DeepSeek Harness: A Free Desktop AI Agent for Code and Files

While some agents live only in the terminal, on September 29, 2026 DeepSeek released a desktop app for its open source agent Harness, version 0.2. The difference from purely console tools is that this is an actual app with a window, not a command in a terminal. Inside it's the same principle: the model decides what to do, and the harness gives it hands for files, code, and commands.

Straight away, honestly, about the "free" part: the app itself and the Harness code are open and free, under the MIT license. But for the agent to actually do tasks, you need access to a DeepSeek model, through an API key paid by usage or through the quota of a personal account. The shell is free, not the model's tokens inside it.

What it is

Harness is an application layer around the DeepSeek model. The model doesn't touch the file system directly. The harness does that: it reads files, edits code, runs commands in the project, and checks the result. Out of the box it aims at three types of tasks: organizing files, analyzing data, and writing code.

The core is open, MIT license, and the code is on GitHub.

Installation

On Mac and Windows it installs with an ordinary installer from the official DeepSeek Harness page. You don't need to install Node.js separately, since the desktop app builds its own runtime inside itself.

There is no desktop version for Linux. Instead there is a web interface, started with the command npx @deepseek-ai/dsh web, and it needs Node 22.19 or newer (or Node 24+). The --no-open flag stops the browser from opening automatically if you only need the backend.

Access modes and what's inside

The app has three levels of file access: read only (read-only), write inside the chosen working folder (workspace-write), and full access to the whole computer (full access). The mode you choose determines what the agent can do without extra confirmation.

Among the features:

• - HTML reports that build themselves, with no external libraries, with charts and tables inside;
• - plugins, built-in and your own, which run with the permissions you gave them;
• - Creator Mode, where a new plugin is built straight from a task description in the chat;
• - a scheduler that can run a task once or on a schedule.

How billing works

There are two separate paths: an ordinary DeepSeek API key paid by usage, or the quota from a personal DeepSeek account. These are different payment mechanisms, and before running anything resource-heavy, check explicitly which one is active in the app right now, so you don't rack up an unexpected bill on the wrong billing path.

Honest downsides

The interface of the desktop app is aimed at developers and may change between releases without warning. At the time of checking, the version was 0.2.0 RC2, which means it is still a release candidate, not a final stable build.

Plugins execute arbitrary code with the permissions they are given, so checking the source of a plugin before installing it is not a formality but a real security measure.

Plugin state is not kept after the app restarts. And scheduled tasks require the app to stay running: closing the window may leave a background process active, but a full quit stops the schedule too.

For my own content factory I'm keeping Claude Code as the main tool for now precisely because of these downsides. I don't want an RC status and interface changes between releases on a project where a chain of many skills is already built. I'll take Harness for a trial on secondary tasks and only then decide whether to move anything important there.

Who it suits

People who have already tried agent tools in the terminal and want the same thing with an app window and visual reports, without having to keep a terminal open all the time. It's too early for those waiting for a final stable version with no interface changes between updates: the RC status means details will keep changing.

Bottom line

DeepSeek Harness v0.2 is a free open source agent with a desktop app for Mac and Windows and a web version for Linux, three file access modes, and plugins and a task scheduler inside. Before trusting it with a real project, check explicitly which billing method is active, and remember that the interface is still changing from release to release.

I write about tools for vibe coding and automation in the channel «Растём с ИИ» ("Growing with AI"). A related topic: how to check code written by AI if you're not a developer.

If you need to go through your own system of agents and tools in person, you can start with a free consultation.