URI, identifierUris, and mcpPlugin.json all realign. Then restart the server against the new tunnel and re-test.
  • 401 in Copilot (not local): server audience must equal $AppIdUri and issuer tenant $TenantId — confirm env/.env.local and that the server loads it. (See §5 of sso-explained.md.)
  • mcpPlugin.json vs ai-plugin.json: this skill is specifically for the ui-widget mcpPlugin.json layout; ai-plugin.json (express-jwt) projects aren't supported here.
  • No OBO here. For Microsoft Graph / downstream APIs, use a separate OBO flow later (out of scope). See §7 of sso-explained.md for what that delta looks like.

来自 microsoft 的更多技能

oss-growth
microsoft
OSS增长黑客角色
agent-framework-azure-ai-py
microsoft
使用Microsoft Agent Framework Python SDK(agent-framework-azure-ai)构建Azure AI Foundry代理。在创建使用AzureAIAgentsProvider的持久化代理、使用托管工具(代码解释器、文件搜索、网络搜索)、集成MCP服务器、管理对话线程或实现流式响应时使用。涵盖函数工具、结构化输出和多工具代理。
development
airunway-aks-setup
microsoft
Set up AI Runway on AKS — from bare cluster to running model. Covers cluster verification, controller install, GPU assessment, provider setup, and first deployment. WHEN: "setup AI Runway", "onboard AKS cluster", "install AI Runway", "airunway setup", "deploy model to AKS", "GPU inference on AKS", "KAITO setup on AKS", "run LLM on AKS", "vLLM on AKS", "set up model serving on AKS", "AI Runway controller".
devops
appinsights-instrumentation
microsoft
使用Azure Application Insights对Web应用进行插桩的指南。提供遥测模式、SDK设置和配置参考。适用场景:如何对应用进行插桩、App Insights SDK、遥测模式、什么是App Insights、Application Insights指南、插桩示例、APM最佳实践。
devops
applicationinsights-web-ts
microsoft
使用Application Insights JavaScript SDK(@microsoft/applicationinsights-web)为浏览器/Web应用添加检测。用于真实用户监控(RUM)——页面视图、点击、AJAX/fetch依赖项、异常、自定义事件,以及与后端OpenTelemetry追踪关联的浏览器端GenAI代理追踪。涵盖SDK加载器脚本和npm设置、框架扩展(React、React Native、Angular)、点击分析、遥测初始化器,以及从浏览器发出的代理/工具/模型跨度所遵循的OTel GenAI语义约定。
devops
azure-ai-anomalydetector-java
microsoft
使用适用于 Java 的 Azure AI 异常检测器 SDK 构建异常检测应用程序。在实现单变量/多变量异常检测、时间序列分析或 AI 驱动的监控时使用。
development
azure-ai-language-conversations-py
microsoft
使用azure-ai-language-conversations Python SDK实现对话语言理解(CLU)。当使用ConversationAnalysisClient分析对话意图和实体、构建NLP功能或将语言理解集成到应用程序中时使用。
development
azure-ai-ml-py
microsoft
Azure Machine Learning SDK v2 for Python。用于机器学习工作区、作业、模型、数据集、计算资源和管道。 触发词:“azure-ai-ml”、“MLClient”、“工作区”、“模型注册表”、“训练作业”、“数据集”。
development