Assistant

Assistant

This page explains the optional assistant column: connecting an AI host, what the assistant sees, how it asks before changing anything, and how to keep it away from what it should not touch.

The assistant does nothing until you add a host. It works with any AI host that speaks the OpenAI API: LiteLLM, vLLM, Ollama on your own hardware, or a cloud service.

Setting up an AI host

Settings, Assistant, Add an AI host:

• Name and Address of the API (https://llm.example.com/v1).
• Key: From the vault (the password of a Vaultwarden item), Kept on this PC (paste the API key), or No key.
• May see sessions in protected folders: for example your company's own AI host.

Save and test asks the host for its models and shows what each can do (tools, images, context size). Then choose the Model for questions and commands and, for Remote Desktop and VNC screens, the Model for screens. Test reloads the list at any time. The model menu at the top of the assistant column switches host and model per conversation.

Asking about what is on screen

Ctrl+Shift+A shows the assistant column at the right. It talks about the tab you are in. In a terminal tab the last lines of the terminal go with your question (how many: Settings, Assistant, Limits per question). In a Remote Desktop or VNC tab a picture of the screen goes with it, unless you switch Send the screen with it off; the picture is taken from Cablebag's own window, so no other window can end up in it.

Each message shows what went with it under Screen sent with it or Picture of the screen sent with it. A command in an answer can be put at the prompt with Paste at the prompt; it runs only when you press Enter.

Working on a server by itself

In an SSH tab, or in the console tab of a Proxmox VM or container, the assistant investigates by itself with commands that only read: status, logs, disk space, container lists. In a VM it goes through the QEMU guest agent (PowerShell on Windows); in a container through pct exec on the node over SSH. Everything in a guest runs as root, so the first command asks once. In a PowerShell remoting tab it runs PowerShell on the Windows server as the login the tab connected with; there is no root step there.

Cablebag classifies every command, not the AI. Commands that write files, run commands inside commands or cannot be checked count as changes. PowerShell is judged as PowerShell: Get-, Select- and the like read, Set- and Stop- change, Restart-Computer, Format-Volume, Clear-Disk and Remove-Item -Recurse are dangerous.

The cards

Every change waits for you. A card shows the command, why, how to undo it and what kind of change it is, with:

• Run it: once.
• Always "" here: commands of that kind run on that server without asking from now on. The rules are listed under Settings, Assistant, What it may do by itself, where Ask again from now on removes one.
• Refuse.

Dangerous commands (deleting folders, formatting disks, rebooting, stopping SSH, firewall flushes, destroying machines) always get a red card and never an Always. While a command waits for your OK, a bar above the question box says so, with Show it. Stop ends the assistant's work at any moment.

The commands run on their own channel of your SSH login, with exact output and exit codes; your terminal shows each one as a dimmed line.

Ask before every change, or only before dangerous ones

Settings, Assistant, What it may do by itself:

• Read by itself, ask before every change: looking around runs at once; anything that changes something waits for Run it, Always or Refuse.
• Do what is needed, ask only before dangerous commands: changes, root and actions on a screen run without asking; dangerous commands still wait for you.

The top of the assistant column shows the choice, with Change.

Root without sharing your password

When a command needs sudo, Cablebag gives sudo the password from the vault, a saved password or a one-time question, after one click per conversation. The AI never sees it. On servers where sudo works without a password, nothing is asked.

Snapshots before a change in a Proxmox guest

Settings, Assistant, Snapshot of a Proxmox guest before the first change: Offer it, Always take one or Don't. The card for it has Take it and Skip. The token needs VM.Snapshot; see Proxmox.

Protected folders

Right-click a folder in the tree and choose Protect from the assistant. In that folder every command asks first, dangerous commands are refused, and only AI hosts that may see protected folders are used. Settings, Assistant, What it may do by itself lists them; click Lift the protection there (or the folder's menu, Protected for the assistant (click to lift)) to lift it.

Remote Desktop and VNC actions

On a remote screen the assistant can click, scroll, type and press keys to do what you ask. Each action waits for your Do it, or Always on this screen for that server; after its actions it takes a new picture and checks the result. With "Do what is needed" these actions run without asking.

Memory

The assistant remembers what it learns per connection (how a server is set up, what fixed a problem), per folder and in general, inside Cablebag, and starts each conversation with what it knows about that server. New memories show in the conversation with Undo. See and edit them in a session's editor under What the assistant remembers about it, and under Settings, Assistant memory (general notes, and everything it remembers per server and folder, with a search), where you can also add a fact yourself. The memory travels with Sync, encrypted; secrets are hidden before anything is stored.

Secrets stay home

Before anything is sent, Cablebag hides the passwords it knows and text that looks like a secret: private keys, tokens, password= values and password hashes. Saved passwords are never shown to the assistant.

Limits

Settings, Assistant, Limits per question (a fold under What it may do by itself): Terminal lines sent with a question, Commands per question, Minutes per question, Command output it reads (KB), Memory that goes along (characters) and Conversation size (characters). Above the conversation size, older questions lose their terminal text and the oldest turns are left out, so the host does not run out of room.

What was sent

Settings, Assistant log lists every question with the screen text that went with it, every answer and every command that ran, newest first (the newest 2000 are kept). Filter it by server, host or word, Show more for older entries, and Clear the log. A conversation can also be saved as a text file from the assistant column.