awesome-deepseek-harness

简体中文 · English · 日本語

The ultimate guide to DeepSeek Harness: resources, tutorials, plugins, and tools

This project follows a curated, quality-over-quantity approach to collecting excellent DeepSeek Harness resources and building a richer Agent ecosystem with AI practitioners.

If this project helps you, please consider giving it a ⭐. You can also follow @李不凯正在研究 on X for more hands-on Agent content.

Table of Contents

• Table of Contents
• Quick Start
  • Launch the Web UI
  • Run from Source
  • Use the Python SDK
  • Install Plugins
• Official Resources
  • Packages and Integrations
  • Source Repositories
  • Official Documentation
  • Community
• Community Resources
  • Guides and Analysis
  • Community Discussions
• Third-Party Clients
  • Desktop Apps and Distributions
  • Terminal, Mobile, and Web Experiences
• Selected Plugins
  • Workflows and Agents
  • Context, Sessions, and Input
  • Browser, Vision, and Interface
  • Sandboxing and Execution
  • Themes and Skins
• External Integrations
• Developer Tools
• Acknowledgements

Quick Start

DeepSeek Harness (also known as DSH or dsh) is an open-source Agent Harness project from DeepSeek AI. Built on Cordis, it follows an Everything is a Plugin architecture: model adapters, tools, session logs, interfaces, and the Agent Loop can all be composed and replaced through a plugin tree.

The latest verified official GitHub developer preview is 0.1.2-rc.1; both npm latest and next now point to 0.1.2-rc.1. This RC adds full-history turn navigation, exact token / duration statistics, subagent model selection, ACP standard controls, schedules, connection retry, and replaces one-way report with continuable two-way send_message. Public WebFetch is enabled by default with SSRF protection; the SQLite Session backend was removed (old data is retained and requires an older version to export); DeepSeek adapter requests include enabled plugin package names and versions by default (configurable off), while incremental Session-log upload remains opt-in and off by default. The official release also states that DSH has not been security-audited and that Sandboxes, Approvals, and Permissions do not guarantee isolation. DSH versions shown below are only the development or test baselines claimed by project authors and must not be treated as automatic compatibility with the latest preview.

Launch the Web UI

Install Node.js 22.19.x or 24+ (24+ recommended), then run:

npx @deepseek-ai/dsh web

Open http://127.0.0.1:3080 by default. Configure a model provider under Settings → Models, then create a session. See the official quick start and model provider guide for details.

Run from Source

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Use the Python SDK

The official Python SDK lets you call Harness programmatically through its bundled runtime, without requiring a system Node.js installation. It currently requires Python 3.10+. Refer to the official Python SDK guide for supported platforms and current limitations.

python -m venv .venv
. .venv/bin/activate
python -m pip install deepseek-harness-sdk

Install Plugins

web and headless are built-in distribution Profiles. External plugins join a target Profile as Bundles that declare dsh.bundle:

dsh plugin --profile web add <package-or-git-spec>
dsh --profile web --dump-config

When installing from a Git repository, pin a commit and inspect its installation scripts first. pnpm may require explicit approval for dependency build scripts, which execute outside the Agent sandbox. See the official plugin packaging and installation guide for the complete mechanism.

Official Resources

The official project provides an open-source repository, a companion paper, extensive reference documentation, and an actively maintained developer community.

Packages and Integrations

• @deepseek-ai/dsh: official npm launcher package for the CLI and Web UI
• deepseek-harness-sdk: official Python SDK for programmatic DSH integration

Source Repositories

• GitHub: source code, issues, releases, and contributors
• Paper: detailed paper on the Cordis-based product architecture

Official Documentation

• Chinese website: product positioning and core concepts
• Documentation: entry point for usage, plugin development, and architecture reference

Community

• GitHub Discussions: questions, usage discussions, and proposals
• DeepSeek Discord: official Discord community, primarily in Chinese
• "DeepSeek Harness": live X search results for DSH
• # dsh-plugin: DSH plugin projects on GitHub

Community Resources

Guides and Analysis

Community Discussions

Public long-form social posts with substantial arguments, practical detail, or first-hand context that supplement gaps in official materials.

Third-Party Clients

The following projects provide standalone user interfaces, distribution formats, or productized assemblies rather than a single tool capability.

Classification note: Distributions and forks directly reuse, modify, or repackage the complete DSH Runtime and cannot be installed with dsh plugin, so they are not plugins. Standalone clients connect to DSH through Web, RPC, ACP, or companion bridge plugins. Both remain important parts of the Harness ecosystem.

Desktop Apps and Distributions

Terminal, Mobile, and Web Experiences

Inclusion does not imply that a project is signed, notarized, self-contained, or production-ready. Check each project's README and Releases for current status.

Selected Plugins

Workflows and Agents

• dsh-toolkit: deterministic tools for time, encoding, JSON, calculations, CSV, regex, Markdown, Diff, and more.
• dsh-deep-research: adaptive deep-research orchestrator for DSH.
• dsh-101: learning mode for reading and understanding the official documentation inside DSH.
• dsh-auto-approval: classifies tool calls with rules and models, then returns allow / deny approval decisions.
• mstar-harness: Skill-driven Harness / Loop Engineering workflow plugin.
• dsh-agent-teams: Agent Teams capabilities for DSH.
• dsh-automation: runs scheduled standalone tasks in fresh root Agents and Sessions, preserving definition revisions, run history, and explicit workspace and permission boundaries.
• dsh-plannotator: annotates Agent plans section by section and submits structured feedback, with draft isolation, version binding, and stale-plan rejection.
• dsh-spec-collab: turns raw product requirements into Git-versioned Ready Specs co-reviewed by product, engineering, and both sides' AI; AI can submit Review Items and candidate patches, while humans retain confirmation and canonical-save authority. Apache-2.0 and npm 0.2.1, declaring DSH 0.1.1-rc.1 compatibility and shipping a test suite; it is still a same-day fast-moving project with no CI, GitHub Release, or independent-use evidence, so it remains Early. The plugin writes a collaboration ledger and separate Git repository under ~/.dsh/spec-collab and starts additional AI review Sessions; its HTTP management surface is loopback- and same-origin-only by default, pseudonyms are not authentication, and external exposure requires separate authentication plus a trusted reverse proxy.
• dsh-context-guard: captures requirements, prohibitions, acceptance criteria, and later corrections as a task contract, allows Goal / turn completion only when current successful evidence matches the current contract, and rebuilds or re-verifies certificates from native Session events. Apache-2.0, npm / Release 0.1.0, pinned to DSH 0.1.1-rc.2; the repository reports 104 tests plus native isolated acceptance on macOS and Windows. It appends no custom event types and is not a sandbox, memory system, or semantic proof engine; conservative command-effect parsing and evidence matching can still miss cases, and the same-day first release has no independent-use evidence, so marked Early.
• dsh-record-replay: records macOS desktop workflows and generates Skills; currently requires Xcode Command Line Tools and a separate local open-record-replay source checkout.
• dsh-science-workbench: reproducible science workbench that records cells, figures, feedback, and rerun lineage in a manifest, together with environment snapshots and input/output hashes; MIT and v0.1.1, still Early.
• dsh-omicos: connects OmicOS bioinformatics to DSH with a persistent Python / R kernel, capability catalog, background jobs, and live execution views; GPL-3.0-only and published to npm as 0.2.1. Analysis tools run with permission_mode: full and may start a local kernel; cloud-backed models and higher tiers require an OmicOS account.
• dsh-crew: dispatches real DSH workers from Claude Code or Codex with progress reporting, status shards, and tier policies; MIT and Release / npm next 0.1.0-rc.4, verified on DSH 0.1.1-rc.1 with MCP and package smoke checks. It writes to ~/.config/dsh-crew/status.d/, and external model services may require API keys; it remains a prerelease without separate test CI, so it is marked Early.
• dsh-trading: DSH workbench for trading research with deterministic indicators, CSV providers, and interactive charts; MIT and npm @dsh-trading/bundle@0.1.0. It exposes no order-execution interface and heuristically blocks fund-moving tools, but that filter is not a complete security boundary, so the project is marked Early.
• oh-story-dsh: brings 13 novel-writing Skills, seven specialist Roles, and a short-drama production workflow into native DSH Sessions, approvals, and a three-column creative workbench. MIT, npm / Release 0.1.3, Node.js 24+, and peer dependencies starting at DSH 0.1.1-rc.1; the Release reports 34 automated tests and Ubuntu / macOS / Windows CI. This version reuses the official Composer height, adds a scroll safety area and anchor-geometry regression coverage, and keeps expanded tasks, streaming state, and located messages above the Composer. The build excludes upstream login / CDP scrapers and the standalone dashboard. The plugin does not read model credentials or open its own listener, but it can read and write creative-project files and invoke specialist Roles within the calling Agent's visible tools, and it adds a DSH Web Server file route protected by Host, Origin, Session, path-containment, and optimistic-concurrency checks. First released on August 21 and explicitly verified only on DSH rc.1, it still lacks long-term maintenance and independent-usage evidence, so it is marked Early.

Context, Sessions, and Input