AI assistant

The left panel is a conversation with a research assistant that has hands. It can search the web and a dozen specialist sources, read what is already on your board, and create, connect, update and delete nodes itself. You direct; it does the legwork and the typing.

It runs on whichever AI provider you have configured. DeepState AI is ready out of the box.

What it sees

Every message you send goes out with context the assistant can rely on:

• Selected nodes - whatever is selected on the canvas, in full. Select the two articles you want compared and ask; you don't need to describe them.
• Selected connections, with their types and reasoning.
• Recent nodes on the board and a summary of the board's actors, organizations and counts.
• The conversation so far, including previous tool results (large results are cached and can be re-read with get_cached_result).
• Images you paste or attach PRO, for models with vision.
• Custom instructions you have set in Settings → Agent - standing orders such as "always cite the URL in the node description" or "write in British English".

Chips above the input box show the current context nodes; Clear drops them all. Context is de-duplicated between turns so a long conversation does not resend the same node repeatedly.

Add nodes (next to the Thinking toggle, in the merged and agent layouts) searches the board by title or type and adds or removes nodes from the context without leaving the chat. In the agent layout the board sits behind the chat: Peek board (top right) clears the glass so you can click nodes on the board itself - Ctrl/Cmd+click adds more - then Back to chat or Esc returns to the conversation. While peeking, double-click a node (or use Open in the bar at the bottom) to open it in the node sheet, and click a connection to edit it; Esc closes the sheet first.

The assistant does not see other boards, your files, or anything you haven't put on the board or into the chat.

How a turn works

1. You send a message. The assistant reasons (visibly, if reasoning is on) and decides which tools to call.
2. Tool calls stream into the chat as they run: what was called, with what, and a compact result. Expand any call to see the details.
3. It almost always starts with search_board to see what already exists, because duplicates are the main way boards rot.
4. It calls research tools, then creation tools. Nodes appear on the canvas as each call returns.
5. It writes a short summary of what it did. Up to 50 tool steps are allowed per turn by default.

Stop a turn at any time with the stop button; nodes already created stay.

Modes

The three buttons in the panel header change what the assistant may do:

Mode	Can research	Can change the board	Use it for
Plan	✓	✗	Talking through approach, asking what's on the board, getting a plan before committing
Investigate (default)	✓	✓	Normal work: "research X and add it"
Autoresearch	✓	✓	Hands-off multi-round investigation of a topic; new chats only

They are described in depth in Modes.

Reasoning depth

The reasoning toggle next to the mode buttons sets how much the model thinks before acting: off, low, medium or high. Higher settings give better plans for open-ended questions and cost more tokens and time. Low or medium is the sweet spot for most research turns; turn it off for quick edits ("rename this node"). On providers whose models don't expose reasoning the toggle has no effect.

Asking you questions

The assistant has an ask_user tool. When a request is ambiguous - two people with the same name, or "which of these four events do you mean?" - it stops and presents a multiple-choice question in the chat, with a free-text option. Answer and it continues. This is the only place it waits for you; otherwise, in Investigate mode, it acts.

Conversations

• Conversations are per board and kept in the database. The dropdown in the panel header lists them; New chat starts another; the trash icon deletes one.
• Compact conversation (the compress icon) asks the summarizer model to condense the history so far into a summary, freeing context for long sessions without losing the thread.
• The model name and token usage for the turn are shown in the header; provider-level usage is under Settings → AI Provider.
• Max response size (Settings → Agent) caps output tokens per turn, default 150 000.

Tool loading

By default all tools the assistant is entitled to are available every turn. Smart tool loading (Settings → Agent) starts each conversation with only the core set and lets the model pull in groups (entities, research, social, media, narratives) as it needs them via get_more_tools. This trims the prompt for small models and local Ollama models; leave it off for frontier models.

Getting good results

• Name the output. "Create actor nodes for the executives and connect them to the bank with employs" beats "tell me about the bank".
• Point at things. Select nodes before asking about them.
• Ask for sources. The assistant is instructed never to invent URLs; if it has none it omits them. Asking "cite sources" makes it prefer fetch_and_create_newsfeed over paraphrase.
• Use Plan first for big jobs. Agree the shape of the work, then switch to Investigate and say "do it".
• Merge, don't duplicate. If it creates a duplicate anyway, "merge X into Y" makes it move connections and delete the extra.
• Give it a region. "Put the new nodes in the Unverified region" keeps its output where you can review it.

What it will not do

It has no memory across boards. It cannot upload files (create a Document node yourself). It cannot browse behind logins except the social accounts you have connected in Settings → Integrations. It cannot see hidden or filtered nodes any differently from visible ones - filters are visual. And it will not touch a board in Plan mode, whatever you ask.