Skip to content
Snotra Agent ManualThis manual describes Snotra Agent 1.23.1

What Snotra can and cannot do

Snotra is a desktop app for macOS, Windows and Linux. You open one folder, the file tree stays in view, and the chat works inside that folder with the model you connected. There is no web version and no terminal session: what the model does, it does through the tools below, and each of them can ask you first.

This page is the overview. Every row links to the page that explains it.

AreaWhat Snotra doesRead more
ReadLists folders, finds files by name, searches their text, reads them, outlines code and reads the text of PDF, Word, Excel and PowerPoint filesSet up the built-in tools
ChangeCreates and edits files in the open folder; before a file is replaced as a whole, a copy goes to the trashSee what Snotra changed
RunRuns shell commands and Python code — off until you switch them on, and isolated in a sandbox on macOS and LinuxRun commands in the sandbox
The webSearches the web (with a Tavily key) and fetches a single pageSet up the built-in tools
ImagesDraws an image with your OpenAI key and saves it in the folder; reads screenshots you paste into a message, when the model understands imagesGenerate an image, Write a message
DictationTurns speech into text in the message field, with OpenAI accessWrite a message
MemoryKeeps facts for this folder or for you across chatsLet Snotra remember
SkillsFollows instructions you give it as skills — built in, for one folder or for all of them — and can write a new skill into the open folderUse skills, Let Snotra write a skill
MCP serversUses the tools of MCP servers you connect: tickets, a database, a calendar, anything a server offersConnect an MCP server
ModelsWorks with OpenAI, Anthropic and Google in the cloud, and with Ollama and MLX-LM on your machineConnect a model
Questions about SnotraLooks the answer up in this manual and links the page it usedSet up the built-in tools

What a chat may do on its own depends on its mode and on the risk class of each tool — see Why Snotra asks before it acts.

  • No audio and no video. Snotra generates images, nothing else.
  • No mail, messages, calendar or tickets of its own. That comes only through an MCP server you connect.
  • No Office files or PDFs written. Snotra reads their text, but no tool creates or changes them.
  • No work outside the open folder unasked. A path outside it is asked about on a card in every mode; your home folder as a whole, the root of a disk and Snotra’s own storage are never offered.
  • No commands that keep running. Each command or Python call runs once and ends: nothing is kept between calls, nothing waits for input, nothing stays in the background.
  • No changes to its own settings. The model can suggest switching a skill or tool on for the chat, but only your click on that card does it — the mode, keys and lasting settings stay yours.

You can chat without opening a folder, but the model then has no tools to change or run anything. It can still read the folders of the skills that are on, search the web, fetch a page, use MCP servers and look things up in this manual. A skill has nowhere to keep its data until a folder is open.

LimitValue
Size of a file the model writes or edits2 MB
Images per answer4
Memory per level (this folder, everywhere)8,000 characters
A page fetched from the webone http or https address per call, text only, never an address on your local network
  • The model says it cannot do something this page says it can. The tool is probably switched off or not set up. Click the skills pill in the chat bar: a missing tool is listed with the reason and a button to the setting.
  • A tool is missing although it is switched on. Some tools only appear once they are set up — web_search with a Tavily key, generate_image with an OpenAI key, run_python with Python 3. See Set up the built-in tools.