Chat Node

The Chat Node is the primary way to interact with AI models. It supports back-and-forth conversations, context from connected nodes, multiple AI providers — and three modes that shape what each conversation can do.

---

Modes: Agent, Chat, and Image

Every chat node has a mode selector at the bottom left of its input. The mode decides what the conversation is allowed to do:

• Agent (⌘1) — the full experience and the default for new nodes. The AI can use tools: writing to the canvas, spawning sub-agents, calling MCP tools, and proposing its own mode and model switches.
• Chat (⌘2) — a focused conversation without tool clutter. No tool calls and no sub-agents — just you and the model.
• Image (⌘3) — turns the node into an image generation studio. Each prompt generates an image into its own connected Image Node, and the model picker swaps to the image model catalog.

A fourth mode, Video, is visible in the dropdown but marked "Soon".

The mode is saved per node, and you can switch anytime — it only affects messages you send after the change. Slash commands, @ mentions, and personas work the same in every mode.

Don't confuse chat modes with context modes (Summarized vs Isolated, which control incoming context) or with Subscription providers (the Cursor, Claude Code and Codex backends). Agent mode works with any regular model.

---

Sending messages

Type your message in the input field at the bottom of the chat node.

• Enter submits your message
• Shift+Enter inserts a new line (also works inside code blocks and lists)
• ⌘⇧Enter submits your prompt into a new branch node (the Split button — tooltip "Submit in new branch")
• ⌘B creates an empty branch node

You can also click the Send button (up arrow) to submit. While a response is streaming, the Send and Split buttons combine into a single Stop button. When the input is empty, a microphone button appears in the send slot instead (title "Start recording") — see Voice input.

While a response is generating, a warm glow wraps the node's title bar and every edge feeding the node runs a comet beam (Connections), so you can spot activity from across the canvas.

---

Slash commands

Type / in the input to open the command menu. Skills, Personas, and MCPs are grouped into browsable sections — click one (or press →) to drill in, and press ← or Esc to step back out. Reasoning presets (Low / Medium / High / Max) and your prompts library sit alongside them as regular entries, and typing filters across everything at once. See Slash Commands for the full tour.

Inserted references render as chips in the input.

If a referenced item is missing or too large, you'll see an error toast before the message is sent.

@ mentions

Type @ to search the nodes on your canvas by name. Selecting a node inserts it as a chip in your message and creates a persistent connection (edge) from that node, so its content feeds this chat as context — a fast way to wire up context without dragging edges by hand. See Connections for how context flows between nodes.

---

Choosing a model

1. Click the model dropdown in the chat input's bottom toolbar
2. Use the Search models... box to filter, or browse the groups:
   • Rhai (built-in) first — Auto ("Picks the right model for each message") at the top, then the hosted catalog, each model with a one-line description
   • Agent providers (such as Cursor and Claude Code)
   • Your own BYOK providers (bring your own key)
3. Click a model to select it for this node

Auto is the default: it routes each message to the model best suited for it, so quick replies go to fast models and hard problems get the heavyweights. See Models for details.

On the free plan the picker is locked to Auto — picking other models prompts an upgrade. Use the Manage AI Models link in the dropdown to open Settings.

In Image mode this dropdown swaps to the image model catalog ("Search image models...") — see Image Node.

When an agent-provider model (Cursor or Claude Code) is selected, a folder button appears next to the input so you can pick the agent workspace — the working directory the agent operates in.

---

Tools dropdown

Click the + button in the chat input (tooltip "Tools") to configure this conversation:

• MCP — a checkbox list of your MCP servers, enabled per conversation. Add New opens Settings. See MCP Tools.
• Personalities — pick a persona for this chat, or choose Add New to create one inline. See Personas.
• Reasoning — radio options Low / Medium / High / Max for models that support extended thinking. The default is Medium.
• RAG Mode — a switch that turns on Retrieval Augmented Generation for this node.

On the Free plan the RAG Mode switch shows a lock, and clicking it opens an upgrade modal: "RAG Mode Requires a Paid Plan".

In Chat mode the MCP section is hidden — tool calls only run in Agent mode.

---

RAG mode

RAG mode retrieves only the most relevant chunks from connected nodes instead of sending everything, which can dramatically reduce input size. Deep dive: RAG Mode.

While RAG mode is on:

• A scissors icon labeled RAG Mode Active appears in the node footer
• The token pill is masked with asterisks (exact counts aren't known until retrieval runs)
• Status banners report progress: Indexing documents…, Documents indexed, Retrieving relevant chunks…, Reranking results…, and an error banner if something fails
• After a response, a summary shows the savings: "RAG Mode reduced your input size by N% from X to Y tokens."

RAG mode with Rabbitholes managed models is temporarily unavailable in the desktop app — use a BYOK provider for RAG conversations.

You can chat with document nodes immediately after adding them. Only RAG mode requires their processing to be complete.

---

Message actions

Hover over a user message to reveal:

• Copy — copy the message text
• Rerun — re-send the same text
• Edit — opens the Edit Message dialog with Save and Save & Rerun (⌘Enter). Save & Rerun removes all later messages and regenerates the response from this point.

Long user messages collapse behind a Show more toggle so a pasted wall of text doesn't take over the thread.

Assistant messages offer:

• Copy and Regenerate
• A collapsible reasoning section when the model produced extended thinking
• Tool-call cards for any tools the model used
• Source citations, shown as stacked source pills

---

Tool calls and approvals

Each tool the AI invokes appears as a collapsible card with a state of Running, Done, or Error.

If a tool requires permission, the card asks "Allow this tool to run?" with Approve and Reject buttons. Long tool output is truncated with a Show more control.

---

When the agent proposes a switch

In Agent mode, the AI can decide a task would go better under a different mode or model — ask it to "draw this" and it may propose switching to Image mode. The proposal appears as a tool card:

• Switch Mode or Switch Model, showing the target and the agent's one-line reason (e.g. "→ Image — you asked for a diagram")
• Nothing changes until you click Approve; Reject carries on in the current setup
• After approval, the switch applies immediately and the conversation continues under the new configuration — you don't need to resend your request

The agent can only change the mode and the model this way — never providers, keys, or other settings.

---

Context modes

A toggle on the node's left edge controls how content from connected nodes is passed in:

• Summarized context (zap icon) — incoming context is condensed
• Isolated context (diamond icon) — the node ignores summarization and treats context separately

A badge on the toggle shows how many incoming connections the node has.

The toggle is disabled while a response is generating — you'll see "Wait for the response to finish".

---

Token summary

The footer shows a token pill. Click it to open the Token Summary popover, which lists:

• Incoming — tokens flowing in from connected nodes
• Own — tokens from this node's conversation
• Input — what will be sent to the model
• Total

Counts are estimates produced with the OpenAI tokenizer. In RAG mode the pill shows asterisks until retrieval determines the real input size.

---

Voice input

1. Click the microphone button (visible when the input is empty) or press Ctrl+M
2. Speak — an overlay shows a timer and a live waveform
3. Click Done or press Enter to finish, or Cancel / Esc to discard
4. Wait for Transcribing… — the transcript replaces your input text
5. Edit if needed, then press Enter to send

Recording uses your preferred microphone from System Preferences. More: Voice Input.

---

Node controls

The node header and toolbar give you:

• Color picker — tag the node with a color
• Duplicate Node
• Collapse — the minimized view shows the last You/Assistant snippet with a See More button
• Open in Sidebar — read the conversation in the side view
• Delete Node

Chat nodes get auto-generated titles based on the conversation.

---

Errors and sub-agents

• Insufficient credits — an amber banner appears with a Reload usage button
• Plan-gated features — the banner includes an Upgrade button
• Other errors — a red banner with a Retry button

In Agent mode, the AI can also spawn child chat nodes as sub-agents to work on parts of a task. A live activity card shows progress ("Subagent is thinking…") with Jump to node and Cancel actions. See Sub-agents.

---

Use cases

• Research — connect web, document, and post nodes to a chat and ask questions across all your sources at once.
• Writing — draft in a chat, then use ⌘⇧Enter to branch alternative directions without losing the original thread.
• Studying — turn lecture PDFs into a Q&A partner: attach them with @, enable RAG mode, and quiz yourself.
• Content creation — use personas to keep a consistent voice, and Rerun/Edit to iterate on hooks and titles quickly.
• Planning — let the AI spawn sub-agents to explore options in parallel, then compare the branches side by side on the canvas.