Azure App Platform
Azure App Platform is a GitHub Copilot canvas for creating, assembling, and operating multi-tier agentic applications centered on ACA Sandboxes and AI Gateway.
What it does
Azure App Platform guides you from a starter architecture or existing codebase to a runnable workspace in an isolated Git worktree:
- Create a new app — follow a three-step flow (Choose a template → Describe app → Review app) using Sandboxes & AI Gateway Starter, App Service Builder - Vite Starter, App Service Builder - Agent & AI Gateway, Dealership E2E Reference Sample, or Empty Starter App, then materialize an isolated Git worktree.
- Open an app — reopen existing app worktrees (or pick from Recent apps on Home) to edit components in Build, configure Connect, test and inspect in Run and Observe (App Map and Logs), or Deploy.
- Assemble existing projects — map existing source directories and compatible Azure resources into a unified
azure-app.yamlmanifest.
Why use it
- Isolated tool execution — separate untrusted tool and document execution in ACA Sandboxes from the agent, with local process simulation for fast development before cloud sandboxes are configured.
- Governed shared access — route model and tool calls through a configured AI Gateway so policy, rate limits, and credentials stay centralized across agents and skills.
- Zero-install plugin — ships as a self-contained GitHub Copilot plugin with bundled templates and browser assets; no manual
npm installor build step is required to open the canvas.
Quickstart
1. Install in GitHub Copilot
/plugin marketplace add coreai-microsoft/azure-app-platform
/plugin install azure-app-platform@azure-app-platform
2. Open the canvas
Ask Copilot:
open Azure App Platform
From Home, choose Create a new app, Open an app, or Assemble existing projects, then use the workspace tabs (Build, Connect, Run, Observe, Deploy) to customize, test, and operate your app.
Home refreshes recent apps in the background without blocking starting journeys, and built-in starter templates are available immediately while reference samples finish loading. Stacked skeleton cards share a 12px gap across Home, the picker, initial controls, and the workspace shell. Selecting an app opens a themed workspace shell while its source graph loads, and workspace creation streams real progress stages while keeping the submitted review form visible. Loading spinners keep rotating even when reduced motion is enabled so pending operations remain visibly active.
When embedded in GitHub Copilot, the canvas uses the host's page surface and typography throughout the workspace, reserving overlay surfaces for dialogs. Primary actions follow the host's button colors, including hover and pressed states, rather than using link colors as button fills. Live host theme changes apply without reloading; the standalone appearance remains independent.
Catalog discovery bounds concurrent worktree inspection and resolves folder aliases once, retaining the complete inventory without launching Git commands for every worktree at the same time. Matching in-flight source-list requests share one read within a canvas and its authorized-folder scope. Completed results are not cached by the server; retries, changed folder scopes, and creation or mutation checks still perform fresh discovery.
In Build, source previews fill the available pane without nested outer gutters. Collections show the filenames that fit, followed by +N more for only the hidden files. Widening the pane restores filenames and removes the disclosure when all fit. Selecting a filename opens Page view; Stack view, per-file line numbers, and Add building block remain available.
In Run, enter a prompt and attach documents on the invocation screen, with Try a boundary test placed directly below Event triggers. Invoke agent opens a separate execution view containing the execution layers, document policy assessment, tool activity, and response output. These outputs are not shown on the prompt-entry screen. The active layer has a clockwise glowing border trail (or a steady highlight when reduced motion is enabled); execution layers remain stacked directly above the response without a parent summary accordion, and each layer opens inline as an accordion to inspect runtime, harness, policy, or MCP tool details. Startup and request states follow actual readiness and HTTP boundaries. Model and tool steps are not streamed: tool reports and policy stages appear when returned, and an unconfirmed tool call stays Unverified. Back to input returns to the prompt without cancelling execution. Use Clear input to reset the task payload without discarding the latest run. View execution in the input header or workspace history revisits that run, while the Output tab exposes its logs. Starting a valid invocation opens Output alongside execution. The command panel docks on the right at viewport widths of 1100px and above, and below the workspace at narrower widths. Drag its divider or use the arrow keys to resize it; each orientation retains its own size. Compact, label-only panel tabs and workspace controls leave more room for content. Collapse it to keep the current request stage and recognized error and warning line counts visible; expand it again for details. Counts cover only the latest loaded shared, bounded log tail, may include other activity, and are not unique issues or a complete invocation trace. They recognize explicit severity labels, tracebacks, and certain HTTP error statuses, so other problems may be missed. Output is grouped by source, not a timeline. Filter by observed prefix or severity, or pause the displayed lines while polling and counts continue. The New output cue indicates changes without claiming an exact line count.
Connect shows this app's saved gateway, model, and sandbox selection from local configuration before refreshing Azure resources in the background. An unconfigured app shows empty connection controls without waiting for Azure. Discovery failures leave the saved connection visible; applying changes still requires current app status and completed discovery.
The header's More options (three-dot) menu contains Docs and Settings on both Home and the workspace. Docs opens in a new tab; closing Settings returns focus to the menu button.
The runtime action bar groups Stop app, Open app, and Logs as icon-only controls with accessible names and tooltips. Stop comes first with a red outline icon and a divider before the other controls. Open app appears only when the web endpoint is ready. During startup, a spinner marks Waiting for the agent and Invoke agent is disabled; the stop control remains Cancel startup. From a stopped app, Invoke agent still starts the app and waits for its agent endpoint before sending the task. Runtime loading spinners keep rotating when Reduce motion is enabled; other animations and transitions still respect that preference.
3. Run canvas tests (maintainers)
From the repository root:
npm test --prefix canvases/azure-app-platform
node --test canvases/azure-app-platform/plugin-portability.test.mjs
Delete a local app
Choose Delete app from a row's actions in Home or Open, or review deletion in App settings. Deletion requires a stopped app and validates only the selected app, without inspecting unrelated worktrees.
All three entry points share a modal confirmation showing the app name and full worktree path, so apps with duplicate names remain unambiguous. Cancel receives initial focus, Tab and Shift+Tab stay in the confirmation, and Escape cancels before submission. Once submitted, deletion cannot be cancelled or submitted again while pending.
Forced worktree removal permanently loses local uncommitted changes and untracked files. It removes the local worktree and its Canvas-created branch, not Azure resources, source repositories, or cloud deployments.
If worktree removal fails, the app stays in the list. If the worktree is
removed but its managed branch cannot be deleted, the app leaves the list and
the result identifies the remaining branch and the git branch -D '<branch>'
recovery command to run from the Canvas repository. Branch cleanup is never
retried automatically.
Home and Open refresh in place, retain the search query, and return focus to a surviving row's actions or search field. A removed app leaves the cached inventory even if refreshing fails; the refresh error does not hide the deletion result or branch recovery instructions. During a refresh failure, search stays visible but disabled and focus moves to an available recovery control. Retrying Open preserves its search query.