DeepSeek Harness v0.2: Desktop App, Plugins and Automation Guide

DeepSeek Harness v0.2 turns DeepSeek’s open-source agent harness into a more practical everyday workspace, adding a desktop-oriented experience, stronger plugin management, file review tools and automation workflows while keeping the project’s “everything is a plugin” architecture.

The release is still explicitly a developer preview. DeepSeek warns that compatibility-breaking changes are expected, so v0.2 is better viewed as a fast-moving platform for experimentation than a finished enterprise desktop product.

DeepSeek Harness v0.2 at a glance

The official source of truth is DeepSeek’s deepseek-harness GitHub repository, which describes the project as an open-source agent harness built on an everything-is-a-plugin architecture.

What is DeepSeek Harness?

An agent harness is the runtime around a model that lets it do work: read files, call tools, run commands, maintain session state and apply an agent loop. DeepSeek Harness, also called dsh, makes those layers composable as plugins.

That is different from a closed assistant app where the model, tools and loop are fixed. In Harness, even core pieces can be replaced or extended, which is why the project appeals to developers building custom agents or local workflows.

What changed in v0.2?

Launch coverage around the v0.2 preview highlights four changes that make the project easier to use outside a terminal-only coding workflow:

• Desktop experience: packaged builds and a more app-like workflow for macOS and Windows users.
• Plugin management: a dedicated interface for installing, enabling, disabling and configuring plugins.
• File and diff review: a sidebar for inspecting generated documents and code changes.
• Automation tasks: recurring prompt/workflow execution with history and editable schedules.

The official repository also includes development paths for Web and Desktop and continues to evolve rapidly.

Quickest way to run DeepSeek Harness

If you already have Node.js, the shortest official route remains:

npx @deepseek-ai/dsh web

By default, that starts the Web UI locally at:

http://127.0.0.1:3080

The command normally opens the default browser automatically. For remote or server use, the CLI supports a no-open mode and the documentation covers host/forwarding options.

How to run from source

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Running from source is the right option if you plan to modify plugins, inspect the architecture or contribute to the project. For normal testing, the npm command is simpler.

Why “everything is a plugin” matters

DeepSeek’s architecture treats major parts of the agent runtime as composable plugins rather than hard-coded components. That can include model adapters, tool registries and agent-loop behavior.

In practice, this means a developer can change more than the model. You can swap providers, add tools, alter how the agent plans and expose specialized workflows without forking an entire monolithic application.

Plugin manager in v0.2

One of the most useful desktop-era improvements is a clearer plugin-management surface. Rather than thinking of extensions as configuration files hidden in a repository, v0.2 pushes plugin discovery and control into the normal workspace.

For teams experimenting with agents, this makes it easier to answer basic operational questions:

• Which plugins are active?
• What capabilities did this session load?
• Can a tool be disabled without rebuilding the app?
• Which provider or compatibility layer is handling a request?

Creator mode

Launch coverage describes a Creator mode that lets a user describe a plugin in chat and have the agent build and install it. That is a powerful feature, but it also increases the importance of reviewing generated code before trusting it with sensitive files or credentials.

An agent that can extend its own toolset can become much more useful — and much more capable of making unintended changes.

Automation Task plugin

The Automation Task feature is aimed at recurring work. Instead of manually sending the same prompt every morning or after every file update, you can define a task, choose a frequency and inspect past runs.

Potential uses include:

• periodic repository checks;
• daily document summaries;
• recurring data cleanup;
• scheduled report generation;
• routine file transformations.

Because v0.2 is a developer preview, important production features such as reliability guarantees, failure recovery and permission isolation should be tested rather than assumed.

Document and file workflows

Harness is no longer only interesting as a coding shell. The v0.2 workspace is designed to handle files such as documents, spreadsheets and PDFs, with generated outputs available for review.

The diff/file sidebar is particularly useful because agentic work needs visibility. A good workflow is not merely “the agent changed a file”; it is “the agent changed a file and the user can quickly inspect what changed before accepting or distributing it.”

Does DeepSeek Harness only work with DeepSeek models?

No. The project’s provider architecture is designed to support third-party providers and custom OpenAI-compatible endpoints. That flexibility is one of the strongest reasons developers may choose an open harness rather than a model-specific closed client.

If you are comparing open model ecosystems, AVARIXO also covers llama.cpp v0.6.0 and Mistral Large 4.

Session modes

Current v0.2 coverage describes several session modes aimed at different jobs:

• Standard: the general-purpose default experience.
• Creator: focused on building and installing plugins.
• PTC: a programmatic tool-calling mode.
• Minimal: a leaner loop used for evaluation and benchmark-style runs.

Exact behavior can change between preview builds, so check the versioned documentation before writing automation that depends on one mode’s internals.

Desktop platform support

v0.2 launch reporting describes desktop builds for macOS on Apple silicon and Windows x64. Linux users can still run the Web UI or CLI path, and community desktop wrappers exist, but third-party wrappers should not be confused with DeepSeek’s official project.

That distinction is important because GitHub already contains multiple independently maintained “DeepSeek Harness Desktop” repositories. Use the official DeepSeek repository and project site when you need authoritative downloads or security guidance.

Safety warning: developer preview means developer preview

DeepSeek’s repository uses unusually direct language: compatibility-breaking changes will happen. That should shape how you deploy it.

• Do not assume plugin APIs are stable.
• Review tools that can execute commands or modify files.
• Keep API keys and sensitive credentials out of prompts and untrusted plugins.
• Use a test workspace before pointing an agent at valuable repositories or documents.
• Pin versions if a workflow depends on a specific behavior.

DeepSeek Harness vs. a normal chat app

A chat app is optimized for conversation. Harness is optimized for an agent that can act inside a workspace. That means the main value is not simply “chat with DeepSeek in a window.” It is the tool/plugin runtime around the model.

If all you need is question-and-answer chat, Harness may be unnecessary. If you need file operations, custom providers, plugins, automation and inspectable agent sessions, the architecture becomes much more relevant.

FAQ

Is DeepSeek Harness open source?

Yes. The official repository is published under the MIT license.

Is v0.2 stable?

No. DeepSeek labels the project a developer preview and warns that compatibility-breaking changes are expected.

How do I start the Web UI?

With Node.js installed, run npx @deepseek-ai/dsh web. The local UI normally opens at 127.0.0.1:3080.

Can it use non-DeepSeek models?

Yes. The provider architecture supports third-party and OpenAI-compatible endpoints.

What is new in the v0.2 desktop workflow?

The release emphasizes desktop packaging, plugin management, file/diff review and recurring automation tasks.

Bottom line

DeepSeek Harness v0.2 is compelling because it moves an open agent runtime closer to an everyday workspace without giving up its plugin-native architecture. The desktop and automation additions make it easier to test, while the MIT license and provider flexibility make it unusually hackable. The tradeoff is maturity: it remains a developer preview, so expect breaking changes and treat permissions, plugins and generated code with the same caution you would apply to any tool-capable local agent.