QuickstartMoving from RabbitholesCore concepts
Canvas OverviewCreating NodesConnectionsOrganizing the CanvasSearchSide ViewExportKeyboard Shortcuts
Chat NodeText NodeWeb NodePost Node (YouTube & more)Document Node (Beta)Image NodeGroup Node
Models & ProvidersRAG ModePersonasSlash Commands (Skills & Prompts)Using MCP Tools in ChatSub-AgentsVoice Input
Cloud ProvidersAgent ProvidersOllama (Local Models)Custom ProvidersMCP Servers (Beta)Default ModelsConnected ToolsPersona PromptsAccount ManagementSystem Preferences
Troubleshooting & Support

Using MCP Tools in Chat

Let the AI call external tools from your configured MCP servers during a conversation.

Last updated: July 6, 2026

MCP (Model Context Protocol) is an open standard for connecting AI models to external tool servers — things like file systems, databases, browsers, or your own scripts. Once a server is configured, the AI can call its tools mid-conversation and use the results in its answer.

Server setup lives in Settings → MCP Servers (Beta). This page covers using those servers in a Chat Node.

Enabling MCP servers for a conversation

MCP servers are enabled per conversation, so tools are only offered where you want them. Three ways to enable one:

  1. Open the + tools dropdown and check servers in the MCP checkbox list

  2. Or type / and use the MCP quick-toggles in the slash menu

  3. Or click an existing server chip in the chat footer

Active servers show as removable pill chips in the chat footer — click a chip's remove control to disable that server for the conversation.

MCP requires a model that supports tool calling. All Rabbitholes managed models do; for other providers, check the model's capabilities in Models & Providers.

Tool calls in the transcript

When the AI decides to use a tool, a tool-call card appears in the conversation showing the tool's name and its state:

  • Running — the call is in flight, with a spinner
  • Done — the call finished; expand the card to inspect the output, with Show more for long results
  • Error — the call failed; the card shows the error message

The AI can chain several tool calls in one turn — each gets its own card, and the final answer follows once the calls resolve.

Approving tool calls

Some tool calls pause for your confirmation before running. The card asks Allow this tool to run? with two choices:

  1. Approve — the tool runs and the conversation continues with its result

  2. Reject — the tool is skipped and the AI is told the call was declined

This keeps you in control of tools that touch your files, network, or accounts.

Use cases

  • Live data in answers: a database MCP server lets the AI query real numbers instead of guessing, right inside your research canvas.
  • File operations: a filesystem server lets the AI read a project folder and summarize it into connected Text Nodes.
  • Web automation: a browser server fetches pages the AI needs mid-conversation, beyond what a Web Node captured.
  • Safe experimentation: enable a powerful server for one conversation only, approve each call manually, and remove the chip when done.

Slash Commands (Skills & Prompts)

Type / in a chat to toggle tools, set reasoning effort, and insert your own reusable prompts and skills.

Sub-Agents

Let the AI delegate work by spawning child chat nodes that run tasks and report back.

On this page

Enabling MCP servers for a conversationTool calls in the transcriptApproving tool callsUse cases