DeepSeek open-sources its plug-and-play agent harness

DeepSeek just dropped DeepSeek Harness (dsh), an open-source framework designed to orchestrate and run its agentic models. Think of it as DeepSeek's official equivalent to what Claude Code is for Anthropic or Codex is for OpenAI.

Previously, running DeepSeek agents meant relying on third-party frameworks like Hermes or OpenClaw. Now, the model maker has shipped its own dedicated harness.

Everything is a plugin

The core idea behind dsh is an absolute "everything-is-a-plugin" architecture. Powered by the Cordis framework, practically every component—models, external tools, skills, session history, sandboxes, storage layers, event loops, schedulers, and even the user interface—is an independent plugin. You can swap them in and out through configuration without touching a line of the core codebase.

Among its four operation modes, the standout is Minimal mode. It strips away everything except a basic bash shell and a text editor, making it an ideal bare-metal playground for running clean agent benchmarks.

Every run is completely reproducible. Prompts, reasoning traces, tool executions, sub-agent dispatches, and context updates are all recorded to an append-only session log. That single event stream powers resuming, branching, searching, and replaying past runs.

Under the hood, Cordis enforces spatiotemporal composability. In time, uninstalling a plugin cleanly unwinds its side effects because every context transformation carries a trackable inverse. In space, dependencies are explicitly declared, and components get reactive notifications whenever context shifts.

How to spin it up

Getting started requires Node.js. You can launch the local web interface directly via NPX using `npx @deepseek-ai/dsh web`. By default, it spins up on port 3080 and opens in your browser, with flags available to suppress browser launching over SSH.

Developers can also clone the repository directly from GitHub, install dependencies with pnpm, build the project, and run the developer environment locally.

Keep in mind that DeepSeek explicitly labels this release as a Developer Preview. Development is moving fast, and breaking changes are definitely on the horizon.

Why it matters

If you are building private coding assistants or custom enterprise agents, dsh is well worth an early look. Its plug-and-play minimalism means you can inject your own tools and sandboxes without wrestling with bloated framework assumptions.

Frameworks built by the model maker usually offer the best synergy with that company's models. While dsh is far too raw for critical production workflows right now, it offers a modular foundation that could make running local DeepSeek agents much more flexible.