Skip to content

Letting a model use your files

overcatbe edited this page Oct 25, 2026 · 1 revision

Letting a model use your files

A persona can be allowed to read files of this machine and to write text files. It can never run commands: when something has to be run or changed in the system, it explains and gives you the command, and you run it.

Switching it on

  1. Settings tab, File access for the AIs: the folders it may read, the folders it may also write in, and whether it may read system files (/etc, /var/log, /proc and the like, so it can look at the machine and explain).
  2. Personas tab, Files: per persona No access, May read, or May read, and write after asking you each time. Save; it counts from the next message.

Without both, a persona has no access.

Each persona's own folder

Every persona has a folder in your AIcove folder, ~/AIcove/personas/<persona>/. Two parts of it are for the persona itself, when it has file access:

  • knowledge/: material this persona works from. It is used in every chat even without file access (see Knowledge in the topic Chat); with file access the persona can also read it with its file tools.
  • notes/: where it writes down what it should still know in a later chat. That is its memory between chats; you can read, change and delete those files like any others.

Its own folder… next to Files on the Personas tab opens it. Other personas and the plain model can't read it, and the persona itself can't touch the rest of its folder (its persona file, its chats). Writing there needs May read, and write and your Confirm each time, like every write. It reads plain text (txt, md, csv, code); not PDF or office files. A persona doesn't look into its folder by itself in every chat: say "check your notes", or put a line into its prompt such as "At the start of a chat, look into your own folder."

It knows this guide

A persona with file access can also read this guide and your notes file (the My notes topic), topic by topic, without any folder being allowed for that. So you can ask it how something on this machine works or what to do, and it looks it up first. Keep the notes file free of passwords.

What happens when it writes

A window opens: which file, how many lines change, and the change itself. Nothing is written until you press Write it; Don't (the default) tells the persona no. The old version of a changed file is kept in ~/.cache/aicove/access-backups/.

What it can never do, whatever the settings say

  • Read keys and passwords: the password store, SSH and GnuPG keys, browser profiles, and any file whose name says it holds secrets (*password*, *.key, .env, shadow, ...).
  • Write outside the folders you allowed, into hidden files or folders, or over a file that is a program. Those would be ways to run something later.
  • Change system files, or delete or move anything.

Everything it reads, searches or writes is listed in ~/.cache/aicove/access.log (which file, when, which persona; never the contents).

Good to know

  • Use a model that is good with tools. Coding models and the larger general ones do this well; small creative models often don't call the tools at all, or call them wrongly.
  • What a model has read is in the chat from then on. If web search is switched on in that chat, a model could put text from a file into a search. Don't combine web search with access to things you'd rather keep on this machine.
  • A model can be wrong about what it read. For anything that matters, look at the file.

Made from the guide inside AIcove (cove/data/guide.md); it opens with the ? button.

Clone this wiki locally