DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS /// DOCS
AI AGENTS·ONNIE AI

Onnie AI

Onnie AI is your personal agent over the entire workspace. It's always there, always scoped to you, and builds context from your workspace automatically as you work.

What Onnie AI is

Onnie AI is the name for the Superagent in your workspace — a bot with no tool restrictions, a directory of all Onniebots, and the ability to delegate work to them. When you open a new chat and leave the bot picker on the default, you're talking to Onnie AI.

Under the hood, Onnie AI is an engine_bots row with is_superagent = true. It's created automatically when your workspace is provisioned and there is exactly one per workspace. You cannot delete it or create a second one.

Onnie AI can do everything agents can do: read workspace data, create and manage Tasks, generate images, search the web, run code, consult Onniebots, and (with admin or editor role) create Onniebots and update workspace context.

What it can see

Onnie AI loads context from your workspace at the start of each chat thread:

  • Workspace and project onnie_md — freeform markdown context your team has written to describe the workspace or individual projects.
  • Workspace overview — entity counts and inventories (projects, tables, recent pages, open Tasks).
  • Workspace knowledge — corrections, facts, preferences, and patterns extracted from prior conversations and stored in engine_workspace_memory.
  • Your preferences — personal preferences stored in engine_user_memory, scoped to you.
  • Skills — keyword-matched skill documents that load when relevant to your first message.

This context is built once per (workspace, project) pair, cached, and reused across turns. When you mention a project or entity in your message, Onnie AI automatically scopes to that project and rebuilds context if needed.

To explore workspace data dynamically — e.g., look up a record, list tasks in a project, or read a page — Onnie AI uses the explore_* tools at runtime. It doesn't load the full workspace into every message; it reads what it needs when it needs it.

Chat modes

Onnie AI operates in one unified agent loop. There is no separate "chat mode" or "agent mode" to switch between — every thread runs the same SUPERAGENT loop with the same tool surface. The quality of the response is controlled by two dials:

Tier (fast / pro) — set per chat in the composer settings popover. fast uses a production-quality model optimized for speed; pro uses a more capable model suited for complex reasoning. You can upgrade to high variant per turn.

Variant (standard / high) — a per-turn dial. The standard variant uses the production-quality model for that tier; the high variant escalates to a more capable model for harder reasoning. The engine may also suggest a tier upgrade via an inline CTA if it determines a harder model would serve the request better.

The loop runs until the model produces a response with no tool calls, until it pauses for your input via request_input, or until a safety cap is hit (max 30 iterations per turn on pro, 20 on fast).

Context and memory

Onnie AI builds memory over time. After each multi-turn conversation, it extracts durable facts, corrections, and patterns from the turn trace and stores them in engine_workspace_memory (shared with your workspace) and engine_user_memory (scoped to you). A weekly compaction job summarizes older traces into condensed memories.

These memories load automatically into Block 3 of the system prompt on the next conversation — no action needed on your part. Up to 5 items per kind (corrections, facts, preferences) are included each turn.

You can also instruct Onnie AI to remember something explicitly: tell it to remember a fact or preference, and it calls the remember tool to write it immediately. Use recall to search memory by semantic similarity when you want it to look up something specific.

Each chat thread maintains its own conversation history. The thread state is persisted on a Cloudflare Durable Object and checkpointed to R2, so long threads survive DO eviction and reconnects. If a thread gets very long, prior messages stay in the checkpoint and continue to provide context.

Privacy

Each chat thread in Onnie is strictly private to the user who created it.

//NOTE

Your chat threads are never visible to workspace admins or other members. There is no admin override — no role in the system can read another user's engine_chats or engine_messages rows. RLS on the engine_* tables enforces this at the database level.

Workspace-level memories (engine_workspace_memory) are shared across the workspace — corrections, facts, and patterns Onnie AI learns are available to all workspace members. Personal memories (engine_user_memory) are scoped to you and are not shared.

If you want to share something Onnie AI produced — a Task, a page, an image — share that artifact directly. The chat thread itself stays private.