DeepSeek Harness: an open-source, plugin-first agent harness

DeepSeek has released DeepSeek Harness, an agent harness that the company says is now in public preview worldwide and open source. It ships as a desktop app for macOS on Apple silicon and 64-bit Windows, and it can also be launched as a web UI from the command line. For anyone who has been running coding agents through a terminal CLI or a vendor's closed app, this is a third option: a harness whose interface, tools and agent loop are themselves swappable parts.

The pitch is about architecture, not the model. DeepSeek describes Harness as built on Cordis's "everything is a plugin" architecture, with composable plugins that extend what agents can do, including the interface itself.

What shipped

The product page lists two installation paths. One is a single command, npx @deepseek-ai/dsh web, run after installing Node.js, which launches the web UI. The other is cloning the full source and following the repository's setup instructions. Desktop downloads are offered alongside links to GitHub, developer docs, a community plugins directory and a Cordis paper.

Out of the box, the harness is pitched at both general work and software engineering. The company's own description of the tool covers everyday work with docs, data and slides, coding tasks such as fixing bugs and running tests, research with cited sources, and background tasks like running scripts and batch-processing files.

The page splits its plugin list into official and installed sections. Official plugins include Terminal, Agent loop and Subagents, plus four marked experimental: Agent teams, Auto approval review, Scheduled tasks and Voice input. The agent loop being a plugin shows how far the composability goes. Most harnesses treat that part as fixed, and here it can be replaced.

Two features stand out for developers. A Creator mode lets users build plugins through chat. The page shows a demo in which a request for a Pomodoro timer plugin leads the agent to load a plugin-development skill, inspect the runtime's providers and slots, write a package manifest and a client file, install the bundle and verify the live plugin state, reported as taking 5m 24s. That is a company demo, not an independent measurement.

The second is Developer tools, which the company says inspects execution traces and detailed runtime information to troubleshoot tool calls and task execution. The screenshots break each turn down into system prompts, user messages, tool calls and results. Each step gets a timing down to the millisecond and its own payload, result and schema views. Debugging is usually the weakest part of agent products, so built-in tracing from DeepSeek itself is a meaningful addition.

An ecosystem that is already large, and already noisy

On GitHub, the dsh-plugin topic lists 17,558 public repositories at the time of writing. JavaScript (10,431) and TypeScript (5,521) dominate the language split. The topic page describes DSH as an agentic, plugin-based software development harness. It says a plugin typically ships a host bundle plus an optional client section and registers into the harness through Cordis. The main repository, deepseek-ai/deepseek-harness, shows 243k stars and a last update of Oct 3, 2026.

Treat the 17,558 figure with care. Several of the top-starred results are general-purpose projects that list dsh-plugin among a long list of tags. Examples include a privacy-focused resume builder and a multi-agent swarm framework whose tags and descriptions also cover MCP servers, agent skills, or Claude Code and Codex integration. GitHub topic tags are self-applied, so the count measures attention, not the number of plugins built for Harness.

What is still unknown

The page does not state a license, so "open source" is the company's own description until someone checks the repository. The page also says nothing about pricing or about usage limits for the preview.

Model support is also unclear. The interface mock-up shows a model selector reading **DeepSeek-V41-Flash** with a "High" setting. The page doesn't say which models are supported, whether third-party or local models can be plugged in, or what the "High" control adjusts. It has no benchmark numbers or evaluation results of any kind.

Platform coverage is narrower than the plugin count suggests. The download links cover macOS Apple silicon and Windows 64-bit only, with no Linux build listed, though the npx and from-source routes may fill that gap. The harness also runs shell commands, writes files and installs plugin bundles, and Auto approval review is still marked experimental. That permission model deserves scrutiny before anyone points it at a real repository. We have not tested it yet.

Sources

• DeepSeek Harness product page
• GitHub: dsh-plugin topic