App Service Custom AI Middleware

Python middleware and tools for AI agents hosted behind Azure App Service Authentication and App Service Hooks.

Install

pip install aw-appservice-custom-ai-middleware-extension

The distribution name is aw-appservice-custom-ai-middleware-extension. The stable Python import remains:

import app_service_extensions as ase

For Azure Functions:

pip install "aw-appservice-custom-ai-middleware-extension[functions]"

For Azure Functions hosted skills:

pip install "aw-appservice-custom-ai-middleware-extension[hosted-skills]"

The hosted-skills integration patches runtime seams that are not yet public API. Its dependency range is intentionally narrow and is expanded only after the integration suite passes against a new runtime release.

Azure Functions

Configure the middleware before creating the hosted-skills app, then attach it to the generated Function app:

import app_service_extensions as ase
from azure_functions_agents import create_function_app

ase.configure()
app = create_function_app()
ase.configure(app)

Or use the convenience factory, which runs that same sequence internally:

import app_service_extensions as ase

app = ase.create_function_app()

Both initialization styles remain supported. Choose one style for an app rather than combining them.

The package preserves the opaque X-MS-Hook-Invocation recursion chain for same-site callbacks. It is not sent to third-party MCP servers.

ase.current_request().request.origin exposes the proxy-aware, client-facing site origin supplied by App Service Hooks for callbacks to the originating site.

Application authentication

Outbound authentication uses the application's configured credentials. Authenticated MCP connections retain the Hosted Skills runtime's managed-identity default; an anonymous MCP connection remains anonymous. Configure each MCP connection's authentication in mcp.json.

ase.call_app() uses the application credential supplied through ase.configure(presenter=...), or managed identity configured with MCP_AUTH_SCOPE and optionally MCP_AUTH_CLIENT_ID (install the [identity] extra for managed identity). These calls use application permissions, not the signed-in user's permissions. Restrict which application APIs an agent can reach accordingly.

The built-in agent tools only edit and answer the current hook request. The library does not provide delegated Microsoft 365 tools, user-grant scheduling, or Teams/deferred trigger integrations.

Publish to PyPI

Publishing is manual because this repository does not have an Actions runner. PyPI does not allow replacing an existing release, so first update the version in both pyproject.toml and src\app_service_extensions\__init__.py.

Save the token once from the repository root in PowerShell:

.\scripts\publish-pypi.ps1 -SaveToken

This prompts for the token, stores it outside the repository at $env:LOCALAPPDATA\AppExtensionsLibrary\pypi-token.dpapi, encrypted by Windows DPAPI for the current user and machine, and exits without publishing.

To publish, run:

.\scripts\publish-pypi.ps1

The publish command verifies that both version declarations match, installs the build tools, runs the tests, rebuilds a clean dist directory, validates the artifacts, and uploads them using the saved token.

Remove the saved token with:

.\scripts\publish-pypi.ps1 -ForgetToken

The script exposes the decrypted value only to the twine child process and restores the process environment afterward. Do not save the plaintext token in the script, commit it, or include it in an issue, pull request, or chat.

After the upload succeeds, tag the exact commit that produced the package:

$version = python -c "import tomllib; print(tomllib.load(open('pyproject.toml', 'rb'))['project']['version'])"
git tag "v$version"
git push origin "v$version"