A virtual concierge for Hotel ZGY, a fictional hotel in Wynwood, Miami. Built with .NET 11, ASP.NET Core, Aspire, Temporal, and Deepgram’s Voice Agent API. Guests can use text or browser voice to ask about the hotel, check weather, find nearby places and more
- .NET 11 SDK (11.0.100-rc.3 or later; pinned in
global.json) - Docker
- just
- Aspire CLI
- Dev Tunnels CLI
- Ollama for local models,
- Microsoft Foundry
- Geoapify account and API key
- Deepgram account and API key with Voice Agent API access
Run all commands below from the repository root.
The repo includes a concierge-setup agent skill
that checks prerequisites, walks you through each account and key, and starts and
verifies the app. Open the repo in your agent and ask it to set up the project,
or invoke the skill directly:
| Agent | Invoke |
|---|---|
| Claude Code | /concierge-setup |
| Codex | $concierge-setup, or pick it from /skills |
| GitHub Copilot | Ask to "use the concierge-setup skill"; Copilot selects skills from your prompt |
The agent never handles your keys. When it asks for one, copy the command it
gives you, replace <your-key> with your key, and run it in your own terminal:
dotnet user-secrets set "Parameters:<parameter-name>" "<your-key>" --project src/Concierge.AppHostWithout an agent, run the setup check to see what's missing:
dotnet run --file .agents/skills/concierge-setup/scripts/check-setup.csThen follow the account steps and the sections below.
Development defaults to Ollama for both chat and embeddings.
ollama pull qwen3.8:27b-mlx
ollama pull embeddinggemma-2:740m
ollama create concierge-chat -f models/ollama/Modelfile.concierge-chatThe Modelfile sets a 32,768-token context. Load the chat model before trying voice to avoid its initial load delay.
To use Foundry instead, configure it below and skip the local models for whichever capability you switch.
Geoapify for mapping features and is required at startup:
dotnet user-secrets set "Parameters:geoapify-api-key" "<your-key>" --project src/Concierge.AppHostDeepgram key is required for voice features:
dotnet user-secrets set "Parameters:deepgram-api-key" "<your-key>" --project src/Concierge.AppHostKeep keys in the AppHost user-secrets store, not checked-in settings.
aspire secret set "<key>" "<value>" --apphost src/Concierge.AppHost
Once per machine:
dotnet dev-certs https --trust
devtunnel user loginAspire creates the public voice callback tunnel and supplies its URL automatically. The web UI uses a separate local HTTPS endpoint.
aspire run --apphost src/Concierge.AppHost/Concierge.AppHost.csprojFor an Aspire-managed background run:
aspire start --non-interactive --apphost src/Concierge.AppHost/Concierge.AppHost.csproj
aspire wait web --status healthy --timeout 180 --apphost src/Concierge.AppHost/Concierge.AppHost.csproj
aspire describe --apphost src/Concierge.AppHost/Concierge.AppHost.csprojStop only this AppHost without resetting its persistent data:
aspire stop --non-interactive --apphost src/Concierge.AppHost/Concierge.AppHost.csproj- Open the Aspire dashboard and wait for the services to become healthy.
- Open the web endpoint.
- Check in using a demo guest's last name and room number from the table below.
Startup prepares the database, demo stays, and Markdown knowledge base.
To restart one guest's demo, use Start over in the chat header. It deletes that stay's
conversation and rides and keeps you checked in. Use just reset only to wipe every stay.
These are seeded demo stays. Use any row's last name and room number to verify the stay and start a concierge chat.
| Guest last name | Room number |
|---|---|
| Rivera | 412 |
| Morgan | 812 |
| Phillip | 112 |
- Ask a hotel question: “What time is checkout?”
- Check the weather: “What's the weather tomorrow?”
- Ask for nearby cafes, then request a ride to “that coffee shop.”
- Use the voice control for a spoken conversation and allow microphone access when prompted.
Development defaults use local Ollama models. Settings, defaults, and validation rules for every service are in the configuration reference. Secrets come from the AppHost user-secrets store.
Commonly changed settings:
| Key | Default | Purpose |
|---|---|---|
AI:Chat:Provider / AI:Embeddings:Provider |
Ollama |
Ollama or Foundry. |
Hotel:Address |
2217 NW Miami Court, Miami, FL 33127 |
Anchor for weather, routes, and rides. |
Deepgram:ListenModel |
flux-general-en |
Speech-to-text model. |
Deepgram:SpeakModel |
flux-hannah-en |
Text-to-speech voice. |
Deepgram:Speed |
1.0 |
Speech rate. |
For chat, set the service endpoint and deployment name:
dotnet user-secrets set "AI:Chat:Provider" "Foundry" --project src/Concierge.AppHost
dotnet user-secrets set "AI:Chat:Endpoint" "https://<resource>.openai.azure.com/" --project src/Concierge.AppHost
dotnet user-secrets set "AI:Chat:Model" "<chat-deployment>" --project src/Concierge.AppHost
dotnet user-secrets set "AI:Chat:RequestTimeoutSeconds" "15" --project src/Concierge.AppHost
dotnet user-secrets set "Parameters:foundry-chat-api-key" "<your-key>" --project src/Concierge.AppHostTo switch embeddings too:
dotnet user-secrets set "AI:Embeddings:Provider" "Foundry" --project src/Concierge.AppHost
dotnet user-secrets set "AI:Embeddings:Endpoint" "https://<resource>.openai.azure.com/" --project src/Concierge.AppHost
dotnet user-secrets set "AI:Embeddings:Model" "<embedding-deployment>" --project src/Concierge.AppHost
dotnet user-secrets set "Parameters:foundry-embeddings-api-key" "<your-key>" --project src/Concierge.AppHostThe embedding deployment must support 768 dimensions. Without an API key,
the app uses DefaultAzureCredential.
Docker must be running for the integration tests. Run just to list all recipes.
just test
just live-place-weather # Geoapify/Open-Meteo check; uses the AppHost user-secret key
just live-concierge-actions # action checks against the running app
just eval-validate
just eval-run # saves a new result; optional run ID: just eval-run my-runRun a single live-check scenario against the running app, or selftest to check
the harness offline:
dotnet run --file tests/Concierge.LiveChecks/concierge-actions-smoke.cs -- <scenario>Scenarios: nearby, nearest, weather-now, ride-conversation, selftest.
Live checks keep the fictional guest, stay, and ride rows they create.
Browser audio worklet tests require Node.js:
node --test tests/voice-play-worklet.test.mjs.
| Problem | Action |
|---|---|
| Tunnel authentication requested | Run devtunnel user login, then retry startup. |
| HTTPS certificate not trusted | Run dotnet dev-certs https --trust. |
| Hotel location fails at startup | Check the Geoapify key and Hotel:Address; inspect WorkflowServer logs in Aspire. |
| Missing AI endpoint/model | Set both values for the selected provider; Foundry uses deployment names. |
| Timeout setting rejected | Keep chat at 18 seconds or less and retrieval embeddings at 8 or less. |
| Demo stays expired or a schema/workflow change requires fresh data | Reset local data using the command below. |
Reset deletes local Postgres data and Temporal history. Use it when you want a fresh demo, then start the app again:
just reset
aspire run --apphost src/Concierge.AppHost/Concierge.AppHost.csproj