DeepSeek released their agent framework dsh

Whenever DeepSeek releases something open source, it immediately grabs attention. This time, the engineers rolled out DeepSeek Harness (abbreviated as dsh) — an open platform for running, debugging, and chaining AI agents.

If you've grown tired of heavyweight frameworks with complex abstractions and tangled call graphs, DeepSeek's approach might feel like a breath of fresh air.

What is dsh and why do you need it

In engineering, "harness" typically refers to a test bench or rigging that holds components together and lets them operate in a controlled environment. That's exactly what dsh solves.

Most modern agent-building tools try to do everything at once: manage memory, invoke tools, orchestrate conversations, construct vector databases. The result is that projects quickly become monolithic. DeepSeek chose a radically different path. Their core principle: "everything is a plugin."

The entire system is built around the Cordis engine. It's a component-based microkernel framework for TypeScript designed for spatiotemporal module composition. The idea is that the system kernel stays minimal while any capabilities — from network connectors to code tools — are added as independent plugins with their own lifecycle.

How the project is structured internally

At its core lies a services and contexts architecture. Plugins in dsh can:

• Register their APIs and share them with other modules
• Dynamically intercept and modify events
• Manage dependencies without tight coupling to specific implementations
• Load and unload on the fly without restarting the host

This provides flexibility during prototyping. If you need to replace how message history is stored or swap out a language model provider, you simply change the plugin in the configuration.

Quick start

You can try the tool in action in just a couple of minutes. You'll need a recent Node.js:

The command will spin up a web interface on http://127.0.0.1:3080 and immediately open it in your default browser. If you're working on a remote server via SSH, you don't need to open the browser. Just add the flag:

If you want to dig into the code or write your own plugin, the project builds easily from source using pnpm:

After building, the pnpm dsh web command will use the compiled artifacts without rebuilding.

Current state and caveats

It's worth considering the project's status. DeepSeek honestly warns: the repository is in developer preview. Code changes rapidly, backwards compatibility breaks regularly, and some documentation sections are still being written.

There are no ready-made enterprise solutions out of the box yet. If you need a stable library for production tomorrow, it's better to wait. But if you're exploring autonomous agent architecture or looking for a modular foundation for your own experiments, studying how dsh works is definitely worth your time.

For publishing your own extensions, the developers suggest using the dsh-plugin topic on GitHub so the community can find and test new modules.

Who will find this project useful right now

The project will be most appreciated by TypeScript developers who feel constrained by standard Python libraries. The microkernel-based architecture is well-suited for creating custom work environments, local assistants with custom tools, and testbenches for agent scenarios.

The source code is open under the MIT license, so you can fork, study, and adapt the code for your own needs without restrictions.