DeepSeek Harness Is More Than a Coding Agent: Everything Is a Plugin
DeepSeek's open-source agent harness separates the model from the tools, sessions, sandbox, storage, loops, and UI around it.
DeepSeek has released DeepSeek Harness, an open-source developer-preview project designed to separate an AI agent's model from the software that lets the agent actually operate. The central idea is simple: everything is a plugin. DeepSeek says models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and even the UI can be swapped or recomposed without changing the Harness source itself. DeepSeek's official announcement describes the system as an environment for keeping agents working in real-world settings rather than simply generating model responses.
What DeepSeek Harness actually does
DeepSeek frames an agent as Model + Harness. The model supplies the intelligence, while the harness provides the surrounding system that lets an agent understand its environment, use tools, maintain sessions, and continue working.
That distinction matters because a coding agent is more than an LLM. It needs access to files and a shell, a way to remember what happened during a run, controls for tool execution, a mechanism for coordinating multiple steps, and an interface for the developer. DeepSeek Harness turns those pieces into separately composable capabilities.
The unusual part: everything is a plugin
The architecture is built on Cordis, a plugin framework that manages plugin mounting, unmounting, dependencies, services, and events. DeepSeek says the Cordis kernel itself does not contain the agent's individual capabilities. Those capabilities live in plugins and work together through the runtime.
- Models can be provided through plugins.
- Tools expose capabilities the agent can use.
- Skills add reusable agent behavior.
- Sessions provide persistent run history.
- Sandboxes define the environment in which work can happen.
- Storage, loops, scheduling, and UI are also treated as replaceable capabilities.
The practical result is that developers can select, replace, or extend capabilities through configuration instead of modifying the core Harness implementation. That makes the project less like a single-purpose coding app and more like an open framework for assembling different kinds of agents.
Four runtime modes are built in
DeepSeek currently provides four runtime modes, each aimed at a different way of working with an agent.
| Mode | What it is for |
|---|---|
| Standard | A full coding agent with file editing, shell, file and web search, skills, planning, goals, subagents, and workflows. |
| Code | Exposes tools through a Code Mode SDK so the model can combine multiple tool operations in one TypeScript program. |
| Minimal | A stripped-down coding agent with a persistent shell and file editor, intended for minimal agent benchmarking. |
| Creator | A development environment for inspecting the runtime, testing plugins in memory, and creating custom agent presets. |
The Creator mode is particularly important to the project's larger idea. Rather than only using a predefined agent, developers can inspect the runtime and experiment with different plugin combinations to create their own presets.
Every agent run is designed to be traceable
Another notable part of DeepSeek Harness is its session model. DeepSeek says everything the model sees is recorded in an append-only session log, including system prompts, reasoning, tool calls and results, subagent scheduling, and context injections.
The Trajectory view can then be used to inspect those records. DeepSeek says developers can resume, fork, search, and replay runs from the same event stream.
For agent developers, that creates a useful debugging model: instead of only looking at the final answer, you can inspect the sequence of events that produced it and continue from an earlier point.
DeepSeek is opening the infrastructure, not just another model
This release also fits into DeepSeek's broader developer ecosystem. The company has already published work around its models and APIs, including the DeepSeek V4-1 Flash native-vision release and the DeepSeek V4 Flash API public beta. Developers comparing the different ways to access DeepSeek can also use Saganote's DeepSeek pricing guide.
DeepSeek Harness adds another layer to that ecosystem: instead of focusing only on model capabilities or API access, it provides open-source infrastructure for assembling the agent around those models.
You can run DeepSeek Harness locally
The project is available on GitHub under the MIT license. DeepSeek's quick-start command is:
- Install Node.js.
- Run
npx @deepseek-ai/dsh web. - The local Web UI starts on
http://127.0.0.1:3080by default.
Developers who want the source can instead clone the repository and build it with pnpm. The GitHub project notes that the APIs are still changing and that compatibility-breaking changes are expected during the developer-preview period.
Why the plugin architecture matters
The important change here is architectural. Traditional AI applications often bundle the model, tools, agent loop, storage, interface, and execution environment into one product. DeepSeek Harness treats those pieces as components that can be recombined.
That could make it easier for developers to experiment with different agent configurations without rebuilding the entire application around each change. It also gives the open-source community a defined place to build and share plugins rather than requiring every experiment to become a fork of the whole agent.
DeepSeek is not presenting Harness as a finished consumer product. It is a developer preview aimed at people building or experimenting with agent infrastructure. But the "everything is a plugin" approach makes it notable beyond coding: it is an attempt to make the software surrounding an AI model as modular as the model itself.
Frequently Asked Questions
What is DeepSeek Harness?
What does "everything is a plugin" mean?
Can DeepSeek Harness run locally?
npx @deepseek-ai/dsh web, which launches the local Web UI on 127.0.0.1:3080 by default.Is DeepSeek Harness production-ready?
Where is the DeepSeek Harness source code?
The bottom line
DeepSeek Harness is best understood as an open-source framework for assembling AI agents, not simply another coding assistant. Its defining ideas are modular plugins, traceable sessions, multiple runtime modes, and a Cordis-based architecture that lets developers replace or recombine major parts of the agent.
The project is still early, so its APIs and compatibility can change. For developers interested in how the next generation of AI agents are built, however, DeepSeek is making the infrastructure itself part of the experiment.