Introduction

Deskfish, created by Iman Reihanian in 2026, is a VS Code extension that gives an AI agent a computer of its own: a small, sandboxed Linux desktop with a browser and a terminal, shown live in an editor tab. You type a task in the chat sidebar. The agent looks at the screen, moves the mouse, types, and gets on with it, exactly the way a person at a keyboard would. When it hits something only a human can do, a login, a two-factor code, a CAPTCHA, it stops and asks you.

Think of it as a junior coworker with a desk in a glass box: you can see everything it does, you can lean in and take the mouse whenever you want, and host folders are not mounted into its desktop. Files move with Attach and Save; the Desktop tab also shares your clipboard when you focus or paste into it.

What it is good at#

Deskfish is built for real-world chores, not for writing code. It shines at anything that a patient person could do in a browser:

• Research that takes many tabs: comparing hotels across a region, collecting prices, reading reviews, summarizing what it found.
• Filling in forms and working through multi-page flows on websites.
• Downloading files, reports and receipts, and handing them to you.
• Running commands in a terminal and bringing back the results.
• Anything else you would explain to a new colleague by saying "open the browser and…".

It is deliberately not a coding assistant. You already have those in VS Code. Deskfish is the one that has hands.

The three things you will see#

• The chat sidebar. Where you give tasks, read what the agent says, and get handed files. It lives in the activity bar under the fish icon.
• The Desktop tab. A live view of the agent's screen. You can watch, and you can take over at any moment. It opens by itself when you send a task; the Open button at the top of the sidebar brings it back any time.
• The tank. The agent's computer itself: a Linux container running on your machine. You turn it on and off with the power button in the sidebar, and Deskfish manages everything else about it. See The tank.

A few words you will meet#

Deskfish has a small vocabulary, and the interface uses it consistently.

It remembers what matters#

Between chats the agent keeps three things, in plain files on your computer. Facts about you and your tasks: preferences, which accounts it uses, quirks of the sites you send it to; yours to read, edit and empty. A journal: one dated line for every task it finished, and the how-to notes it writes for itself. And a page about who it is, written in its own words, which only it can change, when it reflects after a few tasks. Above that page sits a short charter from the person who made it, ten commitments with the reason for each, which you can rewrite. Every chat is saved as well, and the agent can search its own past. None of it ever contains a password. See Memory.

Any model#

Deskfish is not tied to one AI provider. It works with Anthropic's Claude models through their native computer-use tool, with any OpenAI-compatible endpoint (xAI, OpenRouter, LiteLLM, Ollama, vLLM and others), and with a scripted demo model that needs no key at all. See Models and providers.

Safety in one paragraph#

The tank is the boundary. Inside it the agent is free: the browser, the accounts and the files in there are its own. Outside it, nothing: no folder on your machine is shared with it, files move only when you attach one or click Save, its control ports listen on your machine only, and it runs as an ordinary user with no administrator rights inside a container that is rebuilt each time you turn it on. The full picture is in Security and privacy.

Ask Deskfish about itself#

The agent has read access to this documentation. If you ask it something about Deskfish, for instance "how do I get a file out of your desktop?" or "what happens to your cookies when I turn you off?", it reads the relevant page and answers from it, instead of guessing. It only reads a page when you ask such a question, so it costs nothing during ordinary tasks.

Where to go next#

• New here? Start with Getting started.
• Curious how it actually works? Read How the bot sees and acts.
• Something not working? Try Troubleshooting.