Azure AI Proxy

A managed, multi-tenant proxy that sits between workshop attendees and Azure AI services, giving organisers full control over access, capacity, and usage tracking. Run multiple workshops simultaneously with full data isolation between events.

The solution documentation is published here.

Architecture

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#4a90d9', 'primaryTextColor': '#ffffff', 'primaryBorderColor': '#2c5f9e', 'lineColor': '#4a90d9', 'secondaryColor': '#a8c8e8', 'tertiaryColor': '#dce9f5'}}}%%
graph LR
    Attendees --> Reg[Registration Portal]
    Attendees --> T1[Foundry Toolkit]
    Attendees --> T2[SDK Clients]
    Attendees --> T3[REST Clients]
    Attendees --> T4[MCP Clients]
    Attendees --> T5[Microsoft Agent Framework]
    Attendees --> T6[Foundry Agent Service]

    Organiser[Event Organiser] --> Admin[Admin Portal]

    Reg --> Azure_AI_Proxy
    T1 --> Azure_AI_Proxy
    T2 --> Azure_AI_Proxy
    T3 --> Azure_AI_Proxy
    T4 --> Azure_AI_Proxy
    T5 --> Azure_AI_Proxy
    T6 --> Azure_AI_Proxy
    Admin --> Azure_AI_Proxy

    subgraph Azure_AI_Proxy[Azure AI Proxy]
        direction LR
        P1[Auth]
        P2[Rate Limiter]
        P3[Usage Metrics]
        P4[Event Management]
    end

    Azure_AI_Proxy --> A1[Foundry Models]
    Azure_AI_Proxy --> A2[Foundry Agent Service]
    Azure_AI_Proxy --> A3[Azure AI Search]
    Azure_AI_Proxy --> A4[MCP Servers]

Loading

Broad AI Service Support

• Foundry Toolkit integration for hands-on model experimentation
• Azure OpenAI chat completions & embeddings (including streaming)
• Azure AI Foundry Service Agents (assistants, threads, files, conversations, responses)
• Azure AI Search pass-through for RAG scenarios
• MCP Server endpoints with streamable HTTP transport

Event & Attendee Management

• Time-bound events with start/end windows — API keys only work during your workshop
• Self-service attendee registration via GitHub OAuth or shared codes (great for in-person sessions where not everyone has GitHub)
• Per-event resource assignment — choose exactly which models each event can access
• Full admin portal for creating events, managing resources, viewing metrics, and backup and restore

Capacity Controls

• Daily request cap per attendee — prevents any one person from consuming all capacity
• Max token cap per request — stops runaway token usage

Security

• Attendees never see your real Azure API keys or endpoints

• Encrypted storage for all sensitive configuration (AES encryption)

• Managed Identity support (eliminate API key storage entirely with RBAC)

• This update streamlines how the Foundry Agent Service operates by focusing on security and identity management:

  • Managed Identity Integration: Automatically maps Foundry Agent Service Managed Identity requirements to the Event API Key, ensuring seamless authentication.

  • Object Ownership Isolation: Enhances privacy by restricting access so attendees can only interact with their own agents, threads, and files.

Security Architecture

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#4a90d9', 'primaryTextColor': '#ffffff', 'primaryBorderColor': '#2c5f9e', 'lineColor': '#4a90d9', 'secondaryColor': '#a8c8e8', 'tertiaryColor': '#dce9f5'}}}%%
graph LR
    subgraph Clients["Clients"]
        admin_user["Admin User"]
        attendee["Attendee"]
        foundry_toolkit["Foundry Toolkit"]
        sdk["SDK Client"]
        rest["REST Client"]
        mcp["MCP Client"]
    end

    subgraph Apps["Application Layer"]
        reg["Registration App<br/>(Static Web App + Auth)"]
        admin["Admin UI"]
        proxy["Azure AI Proxy"]
    end

    subgraph Backends["Azure Backends"]
        storage[("Table Storage")]
        ai["Azure AI Foundry / OpenAI"]
        ext["MCP / AI Search<br/>(optional)"]
    end

    attendee -->|"GitHub OAuth"| reg
    reg -->|"x-ms-client-principal"| proxy
    foundry_toolkit -->|"api-key / bearer over TLS"| proxy
    sdk -->|"api-key / bearer over TLS"| proxy
    rest -->|"api-key / bearer over TLS"| proxy
    mcp -->|"api-key / bearer over TLS"| proxy
    admin_user -->|"Entra ID or local admin auth"| admin

    admin -->|"Managed Identity"| storage
    admin -.->|"Internal cache invalidation"| proxy

    proxy -->|"Managed Identity"| storage
    proxy -->|"Managed Identity or encrypted upstream key"| ai
    proxy -->|"Managed Identity or encrypted upstream key"| ext

Loading

The proxy is the main security boundary. Attendees authenticate through the registration flow or present an event API key to the proxy, but they never receive direct access to Azure AI resources or the organizer's real upstream credentials.

In Azure, the proxy and admin app use user-assigned managed identities with RBAC for storage and AI access. The proxy also enforces event-scoped authorization, time windows, daily request caps, and token caps before forwarding approved traffic upstream.

Reporting & Analytics

• Per-event usage dashboards: request counts, token usage, active registrations over time
• Per-model breakdown of prompt/completion tokens
• Exportable backup of all configuration data

Deployment

• One-command deploy with azd up (Container Apps + Static Web App + Table Storage)
• Docker Compose for local development
• Multi-tenant — run multiple workshops simultaneously with full data isolation

Developer Experience

• Drop-in compatible with Azure OpenAI SDKs (Python, .NET, LangChain, REST)
• Attendees just swap their endpoint URL and use their issued API key
• Registration page shows available models and copy-paste configuration

End-to-End Tests

Playwright tests are available under tests/playwright.

If you need to refresh Playwright dependencies manually:

npm run e2e:install

From the repository root:

npm run e2e:install
npm run e2e:test

Run authenticated E2E tests too:

E2E_RUN_AUTH_TESTS=true npm run e2e:test

Run the interactive UI runner:

npm run e2e:test:ui