DeepSeek Harness install guide vs Claude Code

DeepSeek Harness is an open-source agent built on plugins

DeepSeek AI released DeepSeek Harness, an open-source agent harness, as a developer preview on August 13, 2026. A harness in this sense is the software that runs an agent around a model, not the model itself. DeepSeek AI designed it so you never patch a fixed core and instead extend it by mounting another plugin alongside the rest.

DeepSeek Harness runs the everything-is-a-plugin design on Cordis

One rule drives the design: everything is a plugin. DeepSeek's announcement says models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and the UI are all plugins, running on a meta-framework called Cordis. The official documentation backs this up: there is no privileged core to patch, and every part, from the model adapter and tool registry to the session log and the agent loop itself, can be replaced through configuration.

The MIT license keeps DeepSeek Harness open source and reusable

The DeepSeek Harness code is open source under the MIT license. MIT is one of the most permissive licenses, so you can reuse the code in commercial products. DeepSeek's announcement pitches the release as an open invitation to other developers building agent harnesses.

DeepSeek Harness is a coding agent that competes with Claude Code

DeepSeek Harness is a coding agent in its own right, in the same category as Claude Code and Codex, not a plugin that runs inside them. Its official repository also ships packages that call those two tools from inside DeepSeek Harness, so it both competes with them and builds on them.

DeepSeek Harness's GitHub repository ships subagents for Claude Code

A dedicated package in the repository runs delegated tasks as fresh, independent Claude Code sessions. It returns only the final answer or a safe failure diagnostic, keeps reasoning, tool traffic and workspace diffs separate from the parent session, and runs unattended without interactive permission prompts using Claude's native authentication. A matching package for Codex starts a fresh Codex process for each delegation, runs a single turn in an ephemeral thread, and also handles approval requests automatically while leaving native Codex configuration and authentication in charge.

How to install and run DeepSeek Harness

There are two ways to launch it, and the npm route is quicker if you just want to try it out. The screenshot shows both routes.

Start DeepSeek Harness with a single npx command

With Node.js installed, one line starts the Web UI. It runs at http://127.0.0.1:3080 by default.

npx @deepseek-ai/dsh web

It needs Node.js ^22.19.0 || >=24.0.0. It will not start on older versions, so check with node -v first.

Run DeepSeek Harness from GitHub source with pnpm

To work on the internals, get the repository from GitHub. It uses pnpm: clone, install, build, then start.

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

What DeepSeek Harness documentation says about profiles and bundles

In practice, two units make DeepSeek Harness swappable: profiles and bundles. A profile is a named set of bundles, and a bundle is the package that carries configuration and code.

A DeepSeek Harness profile lists the bundles it stacks at boot

At boot, the harness assembles a layered tree of plugins. A profile is a named composition kept in the Harness home that lists the bundles it stacks, and five templates come with it: web, headless, sdk, sdk-minimal and acp. This command shows what is actually running in your environment.

dsh --profile web --dump-config

DeepSeek Harness bundles stay patchable by the layers above them

A bundle packages config rows with the code they mount, and anything it adds can still be patched by the layers above it. That keeps the "everything is a plugin" idea intact at every level, from a single tool to a whole profile.

The cost of easy swapping is that behavior can shift a lot with configuration. While the preview still ships changes that break compatibility, try DeepSeek Harness in an isolated environment instead of wiring it into everyday production tools. For the models rather than the harness, see the DeepSeek model articles.