AI assistant

AuroraDocs has AI surfaces for writing, asking about a page, workspace-level analysis, image generation, and external agent access. The exact options you see depend on your workspace mode, encryption state, configured providers, and plan.

AI Surfaces

Surface	Where	Use it for
Selection actions	editor bubble menu	summarize, expand, rephrase, or translate selected text
Generate with AI	editor slash menu / insert dialog	write new text at the cursor
Ask	right panel	page-scoped chat with the current page, block references, and properties attached
Intelligence	sidebar	longer workspace or document questions with source context
Assistant panel	sidebar/slide-over when enabled	multi-turn workspace assistance
Image generation	cover/image/avatar flows	generate covers, inline images, and workspace avatars

The page-scoped Ask tab opens with Cmd/Ctrl+Shift+A when a page is active.

Intelligence page

The Intelligence surface is for longer questions across one or more objects. From the page you can:

• Choose the text-generation provider inline, so you can switch models for a question without changing your global defaults.
• Attach multiple objects as grounded context — search and sort your workspace, pick the objects to include, and the assistant answers from their content with source references.

In encrypted workspaces the usual E2EE consent rules apply before any content is sent to a hosted provider.

BYOK Providers

Open Settings → AI to add provider keys and choose defaults. Aurora supports separate defaults for text, assistant, image generation, and workspace-avatar flows.

Current BYOK provider families include:

• Anthropic/Claude
• OpenAI
• Mistral
• Gemini
• Scaleway
• OpenRouter
• fal.ai, Leonardo, Together, and image-focused providers where configured

You can also add your own OpenAI-compatible BYOK endpoints (name, base URL, key, default model) in Settings → AI. These custom providers are treated as BYOK usage across assistant chat, editor AI actions, Ask, Intelligence, slash /ai, and Inbox triage.

Custom endpoint safety guardrails:

• production requires https://
• loopback/private/local destinations are blocked unless a local dev override is explicitly enabled
• embedded URL credentials are rejected
• connection tests use a minimal token budget request so you can verify safely before relying on a provider

Provider/model availability comes from AuroraCloud's current model catalogue when signed in, so new models can appear without a desktop app update. If a saved model disappears from the catalogue, Aurora warns you and suggests the provider default.

On desktop, provider keys stay session-only by default. Turn on secure storage in Settings → AI if you want Aurora to remember them in the OS keychain.

AuroraAI

AuroraAI uses Aurora's server-side AI gateway instead of your own provider key. The underlying provider and model are chosen by Aurora on the API server and are not shown in the app. It is available only when all of these are true:

• you are signed in to AuroraCloud
• your plan includes AuroraAI and has budget remaining
• AuroraAI is enabled on the AuroraCloud API host
• the workspace mode and encryption state allow it

The AuroraAI card in Settings → AI explains why AuroraAI is available, unavailable, over budget, outside your plan, blocked by workspace policy, blocked by a non-cloud workspace, or waiting for E2EE consent. Paid plan usage has a daily safety cap equal to 20% of the monthly AuroraAI allowance. The current defaults are Plus at $50/month allowance units ($10/day safety cap) and Pro at $200/month allowance units ($40/day safety cap); operators can adjust those fair-use allowances server-side without exposing the underlying hosted provider. AuroraAI also applies operational fair-use guardrails across the hosted provider account: up to three active hosted requests total, up to 240 hosted requests per minute total, and bounded request/response sizes before the provider call.

In the AI Assistant panel, the provider tabs are explicit. Choosing AuroraAI keeps that assistant request on the hosted AuroraAI gateway; BYOK keys are used only when you switch to a BYOK provider tab.

When an object is open, the assistant treats that object as the default target for phrases like “this page”, “this note”, or “the current object”. Explicit rewrite requests such as “rephrase the text in this note” can replace the current object body under the workspace write-action policy. BYOK tool mode uses the object-content update tool; AuroraAI can perform the current-object rewrite handoff from the panel, with confirmation when the workspace or user policy requires it. Rich editor blocks may be simplified to plain paragraphs for this full-body rewrite path; use the editor selection bubble for smaller formatted passages.

For hosted AuroraAI answers, the web Assistant can run server-side read-only workspace tools before the final response. Mobile can also request a fixed structured tool catalog from the hosted gateway; the mobile client executes those reads and confirmation-gated writes locally, preserving the same native policy and UI as BYOK without exposing AuroraAI provider credentials or model configuration. Tool prompts and results retain source ids so visible citation chips still point back to the underlying AuroraDocs objects.

AuroraAI budget is separate from your BYOK usage budget. See Usage & billing.

Workspace owners can set per-workspace AI policy in Workspace settings → Features:

• allow or block hosted AuroraAI for that workspace
• allow or block BYOK assistants for that workspace
• allow or block custom BYOK providers for that workspace
• for E2EE workspaces, opt in to hosted AuroraAI and require explicit consent before hosted requests
• choose assistant write-action policy (Read-only, Require confirmation, or Allow direct actions) for workspace tool calls

Hosted AuroraAI is purpose-bound to AuroraDocs workspace work and can reject unrelated general-chat requests (for example coding help or broad non-workspace Q&A).

Semantic Workspace Retrieval

When Use semantic workspace retrieval is enabled, Ask and Intelligence can include nearby matches from the workspace knowledge index as secondary context. The same settings area shows a knowledge-health badge so you can tell whether retrieval is ready, indexing, disabled, or limited. If the index falls behind recent edits, AuroraCloud quietly starts a repair and shows the calmer indexing state while it catches up.

Retrieved answers can cite source blocks when available, but you should still review the result. AI can miss context.

Workspace Mode Rules

Workspace mode	BYOK AI	AuroraAI	Server retrieval	MCP
Browser storage	available with selected local context	not available	not available	not available
Local folders / Offline folders	available with selected local context	not available unless explicitly promoted through AuroraCloud	not available	not available
AuroraCloud, not E2EE	available	available when plan/gateway allow it	available when enabled	available with token scopes
AuroraCloud, E2EE locked	unavailable for locked content	unavailable	unavailable	locked/unavailable
AuroraCloud, E2EE unlocked	available for client-selected plaintext context	available only when workspace policy allows hosted E2EE and the user accepts the warning/consent step	server retrieval remains restricted	cannot decrypt server-side content

Simple rule: BYOK AI sends only the context you choose to the provider you configured. AuroraAI and server retrieval require AuroraCloud and follow stricter plan, privacy, and encryption gates.

Image Generation

Use image generation from cover, inline image, or workspace-avatar flows. Pick the provider/model in the generation dialog. Generated images are imported into Aurora storage so covers, avatars, and Image archive entries do not depend on temporary provider URLs.

When Aurora offers multiple cover proposals, choose one final image before saving.

Usage Tracking

Open Settings → Account → Usage & Cost to review BYOK and AuroraAI usage. Signed-in usage history persists through AuroraCloud. AuroraAI plan budget remains separate from generic BYOK usage limits, and the plan-limit table shows both monthly and daily AuroraAI usage.

Privacy Notes

• AuroraAI does not use your prompts, retrieved context, or outputs to train Aurora-owned models.
• BYOK provider calls follow the provider you configured.
• Avoid putting secrets or highly sensitive data into hosted or external-provider prompts unless that matches your own data policy.