AI assistant & actions

Lar’s AI features are optional additions to a local workspace. They include AI Chat, Object actions, media analysis, semantic retrieval, and consent-aware Workflow nodes.

Configure AI

On desktop, open Settings → AI Provider:

Choose Anthropic, OpenAI, Gemini, xAI, a compatible endpoint, or local Ollama as offered by the build.
Enter the provider key or local endpoint.
Choose a model and run Test.
Review usage/spend controls before using automatic actions.

Keys are sealed through host secure storage and are not exposed to the renderer as ordinary settings data. Web and mobile use a configured server-side AI backend and therefore do not show desktop BYOK controls.

Use AI Chat

Open AI Chat with Mod+Shift+A or its navigation action. Ask a question or request a transformation. Lar assembles context locally from explicit context, search, graph connections, and—when configured—semantic retrieval.

Assistant responses can include citations that navigate back to source Objects. Tool-activity chips disclose workspace searches, Object reads, and other tool calls. Supported builds can optionally enable provider web search; keep it off when outside information is unnecessary.

Conversations live in memory until you save one as a Chat Thread Object. Saving makes the conversation ordinary workspace content subject to local storage, search, sync, sharing, and deletion rules.

Object actions

Open an Object’s actions to use available commands such as:

Generate or improve a title
Summarize body content
Suggest Tags or Collections and apply a chosen subset
Suggest related Objects or relations
Fill structured properties
Transform selected text
Analyze an Image, PDF, Audio recording, or Weblink when the provider supports the required capability

Generated output is a proposal. Review it before applying changes, especially property values and links.

Semantic search

Desktop can build a local semantic index using a supported embedding model. It stays with the encrypted desktop workspace. Search combines text, connected Objects, and semantic matches.

If no embedder is configured or indexing has not completed, search and AI retrieval continue with full-text and graph sources. Web currently skips the vector arm and uses its available local retrieval paths.

Agentic tools and Workflows

AI Chat can use bounded workspace tools. Workflows can contain classify, extract, summarize, identify, or enrich actions where supported. An automatic Workflow with AI requires standing consent; otherwise it becomes a reviewable suggestion rather than silently dispatching content.

Keep sensitive Objects out

Choose Exclude from AI on an Object to remove it from assistant retrieval, broker dispatches, and agent-facing AI boundaries. Exclusion does not retract content already sent to an external provider.

Read AI consent & privacy before connecting a cloud provider.