AI assistant

The AI assistant is a chat inside RUNOS. It answers questions about your workflows, the log and RUNOS, and it can write, check, test, start and stop workflows for you. It uses the same tools as AI agents over MCP, but you do not need the MCP server or an external program: only an API key of an AI provider.

Turning it on

On a new installation the assistant is already on. The first time you open it, the chat asks for the provider and your API key: choose a provider, paste the key (Get a key opens the provider’s page) and click Save and start. That is all. Without a key nothing goes to a provider.

In the settings (or if you used RUNOS before):

1. Open Settings → AI agents and turn on AI assistant in RUNOS.
2. Under Provider and model, choose your provider. The model field suggests models; you can also type any other model name of the provider.
3. Paste your API key and click Save. Get a key opens the provider’s page where you create one.
4. Click Test under Test connection. “Works” means address, key and model are right.

A round button with the assistant icon now sits in the bottom-right corner of the window, and AI assistant appears in the left bar for the full view.

Using the chat

• Open the chat with the round button. It opens where the button is and stays open on every page (workflows, log, settings, …).
• Move the round button anywhere in the window by dragging it. Move the open chat by dragging its header; resize it at the bottom-right corner.
• Minimize (dash in the header) turns it back into the button, in the place where the chat was.
• Open in full view (arrows in the header) or AI assistant in the left bar shows the chat as a whole page with the chat history on the left. Open as floating window brings it back. It is the same chat: you carry on where you were.
• New chat (plus) starts a new conversation; the previous one stays in the history.
• Answers are formatted (headings, lists, tables, code with a Copy button). An icon under each answer copies its text.
• The assistant knows what you are looking at: with a workflow open in the editor, “this workflow” or “the print node” means that one. It answers in the RUNOS language (Settings → Display language), unless you write in another one.
• Enter sends, Shift+Enter starts a new line. Stop cancels a running answer.

Chat history

Every conversation is saved. Chat history (clock in the header) shows the last ten; you find all of them in the full view, sorted under Today, Previous 7 days and Older, with Search chats (title and content). A click opens a chat; the next message continues it, with everything said before.

With … (or a right-click) on a chat:

• Archive takes it off the list; Archived chats at the bottom shows it again, Restore from archive puts it back.
• Delete removes it for good after asking.

What is saved is what you see and what the provider needs to continue, without attached pictures and PDFs. If you switch to a provider with another format (Anthropic instead of OpenAI-compatible or the other way round), the conversation starts over for the provider; another model or key of the same format continues it.

Attachments

The paper clip attaches files to your next message, the camera a screenshot of the RUNOS window (without the chat). What a message may carry depends on the provider:

Provider	Pictures per message	PDF
Anthropic (Claude)	up to 20, 5 MB each	up to 30 MB
OpenAI	up to 10, 20 MB each	up to 30 MB
Google Gemini	up to 10, 7 MB each	–
Mistral AI	up to 8, 10 MB each	–
xAI (Grok)	up to 10, 10 MB each	–
OpenRouter	up to 10, 10 MB each	up to 20 MB
Groq	up to 5, 3 MB each	–
Together AI, Fireworks AI, Ollama, LM Studio, Other	up to 5, 10 MB each	–
DeepSeek, Cohere	–	–

On top of that, every provider takes text files (TXT, CSV, JSON, XML, LOG, …) up to 200 KB each; they are put into the message. A message carries at most 10 files. Not every model of a provider can see pictures; if one cannot, the provider says so in the chat.

Screenshots

With providers that see pictures, the assistant can also take a screenshot of the RUNOS window itself (tool take_screenshot) to see what you see, for example an error in the editor. It only ever captures the RUNOS window, never the rest of your screen, and not while RUNOS is minimized.

Providers

Provider	Note
Anthropic (Claude), OpenAI (ChatGPT), Google Gemini, Mistral AI, xAI (Grok), DeepSeek, Groq, Together AI, Fireworks AI, Cohere	Key from the provider’s website (Get a key).
OpenRouter	One key for hundreds of models of many providers; type the model as provider/model, e.g. openai/gpt-5.
Ollama, LM Studio	Models on your own PC, no key and no data leaving the PC. Under Server address you can change the address (default http://localhost:11434/v1 or http://localhost:1234/v1). Choose a model that supports tools.
Other (OpenAI-compatible)	Any server with the OpenAI Chat Completions API: enter its address (e.g. https://example.com/v1), the model and, if needed, a key.

Changing the model or key continues the conversation; a provider with another format starts it over (see Chat history).

Cost

The provider charges by tokens (about 4 characters per token). Every request carries the conversation so far, and one question often takes several steps in which the assistant uses tools. To keep this cheap, the tools first return a short overview and details only for what is needed, and long tool results of earlier questions are not sent again. With Anthropic the conversation is cached between steps; OpenAI does that by itself. Smaller models (e.g. gpt-5-mini, claude-haiku-5-5) cost a fraction. For a new topic, click New chat rather than continuing the old conversation.

Safety and privacy

• You confirm every change. Before the assistant saves, runs, starts or stops a workflow, the chat shows what it wants to do with the details and waits for Allow, Allow in this chat or Deny. Reading (workflows, log, building blocks, templates, screenshot) needs no confirmation.
• run_workflow runs the actions for real: mails are sent, files written, documents printed.
• What is sent to the provider: your messages (with the name of the open view or of the workflow in the editor), your attachments and what the tools return (e.g. workflow JSON, log lines, screenshots). Passwords and keys in run results show as ***. With Ollama or LM Studio nothing leaves your PC.
• API keys are stored encrypted (Windows DPAPI) for your Windows user only, in %LOCALAPPDATA%\RUNOS\AssistantKeys.json, never in Settings.json, the log or problem reports. On a terminal server, no other user can read them.
• Keys are only sent over HTTPS; http:// is only accepted for a server on this PC.
• The chat history is stored encrypted (Windows DPAPI) for your Windows user only, in %LOCALAPPDATA%\RUNOS\Assistant\Chats, without attached pictures and PDFs. Delete removes a chat for good. The log only records which tools the assistant used.
• On for new installations, off for existing ones until you turn it on. Without an API key it sends nothing. Turn it off under Settings → AI agents.

Troubleshooting

Message	Cause and fix
“The provider refused the API key.”	Key wrong, expired or for another provider. Paste it again and click Save.
“The provider does not know the model …”	Model name misspelled or not available for your account. Pick one from the list or check the provider’s model list.
“Too many requests, or no credit left …”	Rate limit or empty balance at the provider. A limit that clears within a minute (e.g. tokens per minute on Groq’s free plan) the assistant waits out by itself; if the message still appears, the daily or account limit is reached: try again later, top up your account or choose a bigger plan.
“… cannot be reached”	No internet, proxy/firewall, or the local server (Ollama, LM Studio) is not running.
“… does not take pictures / PDF files here”	Choose a provider from the table above that takes them, or attach the content as a text file.