Developer preview · MIT

DeepSeek Harness
Everything is a plugin · Cordis kernel

As of 2026-11-09, DeepSeek Harness (dsh) is DeepSeek AI’s official open-source agent harness (MIT license): the layer around a model that handles orchestration, tools and context. Current version: 0.2.2-rc.1 (pre-release, 2026-11-04); needs Node.js 24.9 or newer. Run npx @deepseek-ai/dsh web, then enter your own API key and Base URL under Settings → Models → Custom model API. Agent = Model + Harness. This is DeepSeek’s own shell, not a random pip adapter.

Updated 2026-11-10

One key for Claude, GPT and Chinese models — put what this page describes into practice.

#npx dsh web#Cordis#4 modes#MIT preview

Highlights

npx

@deepseek-ai/dsh web

Serves Web UI at 127.0.0.1:3080, needs Node.js 24.9 or newer. Or clone + pnpm.

Cordis

Plugin kernel

github.com/cordiverse/cordis — mount, deps, events.

Minimal

Official bench mode

The official site lists bash + str_replace_editor; the 0.2.2-rc.1 UI describes a terminal plus image reading, with context compaction. That’s the mode behind the Flash 0731 bench footnote.

break

Preview warning

README: THERE WILL BE COMPATIBILITY-BREAKING CHANGES.

What dsh is

dsh is an open-source project from DeepSeek AI (GitHub: deepseek-ai/deepseek-harness, repository created 2026-08-13) built on an everything-is-a-plugin architecture powered by Cordis; besides dsh web it also starts as headless, acp or sdk. Since 0.2.2-rc.1 the interface is also available in Japanese, Spanish, French, German, Russian and Brazilian Portuguese. May hiring said Model + Harness = Agent and named Claude Code / Cursor / Codex as the layer to fill. August’s developer preview is that line shipping.

Why it is trending

Do note the distinction from the third-party "deepseek-harness" protocol wrappers on PyPI: the official project lives under the deepseek-ai GitHub organisation in a repository named deepseek-harness. Community plugins are usually tagged with the dsh-plugin topic.

Timeline

2026-05

Public Code Harness hiring.

2026-07-31

Flash 0731 changelog names upcoming Harness Minimal.

2026-08

dsh developer preview + Cordis site.

Confirmed vs caution

Confirmed

Checked against 0.2.2-rc.1 and the official site: preview, open source under MIT; four modes — Standard, PTC (Code mode on the English site), Minimal and Creator; an append-only session log of everything the model sees, viewable by source in Trajectory; the bench footnote in DeepSeek’s changelog (Minimal mode + max effort).

Caution

Three things to know before you start: dsh is a developer preview, and the README warns of compatibility-breaking changes; its feature scope is not equivalent to Claude Code; the software is open source under MIT, and model tokens are paid for separately.

How to choose

Need to reproduce DeepSeek’s official scores

Use the setup DeepSeek documented in its changelog: Minimal mode + max effort + temperature=1.0 + topp=0.95.

Need a daily coding agent

Standard mode for everyday tasks, PTC mode for tasks that call tools in batches and then filter or summarize the results, knowing it’s a preview. Or Claude Code / OpenCode.

Install and connect your own API

As of 2026-11-09, per the official 0.2.2-rc.1 docs: 1) install Node.js 24.9 or newer; 2) run npx @deepseek-ai/dsh web, which opens the Web UI at http://127.0.0.1:3080 (an SSH launch only prints the address; --no-open skips the browser); 3) go to Settings → Models → Add model provider, switch the card to Custom model API, and fill in a lowercase Provider ID (permanent once created), Base URL, API protocol and API key, plus at least one model ID; 4) save and pick the model in the model picker. Changes apply on the next request, no restart needed. The API protocol is one of OpenAI Chat Completions, OpenAI Responses or Anthropic Messages. A provider speaks one protocol, so a gateway that serves two needs two providers. The settings are written to $DSH_HOME/profiles/web/cordis.patch.yml (fields api, baseURL, apiKeyEnv) and the key is stored in $DSH_HOME/.credentials.yaml. To run from source: git clone github.com/deepseek-ai/deepseek-harness, then pnpm install, pnpm run build, pnpm dsh web. The official site deepseek.com/harness also offers desktop downloads for macOS (Apple silicon) and Windows (64-bit).

On QCode

dsh runs on your own key: one QCode key works in dsh for DeepSeek / GLM / Kimi / Qwen / GPT (OpenAI-compatible) and Claude (Anthropic-compatible), billed per token, with each model’s price on /models; if you would rather not pay monthly, quota packs are on /pay-as-you-go. Create two providers: OpenAI Chat Completions with Base URL https://api.qcode.cc/openai/v1 for GPT plus GLM, Kimi, DeepSeek and Qwen (for example gpt-6-sol, deepseek-v4-pro, glm-5.3, kimi-k3, qwen3.8-max; add the Chinese model IDs by hand), and Anthropic Messages with Base URL https://api.qcode.cc/api for Claude (for example claude-sonnet-5). Claude uses the Anthropic protocol; do not end a Base URL with /, and enter model IDs exactly as on /models. As of 2026-11-09, QCode /models lists deepseek-v4-flash, deepseek-v4-pro and deepseek-v4.1-flash; point dsh at them like any OpenAI- or Anthropic-compatible endpoint. Step-by-step setup, a test command and a troubleshooting table: https://docs.qcode.cc/en/docs/ide/deepseek-harness.

FAQ

What is DeepSeek Harness?

DeepSeek Harness (dsh) is DeepSeek AI’s official open-source agent harness (repository deepseek-ai/deepseek-harness, MIT license). It is software, not a model: it handles orchestration, tools and context, and you connect a model API separately. Start the web version with npx @deepseek-ai/dsh web, or download the desktop app from the official site. The official status is developer preview, and the README warns of compatibility-breaking changes.

What is Cordis?

Cordis is the plugin kernel under dsh (github.com/cordiverse/cordis): dsh uses an everything-is-a-plugin architecture, so tools, skills and the interface can all be extended with plugins. The official site says you can install existing plugins or have the agent write new ones through chat in Creator mode; community plugins usually carry the dsh-plugin topic.

What is Minimal mode for?

Testing the model itself: per the 0.2.2-rc.1 UI, Minimal mode gives the agent a terminal and image reading, with context compaction, and is meant for testing and comparing basic performance. DeepSeek’s 2026-07-31 and 2026-08-21 changelogs say its Code Agent benchmarks used it as the framework, with the max effort level, topp=0.95 and temperature=1.0. To score the model, not the kitchen sink.

Can dsh use a third-party OpenAI-compatible or Anthropic-compatible API?

Yes. Settings → Models → Add model provider → Custom model API is meant for a relay, a company gateway, a self-hosted server or any provider absent from the catalog, and the protocol can be OpenAI Chat Completions, OpenAI Responses or Anthropic Messages. The software is open source under MIT, and model API calls are paid per token. On QCode, for example, GPT and Chinese models use https://api.qcode.cc/openai/v1 and Claude uses https://api.qcode.cc/api, with the same key.

Can dsh replace Claude Code?

Same category, preview maturity: the official Web UI guide says the dsh agent can read and edit workspace files, run commands, delegate work and maintain a plan, while the README warns of compatibility-breaking changes. Before switching, try both side by side on your own project; the same QCode key works in Claude Code and in dsh.

How does DeepSeek Harness relate to harness engineering?

As concept and implementation: harness engineering covers the generic concept of a harness — the scheduling, tooling and context layer that sits around a model — explained in the on-site /harness-engineering-guide; dsh is DeepSeek’s own concrete implementation of that idea.

Sources

deepseek.com/en/harness/ ; github.com/deepseek-ai/deepseek-harness ; 0731/0813/0821 changelogs ; npm @deepseek-ai/dsh ; GitHub docs/user/guide/providers.md ; GitHub 0.2.2-rc.1 source (Web UI mode names) ; docs.qcode.cc.

Run DeepSeek Harness on your QCode key

Sign up for QCode, get your API key and connect Claude, GPT and leading Chinese models in dsh through Custom model API. Billed per token; per-model rates are listed on /models.

Try first, then decide

Not sure which tier? Start with Starter ($9.00/mo) and upgrade when you're happy — the unused value of the old plan goes back to your balance.