Skip to main content
Chat is the primary way to interact with AI in Runchat. The agent composer sits at the bottom centre of the editor and gives you access to an AI agent that can use tools, run generative models, search the web and create nodes directly on the canvas.
The editor with the Agent panel open on the right, showing a conversation with its tool steps

What you can do

The agent is particularly well suited for:
  • Research: Ask the agent to search the web, summarize findings and compile results
  • Design exploration: Describe a concept and iterate on it through conversation, generating images and variations as you go
  • Workflow creation: Describe what you want to build and let the agent set up nodes on the canvas
  • Code generation: Ask the agent to write HTML pages, React components, or scripts for CAD plugins
  • Quick generation: Generate a single image, text or other output without building a full workflow

Default settings

The agent starts on GPT 6 Luna with low reasoning effort and Edit canvas permission, so you can send your first message straight away. Change the model and its settings from the agent menu, or type /model in the composer. To run the agent from your Claude or ChatGPT subscription instead, see Running the agent from your own subscription.

Starting a chat

Click the composer at the bottom of the editor, type what you want and press Enter (or click Send). Use Shift+Enter for a new line. You can ask questions, give instructions, or describe what you want to create. Sending a message opens the Agent panel on the right edge of the canvas. Drag its left edge to resize it. Press Esc to collapse it; the session pill above the composer reopens it, and its trash button (End agent) stops and closes the agent. Click Stop generating to stop a response early. Each step in a response shows what the agent did. Click a step to expand its details, then Copy them if you need to. Steps that changed nodes have a Show on canvas button, and steps that spend credits show a credits badge. A step only shows an icon while it is running or if it failed.
You can also run a tool loop or single turn of a conversation from the Agent node. This is useful for automating agents within workflows, and for using multiple sub-agents from chat.

The agent menu

Click the chat title at the top of the Agent panel to open the agent menu:
  • New chat: start a fresh conversation
  • History: recent chats in this workflow, plus Search all chats to search chats in this Workflow or All workflows. Opening a previous chat restores its messages, tools and model.
  • Model: : choose the model the agent uses
  • Other agents: switch between agents running in parallel
  • End agent: stop the current agent if it is still working, and close it. The chat stays in History.
The agent menu open from the chat title, with the History submenu listing recent chats and Search all chats

Model selection

Open the agent menu and click Model: , or type /model in the composer and press Enter, to open Choose a model. /model works before you send your first message, so you can pick the model for a new agent without opening the panel. It lists the text models you have installed, with each model’s credit price. Filter by quality tier (Any, Draft, Balanced or Quality), search by name, or install another model with Add Fal or Add OpenRouter. Models that draw on your free monthly pool rather than credits are marked with a Free badge.
The Choose a model picker listing agent models with their prices, quality tier filters and Add Fal and Add OpenRouter buttons
The selected model persists within the current chat session. When you restore a previous chat, the model used in that session is restored too.

Model settings

Picking a model opens its settings card above the composer. This is where you set how hard the agent thinks and how many credits it can spend. Models that spend credits also show a credit limit per response. Free models skip it, since they don’t draw credits. Every control applies as you set it, so there is nothing to save. Close the card with the ✕, or just send your message and it closes itself. The credit limit saves when you click out of the field, and again if you close the card while still typing in it.
The model settings card above the composer, with permissions, reasoning effort and a credit limit

Reasoning effort

Each model declares its own default Reasoning effort, and switching models adopts it. Most default to a low level so the agent acts quickly. Raise it, up to Max, for complex requests where quality matters more than speed, or turn it Off to have the model answer immediately.

Credit limit per response

On models that spend credits, the agent can use unlimited credits in its responses by default. Type a Credit limit to cap this. The model’s average spend per response is shown underneath as a guide. Once the agent reaches this limit it immediately returns its response and stops reasoning, searching or using other tools. Clear the field to go back to no limit.
A single agent step costs around 25 credits, so a lower limit would stop the response before it starts. Runchat won’t save one. If a stored limit is too low to send, the settings card reopens and tells you so.

Running the agent from your own subscription

The top of Choose a model has three ways to run the agent without spending Runchat credits:
  • ChatGPT plugin: install the Runchat plugin in ChatGPT and drive the agent from your ChatGPT subscription.
  • MCP: connect Claude or another MCP client to Runchat over MCP.
  • OpenRouter key: bill model calls to your own OpenRouter account. See Linked Accounts.
Working in a coding agent such as Claude Code or Codex? Use the Runchat CLI instead of MCP. It exposes the same tools as shell commands, so there’s no MCP server to configure.

Permissions

The agent works at one of three permission levels, set from Permissions in the model settings card:
  • View only: The agent can browse and inspect the canvas, discover models, and answer questions, but cannot create or modify anything.
  • Edit canvas: The agent can create, connect, organize and update nodes on the canvas. It cannot run nodes; it presents the completed workflow for you to run.
  • Edit and run: Full access. The agent can create, connect, organize, and also run workflows on your behalf.
The agent starts on Edit canvas. When the agent needs more access than it has, it asks with a Permission needed card. Choose Allow editing, Allow editing and running nodes, or No thanks. For the full list of tools available at each permission level, see Agent Capabilities.

Agent teams

With agent teams turned on, the agent runs on Runchat’s servers rather than in your browser tab. Two things change:
  • The agent keeps working when you close the tab. Reload the page or come back later and the chat picks up where the agent got to. The Agents page of the dashboard lists every agent still running, on any workflow, and can stop them.
  • The agent can bring in teammates. For a bigger job it can split the work between specialist agents that run at the same time, each with its own brief.
Agent teams are experimental and on by default for Hobby plans or higher. To turn them off, open the account menu, choose Preferences and switch off Agent teams (experimental) under Features. Teams work in chats on a saved workflow.

Starting a team

Ask for a team (“use a team to research these five sites”), or let the agent suggest one. A team uses more credits than one agent working alone, so unless you asked for a team the agent checks with you first: it names the teammates it would start and offers to work alone instead. The agent you chat with is the lead. It can start up to five teammates, each one a specialist:
  • Researcher: finds and reads sources, and gathers notes and reference images
  • Designer: writes prompts, picks models and generates images and video
  • Modeller: builds 3D and CAD geometry, in Rhino, Grasshopper or Blender when connected
  • Coder: writes and runs code, from code nodes and apps to data analysis
  • General: the full agent, for anything else
Teammates use the lead’s model and permission level. Each has its own credit budget (500 credits unless the lead sets another), and reports back to the lead when it finishes.

Following the team

While a team is active, a Team section appears at the top of the Agent panel:
  • Each agent has a card with its name, what it is doing and the credits it has used. Click a card to see that agent’s activity, or to send it a message directly.
  • Tasks lists the work the lead has handed out and who has it.
  • Messages shows what the agents have said to each other.
Each agent also has its own glyph, which sits on the node it is working on so you can see where the team is on the canvas. Click Stop all to stop every running agent in the team. To clear the team’s agents, tasks and messages, open Team options and choose Dismiss team. This clears the teams from every chat in the workflow, and keeps your chats and everything on the canvas. Teams can also use the agent workspace, a Linux sandbox for running Python and Node.

Learning about Runchat

The agent knows about your canvas and the Runchat interface and can help teach you how to use Runchat. Explain what you want to do and have the agent walk through how to do it, or build the workflow for you. If you’re unsure what a node does or why it works in a particular way, select it and ask the agent to explain its behaviour.

Running generative models

Ask the agent to generate images, videos, 3D models and other media, or perform any other task you might take on the canvas. The agent selects an appropriate generative model for your request and creates nodes on the canvas using that model. You can then change the node in any way (e.g. edit the prompt or swap the model) and click Run in the node header to run it like any other node. The agent selects from the models you have installed. To install any model from Fal, open the node picker and click Add Fal. See Models. You can also name a specific model in your message.

Using tools

The agent searches for and uses the tools it needs on its own. This includes web search, page fetching, any tool you have added to your nodes, and tools that have been published and approved by the Runchat admins. See Tools and Apps. The agent can inspect a tool’s workflow to learn from it and make changes (for instance to adapt an example to your specific design task), and run the tool directly when it has permission to run nodes.

Attaching images and context

You can give the agent visual context in several ways:
  • From your computer: Paste or drop an image onto the composer.
  • From the canvas: Select nodes on the canvas. They appear as chips on the composer and are sent with your next message. Selected image and PDF nodes attach their files the same way.
Selected images and PDFs are cleared after each message is sent. To speak a message instead of typing, click Dictate. Typing at any point stops the recording.
The agent composer with selected canvas image nodes attached as chips
Ask the agent to search the web to look up information, find reference images, and research topics. See Agent Capabilities → Custom Tools for patterns that work well.
Each web search costs 10 credits, and reading a page costs 1. The agent’s searches always run on Runchat’s keys; search keys you link are used by Code nodes instead.

Creating nodes on the canvas

The agent can create and configure nodes directly on your canvas. Describe what you want to build and the agent creates the appropriate nodes and configures their inputs and settings. Some things that work particularly well:
  • Asking the agent to generate custom nodes with a specific interface (e.g. a tool for drawing sketches or adjusting image colours)
  • Asking the agent to generate websites and portfolios
  • Asking the agent to create interactive charts and visualisations
  • Asking the agent to create variations on all of the above as multiple nodes
For the underlying tools the agent uses to read, create, connect, and run nodes, see Agent Capabilities.

Debugging with the agent

If a node errors, select it and ask the agent to take a look. The agent can read the error, inspect upstream inputs, and in many cases apply a fix directly. See Running and Debugging.

Skills

The agent can load domain-specific skills to get specialized knowledge before creating code nodes. Available skills include guidelines for Rhino Python and C#, Blender Python, Revit C#, HTML and JavaScript. The agent loads these automatically when creating code for these environments.

Next steps