You want a coding agent in a repo
You need to read and edit files, run a shell, search the web, and let the model finish multi-step engineering work.
Developer preview · MIT open source · Aug 2026
DeepSeek Harness: keep the model working in the real world
DeepSeek Harness (dsh) is not a new chat model. It is an open-source agent runtime. The model thinks; the harness supplies tools, sessions, sandboxes, and replayable logs. The formula is simple: Agent = Model + Harness.
Updated 2026-08-18 · Compiled from the public developer preview
What is DeepSeek Harness?
Separate the model layer from the harness layer before you decide what to install.
Agent = Model + Harness
The model is the mind: it reasons, splits work, and chooses the next step.
The harness is the hands and the job site: it reads the environment, calls tools, and keeps working in a real repo.
DeepSeek Harness is DeepSeek’s agent runtime infrastructure, released in August 2026 as a developer preview for developers worldwide. The source is MIT-licensed. The npm package is @deepseek-ai/dsh, and the repository is deepseek-ai/deepseek-harness on GitHub.
It does not ship another chat model. It answers a different question: once a model can reason, how can it edit files, run commands, call tools, write logs, and resume after failure inside a real directory? Without a harness you usually stop at the chat box; with a harness the model can keep executing inside controlled boundaries.
The official idea is “everything is a plugin.” The default path still needs a DeepSeek API key. The Web UI can open for free; actually running an agent consumes model tokens. This is the v0.1 preview series, so core plugins and APIs may include breaking changes.
Who is DeepSeek Harness for?
It is not the same kind of entry as this site’s web chat. Match the tool to the job.
-
-
You want plugins or custom presets
You want to swap models, tools, sandboxes, or UI and assemble your own agent without changing the Harness trunk.
-
You only want to chat in a browser
You do not need to install Node.js. Use this site’s DeepSeek web entry and talk to DeepSeek-V4 directly.
Architecture: everything is a plugin
DeepSeek Harness is built on Cordis. The kernel only loads plugins and dependencies; every capability lives in a plugin.
-
01
Cordis kernel
It only handles mounting, unmounting, and dependencies. Agent business logic stays out of the trunk, so you do not fork the whole repo to change a capability.
-
02
Capabilities as plugins
Model adapters, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI can all compose as plugins and work together through services and events.
-
03
Compose in config
Developers can select, replace, or extend any capability in configuration and assemble an agent for the scene—without editing Harness source.
Four run modes
The same plugin kernel loads different defaults, covering daily coding, programmatic tool use, benchmarks, and custom presets.
Standard
A full coding agent.
Supports file editing, shell, file and web search, Skills, planning, goals, subagents, and workflows. Built for everyday repo tasks.
PTC
Combine many tool calls with a program.
Includes Standard capabilities, then presents tools through a Code Mode SDK so the model can orchestrate multi-step work in a TypeScript program.
Minimal
The smallest tool surface, for benchmarks.
Usually keeps only persistent bash and file editing (str_replace_editor), so you can watch the model in a minimized environment.
Creator
Inspect the runtime and author custom presets.
Use it to inspect current plugins, try Cordis plugins in memory, and compose new run modes.
Every run is auditable
Sessions are recorded as an append-only event stream. Resume, fork, search, and replay all share the same log.
-
System prompts, chain-of-thought, tool calls, and results are written into the session log
-
Sub-agent dispatch and every context injection can be inspected by source
-
The Trajectory view helps you audit what the model saw and what it did
-
High-risk actions should use permissions and approval, not a bare repo handed to the model
Get started
Try it quickly or install from source
Node.js must already be installed locally. Version requirements follow the official repository README.
Quick start
After installing Node.js, launch the Web UI with npx. It opens on local port 3080 by default: http://127.0.0.1:3080
$ npx @deepseek-ai/dsh webInstall from source
Get the full project source and follow the repository docs to finish setup (pnpm install, pnpm run build, pnpm dsh web).
$ git clone https://github.com/deepseek-ai/deepseek-harnessConfigure an API key and pick a model
In the UI, configure a DeepSeek API key and choose DeepSeek-V4-Flash or DeepSeek-V4-Pro. Start with Flash.
Set the workspace and permissions
Pick a repo directory and tighten file and command permissions before assigning work. Commands and config may change during the preview; follow the repository docs.
How is web chat different from Harness?
This site’s DeepSeek web chat is for opening a browser and talking. Harness is for continuous execution inside a repo.
| Topic | DeepSeek web chat | DeepSeek Harness |
|---|---|---|
| What you are looking for | A no-install online chat entry | A local / source-level agent runtime |
| Typical users | Study, office, writing, Q&A | Developers building agents and plugins |
| Can it edit repo files? | Conversation first | Read/write files and run commands within policy |
| How to start | Open this site’s entry in a browser | Start the Web UI with npx or clone the source |
DeepSeek Harness FAQ
Open-source license, cost, model choice, and how this differs from web chat.
Is DeepSeek Harness a new official model?
No. It is DeepSeek’s open-source agent runtime (a harness) that connects models, tools, sessions, and UI. Chat and reasoning still come from models such as DeepSeek-V4.
Is DeepSeek Harness free?
The framework source is MIT-licensed and free to obtain. Running an agent usually calls the DeepSeek API and consumes tokens. Follow official pricing and the balance on your key.
Is it the same as this site’s DeepSeek web chat?
No. Web chat is for opening a browser and talking. Harness is for developers running a coding agent in a real environment. Use this site’s entry to chat; use Harness to edit code and run tools.
Do I have to host a large model locally to use Harness?
No. The common path in public docs is the DeepSeek cloud API. You need Node.js, an API key, and a browser. You do not have to bring your own GPU.
Can I use it in production now?
Official docs mark it as a developer preview and warn about compatibility-breaking changes. It is suitable for learning, experiments, and plugin development. Evaluate stability yourself before production, and pin a version you have verified.
What is dsh, and how do I start it?
dsh is the common short name and CLI for DeepSeek Harness. After installing Node.js, run npx @deepseek-ai/dsh web. The Web UI opens at http://127.0.0.1:3080 by default.
Pick the scene first, then the entry
Need a conversation now? Open DeepSeek web chat. Need an agent in your repo? Use official DeepSeek Harness.
This page is compiled by the DeepSeek-V4 site from DeepSeek’s official Harness page and GitHub repository so you can search and compare. It is not an official DeepSeek website. Follow the official docs and repository README for install, configuration, and compatibility.