DeepSeek Harness

DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek that runs as a desktop app or a local web UI. It hands file work, coding, research and recurring jobs to an agent working in a workspace you choose. Its core idea is that nearly every part of the agent, from tools to the interface, is a plugin you can swap or build.

Key Highlights

• Open source under the MIT license, in public preview
• Desktop app for macOS (Apple silicon) and Windows (64-bit), plus a web UI started with npx @deepseek-ai/dsh web
• Agent that reads and edits workspace files, runs commands, delegates work and keeps a plan
• Plugins you install, or build through chat in Creator mode
• Background jobs with progress tracking, plus scheduled tasks (experimental)

What Makes It Different

Most agent apps ship a fixed product. DeepSeek Harness is built on Cordis, an "everything is a plugin" architecture, so tools, skills and even parts of the interface are plugins. That makes it closer to a kit for assembling your own agent, and plugin authors share their work under the dsh-plugin GitHub topic.

It is not limited to coding. The same app organizes files, analyzes spreadsheets, drafts documents and slides, and previews Word, Excel, HTML, PDF and Markdown files in the conversation.

Features & Capabilities

You start a session in a selected workspace, and the agent works there under a permission policy that asks before sensitive operations. For coding, it explores repositories, fixes bugs, builds features and runs tests. For research, it looks up information and cites what it finds. Background tasks run scripts or batch-process files, and recurring work such as a weekly report can be scheduled.

Developer tools let you inspect each turn's tool calls, payloads, results and timing. You add a DeepSeek API key under Settings, and other providers and custom OpenAI-compatible endpoints are also supported. A Python SDK and extra CLI modes are available. Agent teams, auto approval review, scheduled tasks and voice input are marked experimental.

User Ratings and Testimonials

DeepSeek Harness has no established rating on review sites yet. In our hands-on test in November 2026, signing in with Google failed, but setup with an API key worked. The interface felt close to familiar coding agents, work stayed inside the chosen folder, and model costs were low. DeepSeek calls it a preview with evolving core plugins and APIs and warns of compatibility-breaking changes, so expect rough edges and plan for updates.

Pricing & Value

• DeepSeek Harness: free and open source (MIT), for the desktop app, web UI and source code
• Model usage: billed separately by the model provider you connect, such as the DeepSeek API

The harness itself costs nothing, so your spend depends on the model and how much the agent runs. It suits users who want a customizable, self-run agent and accept preview software.