Assistant
The console assistant and what it may do.
What it is
A chat panel that opens on the right of any console page from the sparkles button in the header. It answers questions about Deutsche Cloud and about cloud concepts, in prose, from platform documentation it is given at the start of every conversation. With Read live data switched on for the organization, it can also look up your own resources through the platform API, as you and with your permissions, and it shows which reads it made. It performs nothing on your behalf: everything it produces is text for you to act on yourself.
Concepts
| Term | Meaning |
|---|---|
| Panel | The right-hand column. It shares the window with the page rather than covering it, and stays open as you navigate. |
| Capability | An organization-wide switch. Read live data is the one that can be changed; Answer questions is always on. |
| Documentation context | The platform documentation handed to the assistant with each message. It is the same for every organization. |
| Page context | The path of the page you are on, sent with each message so answers can refer to what you are looking at. |
| History | The conversation, kept in your own browser under the active organization, capped at the most recent 50 messages. |
| Provider | The OpenAI-compatible model endpoint the installation is pointed at. It belongs to whoever runs the platform, not to your organization. |
How to
Ask a question
Press the sparkles button in the header. Type in the box at the bottom: Enter sends, Shift+Enter starts a new line. Answers stream in as they are produced, formatted as Markdown; a link to a console page opens it beside the panel.
While an answer streams, the panel follows it only if you are at the bottom of the conversation. Scroll up and it stays where you are, with a button to jump back to the latest line. The square button beside the box stops an answer, and the model stops with it.
With Read live data off, a line above the box says so and links to the switch.
The assistant is told your organization name, your IAM username, the path of the page you are on, and whether it may read live data. Anything else it knows about your account, it looked up through the API while answering, and the panel shows each read as a chip above the answer.
Clear the conversation
Use the bin icon in the panel header. This empties the stored history for the active organization in this browser. Nothing is stored on the server, so there is nothing else to delete.
Change what the assistant is allowed to do
Settings → Assistant, which opens /settings/assistant. The switch there applies to the whole organization, not to you alone, and changing it needs ai:ConfigureSettings. Without that action it renders read-only. Or:
PATCH /ai/api/settings
{ "aiReadLiveData": true }With live data on, the assistant gets one tool per read operation of the API (the same operations the CLI and the MCP server expose, minus secret values and source code). Each call goes to the route with your own session, so a read you are not allowed is refused for the assistant too, and the audit log records the same caller.
Read the current settings
GET /ai/api/settingsThe response carries all four flags, of which aiFillForms and aiExecuteActions are recorded and have no effect. The call needs ai:Chat.
Reference
Capabilities
| Capability | Default | Effect today |
|---|---|---|
| Answer questions | on, fixed | The assistant replies in prose from its documentation context. No switch turns it off. |
| Read live data | off | The assistant may call the read operations of the API as you, up to six per answer. Secret values and source code are never offered to it. |
Permissions
| Operation | Permission |
|---|---|
POST /ai/api/chat |
ai:Chat |
GET /ai/api/settings |
ai:Chat |
GET /ai/api/assistant-models |
ai:Chat |
PATCH /ai/api/settings |
ai:ConfigureSettings |
ai:Chat is a read, so an allowed message writes no audit entry; a refused one does. Ownership is not a factor anywhere in this flow — an owner may configure the assistant because the administrator policy carries ai:ConfigureSettings, and any policy that carries it works the same way. AI resource names are regionless.
Provider configuration
Set by whoever runs the installation, through its environment:
| Variable | Meaning | Default |
|---|---|---|
AI_BASE_URL |
An OpenAI-compatible chat completions endpoint | none |
AI_API_KEY |
The key presented to that endpoint | none |
AI_DEFAULT_MODEL |
Optional: the model Auto stands for, when the provider's list offers it | none |
One provider serves the whole installation, and no model name is configured: the list comes from the provider's API (AI_BASE_URL/models for what the key may use, AI_BASE_URL/models/all for each model's kind and hosting region), cached for 5 minutes and served up to 24 hours stale if the provider cannot be reached. Which models the key reaches is a setting on the key at the provider. Of those, Deutsche Cloud offers only chat models the provider reports as hosted in Germany or the EU. An organization cannot point the panel at its own endpoint.
Choose the model
Account → AI settings → Model, or:
dcloud ai assistant-models list # the live list, and what Auto is now
dcloud ai settings update --ai-assistant-model gemma-4-31b
dcloud ai settings update --ai-assistant-model auto # back to the default| Value | What a message uses |
|---|---|
auto |
The default. AI_DEFAULT_MODEL if the list offers it, else the model the provider marks recommended, else the first on the list, skipping any that failed the tool check (below). |
| a model id | That model. It must be on the list when you set it (400 otherwise). If the list later drops it, Auto stands in and the settings say so. |
GET /ai/api/settings returns the choice (aiAssistantModel) and the model a message sent now goes to (aiAssistantModelInUse, with aiAssistantModelVia: chosen, auto or fallback). The same choice is used by CI's workflow conversion. Terraform: ai_assistant_model on dcloud_ai_settings.
The tool check. Some models answer nothing once a request carries the assistant's live-data tools (measured: the provider's recommended gpt-6.1-terra and gpt-6-luna return an empty answer with ~275 tools). Before Auto settles on a model, the platform sends it one real request with the assistant's system prompt and read tools, and passes over a model that gives no answer, trying the next candidate. The verdict is kept 24 hours. aiAssistantModelSkipped lists what Auto passed over, GET /ai/api/assistant-models carries each checked model's toolCheck, and the picker labels a failed model "fails with live data". A model you choose yourself is always used as chosen. With Execute actions on, the request carries every write tool too (~730 tools); the provider currently refuses that many for every model, so the assistant does not answer until the tool set is reduced.
What the assistant never gets
- Secret values, key material, and private keys. Revealing a secret is a separate permission on a separate call, and the assistant makes no calls.
- Any record from your account: instances, databases, repositories, policies, invoices.
- Credentials of any kind — no access key, no session cookie is placed in its prompt.
What is not supported
- Filling in forms, or executing actions. The assistant can say what to enter or which call to make; it cannot do either.
aiFillFormsandaiExecuteActionsare stored and change nothing. - Writes of any kind through the assistant. Its tools are the read operations only.
- Per-user assistant settings. All four flags belong to the organization; every member of it gets the same assistant.
- Choosing a temperature or a system prompt. The model is chosen per organization (above), not per user or per conversation.
- Conversation history on the server. It lives in one browser, under one organization, and does not follow you to another machine.
- Attachments, images, or file uploads in the panel.
- Citations or links back to the documentation an answer came from.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| the typing indicator appears, then no answer arrives | no provider is configured for the installation, or its endpoint is unreachable | ask whoever runs the platform to set AI_BASE_URL and AI_API_KEY |
| the assistant says it cannot see your resources | Read live data is off for the organization | turn it on at /settings/assistant |
| a read chip shows an error | the route refused the call, usually a 403 for an action your policies lack | the same call would fail for you in the console; fix the policy |
| the switches will not move | the caller lacks ai:ConfigureSettings |
attach a policy allowing that action |
| the settings page says you have no access to AI features | the caller lacks ai:Chat |
attach a policy allowing that action |
| an answer about your resources is wrong with live data off | it answers from documentation only | turn on Read live data, or read the console page |
| the conversation is gone | history is per browser and per organization, and holds the last 50 messages only | none — start a new conversation |
Actions with confirmation
With Execute actions on (/settings/assistant), the assistant has a tool for every non-sensitive write of the API: creating a hosted zone, adding a domain to an app, deploying, rotating a password, and the rest. A proposed action appears in the chat as a card showing the operation and its exact inputs, with Confirm and Cancel. Nothing runs until Confirm; the call then goes through the same route, IAM check and audit entry as a click in the console, as the person chatting. Sensitive operations (secret values, source code) are never offered. The assistant's product knowledge for guiding a setup lives in docs/ai-context/product-guide.md.