DeepSeek Harness Everything is a Plugin · Agent Runtime

DeepSeek Harness (command name dsh) is the Agent runtime (agent harness) open-sourced by DeepSeek AI. It lets the model understand its environment, call tools, and keep working in real scenarios. The architecture slogan is “everything is a plugin”; the kernel is Cordis.

The DeepSeek web version and DeepSeek-V4 solve conversation, reasoning, and long context; Harness solves tools, sandboxes, session logs, and a replaceable runtime. Recommended path: spell out the task on the web version first, then run a tool-using Agent locally with dsh as needed.

Officially it is a developer preview, with an explicit warning that breaking changes will land. It is suited to plugin experiments, evaluations, and learning the architecture—not to pinning a critical production system to the current version.

After installing Node.js, run npx @deepseek-ai/dsh web; it opens http://127.0.0.1:3080 by default. You can also git clone the official repo and install and build with pnpm following the README. To validate in conversation first, use Try Now on this site to enter the DeepSeek web version.

Cordis is the plugin kernel used by DeepSeek Harness. It handles plugin load, unload, and dependencies. Concrete Agent capabilities such as models, tools, and UI do not live in the kernel—they live in plugins, collaborating through services and events.

The official Chinese page calls it PTC Mode; the English page calls it Code mode. Both mean composing multi-turn tool calls with a snippet of model-generated code (via the Code Mode SDK), while keeping all Standard Mode capabilities.

Minimal Mode keeps only persistent bash and str_replace_editor, for model benchmarks in a minimized environment. In public evaluation notes, some Code Agent scores for DeepSeek-V4-Flash are tied to the Minimal Mode framework.

Add the GitHub topic “dsh-plugin” to the plugin repository so it can be discovered. Feedback and bugs should go through official GitHub Discussions. Core plugins and the base API will keep iterating.

Yes. Prompts the model sees, chain of thought, tool results, and sub-agent scheduling are all written to an append-only log. Resume, fork, search, and replay share the same event stream, and you can inspect by source in the Trajectory view.

No. Official guidance stresses that developers can select, replace, or extend any capability at the configuration layer—including model adapters and UI—without modifying DeepSeek Harness source.

The official repository is open-sourced under MIT; third-party dependencies are listed in THIRD_PARTY_NOTICES in the repo. Read the license and preview disclaimer yourself before commercial use.

Open the DeepSeek web version on this site, chat with DeepSeek-V4 directly, and confirm prompts and task breakdown. When you need local tools and a sandbox, install Node.js and run dsh. The two paths complement each other; they do not replace each other.