DeepSeek Harness

Harness is the open-source agent runtime DeepSeek previewed in 2026. Formula: Agent = Model + Harness. The model thinks; Harness reads files, runs the terminal, and calls tools.

DeepSeek API docs Official GitHub repo

Developer preview — APIs will change. Commands and source follow the official GitHub repo (deepseek-ai/deepseek-harness). This is not a new model.

What Harness actually is

Think of it as tack for the model: V4 Pro or Flash supplies the pull, Harness decides which files it may touch, which commands it may run, and how to retry. Official slogan: Agent = Model + Harness. Another: everything is a plugin — model adapters, tools, skills, sessions, sandbox, storage, the agent loop, scheduling, and UI can be loaded or swapped without rewriting the runtime.

It is not the web chat box, not Deep Code, and not Claude Code. Those are other shells. Harness is DeepSeek’s own execution layer. Without a model it is empty; without Harness the model only emits text. Default brain is the DeepSeek API; other OpenAI-compatible endpoints also work.

When to use Harness

Use it when the model must work in a real repo, not just talk:

Read a local codebase, edit files, run tests, and keep a replayable session.
Build a coding agent on DeepSeek’s runtime instead of renting another IDE agent.
Swap models behind the same tools: default DeepSeek API, or another OpenAI-compatible endpoint.
Treat everything as a plugin — swap tools, skills, or UI without forking the runtime.

What Harness is not

Do not put it in the same bucket as models or other agent shells:

Not a new DeepSeek model. There is no harness model ID. You still call V4 Pro, V4 Flash, or whatever you configure.
Not the web chat box. Browser chat is for trying copy and reasoning first.
Not Deep Code. That is a free coding-assistant shell (CLI + VS Code) talking to the same API.
Not Claude Code. That keeps Anthropic’s UI and points the brain at DeepSeek.
Not a frozen production API. This is a developer preview — APIs will change.

At a glance

What it is: Open-source agent runtime
Formula: Agent = Model + Harness
Slogan: Everything is a plugin
Quick start: npx @deepseek-ai/dsh web
Source: git clone https://github.com/deepseek-ai/deepseek-harness
Needs: Node.js + DeepSeek API key
Status: Developer preview — APIs change

Quick start

Developer preview — APIs will change. Typical path:

Install Node.js.
Run npx @deepseek-ai/dsh web and open the local Web UI.
Need full source? git clone https://github.com/deepseek-ai/deepseek-harness
Put in a DeepSeek API key. Default models are DeepSeek; other OpenAI-compatible endpoints also work.

npx @deepseek-ai/dsh web

Harness vs Claude Code vs Deep Code vs web vs API

Same DeepSeek API, different shells. Open the one you actually need:

Option	Role	When to pick it
DeepSeek Harness	DeepSeek’s open agent runtime (files, terminal, plugins)	You want DeepSeek’s own runtime and plugin model
Claude Code	Keep Claude Code’s UI; DeepSeek is the brain	You already live in Claude Code
Deep Code	Free OSS coding assistant, CLI + VS Code	You want a V4-native terminal agent, not a $20 seat
Web app	Browser chat	Try copy and reasoning before you wire an agent
DeepSeek API	Hosted endpoint for your product	You are shipping, not just chatting

Boundaries and traps

Preview APIs change. Do not treat npx @deepseek-ai/dsh web or the plugin surface as a frozen contract.
Harness without a model is an empty shell. You still need a DeepSeek API key (or another OpenAI-compatible endpoint).
A GPU is not required for the runtime when inference runs on the cloud API. Local Node.js is.
Do not confuse Harness with Deep Code or Claude Code. Same API, different shells — pick one on purpose.
If the README and this page disagree, the official GitHub repo wins.

Related guides

Claude Code DeepSeek

Write and edit code in the terminal

Deep Code

Free terminal coding assistant built for V4

DeepSeek API docs

OpenAI-compatible API for your apps

DeepSeek V4 Pro

1M context, deep reasoning and vision

What is DeepSeek Harness? Open-Source Agent Runtime and API Setup

Longer npx start and plugin notes.

FAQ

Is Harness a new DeepSeek model?

No. There is no harness model ID. Harness is an agent runtime. The model is still V4 Pro, V4 Flash, or whatever you configure. Official slogan: Agent = Model + Harness — the model thinks; Harness reads files, runs the terminal, and calls tools. Without a model it is empty; without Harness the model only emits text.

How do I start Harness?

Install Node.js, then run npx @deepseek-ai/dsh web and open the local Web UI. Need full source? git clone https://github.com/deepseek-ai/deepseek-harness and follow that repo. Put in a DeepSeek API key. Default models are DeepSeek; other OpenAI-compatible endpoints also work. This is a developer preview — APIs will change.

Do I need a GPU?

No for the cloud API. Inference runs on DeepSeek (or your chosen endpoint). The runtime itself is a Node.js process on your machine. You need Node.js locally and a key; you do not need a local GPU unless you point Harness at a self-hosted model.

Do I need a DeepSeek API key?

Yes, unless you point it at another OpenAI-compatible endpoint you already run. Harness is the execution layer, not a hosted chat. Keys come from DeepSeek’s open platform. The web app can try prompts without a key; Harness cannot do useful work without a model behind it.

Harness or Claude Code?

Want DeepSeek’s own runtime and the “everything is a plugin” model? Harness. Want to keep the Claude Code UI and only swap the brain? Use the Claude Code DeepSeek setup. Same DeepSeek API, different shells. You can run both; pick by workflow, not by camp.

Harness or Deep Code?

Harness is DeepSeek’s open agent runtime — files, terminal, plugins, replayable sessions. Deep Code is a free coding-assistant shell (CLI + VS Code) adapted for V4, billed on API tokens instead of a $20 IDE seat. If you want DeepSeek’s own plugin runtime, open this page. If you want a terminal / VS Code coding assistant, open Deep Code.

How is Harness different from the web app or the API?

The web app is a browser conversation. The DeepSeek API is the hosted endpoint you call from a product. Harness sits on that API (or another OpenAI-compatible endpoint) and gives the model hands: files, terminal, tools. V4 is the brain; the API is the door; Harness is the workshop.

Will the APIs stay stable?

Not yet. This is a developer preview — APIs will change. Treat npx @deepseek-ai/dsh web, plugin names, and session formats as documentation, not as an SLA. Watch the official GitHub repo (deepseek-ai/deepseek-harness) before you pin a production workflow to a preview surface.