Agent
Agent. A chat surface over the shared AI broker: run a local Ollama model or bring your own key for Anthropic, OpenAI, Gemini, or GLM. The agent builds real things with you — notes, tasks, events, whole databases, even apps — proposed as cards you approve before anything is saved. Conversations are vault objects, answers cite the objects they drew from, and the agent can only act within capabilities you granted.
What it can do
Section titled “What it can do”- Propose & approve — ask for a note, task, event, bookmark, or contact and the agent drafts it as a card in chat. Edit a field, approve what’s right, discard the rest. Nothing is written to your vault until you approve it — even a hidden instruction in a document can only draft a card, never save one.
- Propose a database row — or design a whole new database — the same way: as a card you review before anything lands.
- Draft a reply as an email and hand it straight to the Mailbox composer, ready to send.
- Ask it for an app and it drafts the manifest and code as files you read and approve, then install from your vault.
- Hybrid retrieval grounds every answer in your vault, with citations that open the source object.
- Attach context explicitly: @-mention documents and people, drop in text files, or add images for vision models.
- Long-term memory is opt-in and off by default — every remembered item is visible, redactable, and deletable.
- “Summarize with the agent” and “Ask the agent about this” appear on objects across the other apps.
- Save a conversation as an automation: its tool trace becomes a workflow, opened in the builder for review.
- Fail-closed capability model: the agent’s tools are always a subset of what you granted the conversation, re-checked on every call.
Related
Section titled “Related”- Apps & permissions — the same grants govern agents
- Your data & security — why local AI matters