DeepSeek open-sourced the agent runtime
DeepSeek Harness (CLI name dsh) is not a new model, and it is not chat.deepseek.com. It is the layer around a model: tools, files, sessions, a loop, a UI. Official line: everything is a plugin.
Updated 2026-08-13. Numbers are a same-day snapshot of the official repo. No invented benchmarks.
Three things people mix up
Chat product
chat.deepseek.com is the web chat. Use it for questions and copy.
Model API
deepseek-v4-pro / flash are models on the API. Your backend calls them.
Harness / dsh
The runtime that lets a model read a repo, run commands and swap plugins. This page is about that.
Founders
Only verified people pages, official X accounts, and interviews already on this site. Unverified personal accounts stay off.
Founder
Liang WenfengDeepSeek founder
DeepSeek founder; shocked the world with restraint and open source
Lab official X: @deepseek_ai(Liang Wenfeng's personal account is unverified and not listed under his name)
Get it running
The official shortest path is one command. After boot you still need two things: an API key and a workspace. The composer stays closed until a workspace is selected.
npx @deepseek-ai/dsh web # Web UI → http://127.0.0.1:3080 # Settings → Models → DeepSeek API key # Choose workspace → send: Summarize this repository
Command and default port come from the official README and Web UI guide. Compatibility will change; trust the current repo docs.
A boot is four stacked layers
Do not picture Harness as a locked app. Boot stacks an empty list: base bundle, then web or headless, then your patch. Any layer can be swapped.
- 01
cordis.patch.yml
Your overlay: swap models, add plugins, change policy.
- 02
profile · web / headless
Which stack boots: a UI, or a one-shot runner with no server.
- 03
bundle · dsh-web-app
The browser app. Headless profiles do not stack this layer.
- 04
bundle · dsh-base
First layer of every profile: model adapters, tools, persistence, sandbox, credentials.
Tutorials
All tutorialsBeginner · 2026-08-13
DeepSeek Harness Quickstart: Boot the Web UI, Then Close One Repo Loop
Do not start by writing a plugin. Follow the official README: boot the dsh Web UI, add an API key, pick a workspace, then ask the agent to summarize the repo.
Intermediate · 2026-08-13
DeepSeek Harness Architecture: Why ‘Everything Is a Plugin’ Is Not a Slogan
Use the official architecture.md to explain profiles, bundles, the session log and turn/step. No package table dump — only where new behavior should attach.
Intermediate · 2026-08-13
Your First DeepSeek Harness Plugin: Mount It Before You Productize
Follow the official first-plugin tutorial through apply(ctx) and a Cordis overlay. No full step dump — only boundaries and what to do next.
Beside Claude Code and Cursor
Orientation only — not a ranked review. Harness license and architecture come from the official README and architecture.md; Claude Code and Cursor shapes come from their product pages. We have not run a shared eval, so we do not claim which is faster, more accurate or cheaper.
| Product | License | Shape | Replaceability | Maturity (official) |
|---|---|---|---|---|
| DeepSeek Harness | MIT | Open agent runtime + Web UI / headless | Model, tools, session, loop and UI are plugins | Developer preview; breaking changes expected |
| Claude Code | Proprietary product | Terminal coding agent | Skills / hooks / MCP; core loop is closed | Generally available to developers |
| Cursor | Proprietary product | AI-native IDE | Rules, Skills, extensions; the editor is closed | Generally available to developers |
Directory
Official
- Official site
Product page for DeepSeek Harness (Chinese at /harness/zh/).
- GitHub repository
Official source, architecture docs, Web UI and plugin tutorials. MIT licensed.
Docs
- Official README
One-command run, source checkout, community links and license.
- Architecture docs
Cordis, profiles/bundles, session log, and the turn/step lifecycle.
- Web UI guide
Configure a model, pick a workspace, run the first task.
- Model providers
DeepSeek keys, catalog providers (Anthropic / OpenAI) and custom OpenAI-compatible endpoints.
- First plugin
Write an apply(ctx) plugin and mount it into the Web UI with a Cordis overlay.
Community
- GitHub Discussions
Official feedback and bug reports.
- Discord community
English community listed in the official README.
- dsh-plugin topic
Tag plugin repos with this topic so they can be discovered.
Related
- Cordis framework
The plugin runtime under Harness: services, typed events, reversible effects.
- DeepSeek API docs
V4 Pro / Flash, billing, and Codex / Claude Code integration notes.
- DeepSeek Chat
The official chat product — not the same thing as Harness.
Latest news
2026-08-13 · GitHub · deepseek-ai/deepseek-harness · Official
DeepSeek open-sources DeepSeek Harness (dsh): everything is a pluginOfficial repo deepseek-ai/deepseek-harness is public: MIT license, developer preview, one-command `npx @deepseek-ai/dsh web`. Built on Cordis so the model adapter, tools, sessions and agent loop are all replaceable plugins.
2026-08-03 · South China Morning Post · Press
SCMP: DeepSeek invites open-source developers to test its upcoming harnessThe report said DeepSeek was inviting testers for a harness layer that turns LLMs into executable agents, in the same narrative as V4's agentic work.
LaoJin's pick
FAQ
- What is DeepSeek Harness?
- DeepSeek Harness (CLI name dsh) is DeepSeek AI's open-source agent runtime: tools, filesystem, sessions, a loop and a UI around a model. It is not a new model, and it is not the chat.deepseek.com product. The official design is “everything is a plugin,” powered by Cordis.
- How do I install and start DeepSeek Harness?
- Install Node.js, then run `npx @deepseek-ai/dsh web`. The Web UI defaults to http://127.0.0.1:3080. From source: clone, `pnpm install && pnpm run build && pnpm dsh web`. After boot, add a DeepSeek API key under Settings → Models and choose a workspace before sending a task.
- Does DeepSeek Harness only work with DeepSeek models?
- No. The official provider guide covers a DeepSeek key, catalog providers such as Anthropic and OpenAI, and custom OpenAI-compatible endpoints. Keys live in `$DSH_HOME/.credentials.yaml`; model changes apply on the next request without a restart.
- How does DeepSeek Harness differ from Claude Code and Cursor?
- Claude Code and Cursor are mature commercial coding-agent products. Harness is an MIT-licensed runtime still in developer preview: you can replace the model adapter, tools, sandbox and UI. Think of it as a take-apart agent chassis, not a finished IDE. The official README warns that compatibility-breaking changes will happen.
- What are dsh, a profile, and a bundle?
- `dsh` is the launcher. A profile is a named plugin composition (shipped templates: web and headless). A bundle is a patchable plugin layer: `dsh-base` owns models/tools/sandbox, `dsh-web-app` adds the browser UI, `dsh-headless` is a one-shot runner with no server. Your own overrides live in `cordis.patch.yml`.
- Is it production-ready?
- The official README is explicit: it is a developer preview, iterating rapidly, and there will be compatibility-breaking changes. Use it to learn the architecture, write plugins and experiment. Pinning core business flows to the current API is risky. The license is MIT.
How this hub stays current
Append news to the head of HARNESS_NEWS in src/content/deepseek-harness.ts. Links, FAQ and the comparison table live in the same file. Tutorials, directory cards and picks stay in their append-only files and hang off this page by slug. Unverified numbers stay out.
Playbook: docs/deepseek-harness/PLAYBOOK.md