Snotra 1.15.1 to 1.18: a new name, images, and a sandbox that asks

Since the last round-up on Monday, seven more releases have come out. As before, this is the short version: release by release, only what you will notice in daily use. The biggest change is in 1.18.0: the sandbox no longer just says no, it asks you.

1.15.1 and 1.15.2: polish in the preview

• Syntax highlighting. Code and configuration files open in colour in the preview: JavaScript and TypeScript, JSON, YAML, Python, shell scripts, HTML, CSS and many more, following the light and dark theme. The Source mode of Markdown, HTML and SVG files is coloured too. Copying still gives you the file exactly as it is.
• Release notes, formatted. The update dialog shows the notes of a new version with headings, lists and bold text instead of raw Markdown.
• The Markdown preview uses the width of the pane instead of stopping at a fixed line length.
• Choosing a reasoning level closes the model menu, just as choosing a model does.
• No more lost questions. A question sent right after the start, or right after opening another folder, used to vanish into a chat that was replaced a moment later. It now waits until the right chat is on screen.

1.16.0: images, and skills Snotra writes itself

• Image generation. Ask for a logo, an illustration or an icon, and Snotra draws it with OpenAI’s image model and saves it as a PNG, JPEG or WebP file in your folder. The image appears in the chat, and a click opens it in the preview. It uses the OpenAI key you have already set up; the image model is chosen under Settings › Tool setup. OpenAI bills every image, so each one asks for your approval first.
• Skills written by Snotra. Snotra can write a skill into the open folder, under .agents/skills/, checks it right after writing it, and it shows up in Settings › Skills straight away. A new built-in skill explains what a good skill looks like.
• Settings regrouped. “Security” is now Tools & security — what Snotra can and may do. The former Tools and MCP pages are one page, Tool setup: interpreters, keys and MCP servers.
• Steadier with OpenAI models. Reading a range of lines from a file failed every time with OpenAI models, and patches in a slightly unexpected format were refused. Both now work on the first try.
• Long tool outputs no longer overflow the context. The budget now counts the system prompt and the tool definitions too, and an output that is too long is shortened to what still fits — the tool line in the chat says so.

1.16.1 and 1.17.0: Snotra AI is now Snotra Agent

The desktop app has a new name: Snotra Agent. Snotra AI stays as the name of the platform it is part of — and of this site. The app bundle, the program file, the menu bar, the window title and the downloads carry the new name. Your settings, chats and stored API keys stay where they are and stay readable.

1.16.1 changed nothing you would see; it taught the self-update to install the renamed package. What that means when you update:

• macOS: the update replaces Snotra AI.app with Snotra Agent.app. Add it to the Dock again. If you install from the DMG by hand instead, delete the old Snotra AI.app afterwards.
• Windows and the Linux tar.gz, from 1.16.1: the update installs the renamed package. Shortcuts to Snotra AI.exe have to be created again.
• Windows and the Linux tar.gz, 1.16.0 or older: install the new version by hand once, unzipped in place of the old folder. Settings and chats are kept.
• AppImage and .deb: nothing changes.

1.17.1: zoom into images

The image preview zooms like the PDF preview: −, percentage, + and Fit in the header, from 25 to 400 percent, or Cmd/Ctrl with +, − and 0. An image larger than the column can be dragged around with the mouse. Handy for the images from 1.16.0.

1.18.0: the sandbox asks

Commands and Python scripts run in a sandbox: they may write in the open folder and nowhere else, and reach only the hosts the call names. Until now, when the sandbox refused something, the command simply failed, and neither you nor the model saw why. 1.18.0 changes that in five steps.

• You see what was blocked. A box under the tool log lists each refused write, read and connection: what, where, and why — the raw sandbox lines folded away underneath.
• A refused write or read: a card after the run. It offers exactly the blocked path or the folder one level up, for this run only or for this session, and then Allow and run again. The sandbox stays on for everything else.
• A refused connection: a card while the command waits. When a command reaches for a host it was not given, say pip fetching from download.pytorch.org, the connection is held and a card asks right away: this host, or every host of its domain. Allow it, and the command simply carries on — nothing runs twice.
• Files outside the open folder: a card before the call. When the agent wants to read or write a file outside the folder, it no longer hits a wall: a card offers that file or the folder around it. Some places are never offered — your home folder as a whole, the root of a disk, Snotra’s own storage. A place with credentials, such as ~/.ssh, is offered for reading only exactly as named and never for writing.
• A notification when a card waits. A run stands still until you answer. If Snotra is in the background, or the card waits in another chat, the system now says so, and a click brings you to it. The switch is under Settings › General.

These cards appear in every mode, Auto included: they widen the sandbox, and that is your call. Saying no ends it: the command is not run again, and the model is told not to work around it but to tell you. What you allow for the session is listed under Settings › Tools & security › Session allowances, where you can revoke it.

One smaller thing in the same release: the preview goes back and forward. ‹ and › in front of the file name step through the files the preview showed, with the scroll position kept. Cmd+[ and Cmd+] on macOS, Alt+← and Alt+→ on Windows and Linux, or the mouse’s side buttons; a right-click on a button lists the whole way.

A user manual

Snotra Agent now has a manual, at docs.snotra-ai.dev, in English and German. Five chapters are written: getting started, chatting, the workspace, modes, permissions and safety, and customising. Its screenshots are generated by script, so they keep up with the app, and every change to the app now comes with the matching change to the manual.

Getting it

Snotra checks for a newer version at startup and walks you through the update, one confirmed step at a time — mind the note on the rename above if you are on Windows or the Linux tar.gz. Otherwise, download it for macOS, Windows or Linux. Every release has its notes on GitHub, and the Discussions are open for anything that doesn't work the way this article says.