Deep-diving the GitHub dsh-plugin Top 50 - mapping a 1.7M-star plugin ecosystem

DeepSeek Harness (dsh) went public on August 13, 2026. DeepSeek's open-source agent harness, built around "Everything is a Plugin," has watched the GitHub topic dsh-plugin grow to 26,895 repositories in just two and a half months. This post is an ecosystem map: I read the READMEs of the top 50 repos by stars (1,723,509 combined) and grouped them by category. Snapshot date: 2026-11-05, sorted by stars. Method: GitHub Search API for the list, raw README fetches for every single repo, then summary and translation. All numbers and feature claims come straight from the source material that day.

One honest caveat first. High stars does not mean "DSH-exclusive." The #1 repo (rea) is a general-purpose reverse-engineering tool that merely carries the topic, and legacy favorites like Reactive Resume (#7, born 2020) and PicGo (#16, born 2017) ride the topic into the ranking too. Still, looking across all 50, the phrase "DSH plugin ecosystem" starts to mean something quite specific.

Core & Frameworks

2. deepseek-ai/deepseek-harness - ⭐272.9k

The center of this map. DeepSeek's official open-source agent harness, built on the Cordis-based "everything is a plugin" architecture, complete with a published paper on the paradigm (arXiv 2608.25512). With Node.js installed, npx @deepseek-ai/dsh web boots a local Web UI (default 127.0.0.1:3080) in one line. The README warns this is a developer preview with compatibility-breaking changes ahead.

• everything-is-a-plugin: models, tools, sandboxes, session storage, UI - even the agent loop itself
• one-command boot: npm one-liner opens the web GUI

5. ruvnet/ruflo - ⭐80.9k

A meta-harness for Claude Code and Codex that defines "agent = model + harness." Ships 100+ specialized agents, collaborative swarms, adaptive memory that learns per task, and federated cross-device communication. Descends from claude-flow (first commit June 2025) - the original-harness lineage.

• swarm + adaptive memory: learning accumulates across sessions
• MCP-first: npx ruflo@latest mcp start connects any MCP client

12. esengine/DeepSeek-Reasonix - ⭐36.1k

An open-source coding agent for terminal, desktop, browser and editor. Reads your project, edits files, runs commands and tests. You choose between ask-before-every-step and autonomous-within-limits, and edited files can be rewound. MIT, local, bring-your-own-model.

• rewindable edits: step-wise undo of what its tools changed
• Studio desktop app: a friendly entry point for newcomers

38. Q00/ouroboros - ⭐6.3k

An "Agent OS" that gets smarter on its own. For replayable coding workflows it runs interview-gated, staged evaluation with a budgeted evolution loop - and the grading command and expected result never enter the success contract handed to the agent. Supports 14 runtimes: Claude Code, Codex CLI, Gemini CLI, OpenCode, Copilot, Kiro and more.

• loop engineering: failed generations feed the next one
• one-line install: curl script, then ooo setup inside your coding agent

Clients & UI

14. anywhere-labs/dsh-desktop - ⭐34.9k

A community desktop client that wraps DSH into a native app - "the desktop itself is a plugin." Handles window, tray, terminal and updates while pinning and running a specific upstream version verbatim. Notably, the README explicitly disclaims any affiliation with DeepSeek.

• installers for 3 platforms: Windows (NSIS), macOS (DMG), Linux (x64/arm64)
• pinning strategy: runs a fixed dsh version for predictability

21. dataelement/dsh-desktop - ⭐17.3k

Dataelem's desktop agent app built on the Harness framework. Auto-starts and cleans up the runtime, keeps profiles, plugins and sessions outside the app directory so upgrades survive. macOS builds are signed and notarized; Windows installers are code-signed. A PPT mode (16 templates, 192 layouts) turns source material into editable PPTX.

• phone pairing: QR scan to continue sessions on mobile (Cloudflare tunnel + 5-minute pairing code)
• safe mode: on plugin conflicts, identifies the culprit and restarts in an isolated profile

25. zhu1090093659/dsh-web - ⭐10.1k

An all-in-one plugin and theme bundle for the DSH Web GUI that attaches via the official profile mechanism without touching upstream source. Adds a task automation kanban (cron schedules, GitHub issue→PR), web SSH/SFTP, full session search and token cost monitoring in one shot.

• "Liangshen" mode: minimal persona + programmatic tool calling to cut context overhead
• remote ops: app QR pairing, web SSH terminal for cluster-scale commands

27. ccch1mneyyy/dsh-Go - ⭐9.9k

DSH in your pocket: an Android client. Without root or Termux, a built-in Linux environment runs DSH on the phone itself, or you switch to your PC's DSH in one tap. Hit #1 on Trendshift's all-language daily chart on 2026-10-30.

• local run + remote control: phone-standalone and PC takeover in one app
• conversation replay: full chats become browser-playable E2E-encrypted links

36. ccch1mneyyy/dsh-TUI - ⭐6.9k

Same author, terminal edition: pixel-whale header, live work status, streaming thinking, double-Esc time rewind, a context progress bar and a TPS gauge. The key design point is zero core patches - it consumes only DSH's public exports and is evolving into an app with its own entry point.

• clean uninstall: no patches left behind
• UI craft: split diff cards, one-line compaction status in narrow panes

46. omdsh-dev/DSH-better-sidebar - ⭐4.9k

A dual workbench: native right sidebar plus a self-drawn bottom panel, with the ctx.betterSidebar service open to every plugin. File management (lazy tree, git-status colors, global search), an editable CodeMirror editor (Mermaid, sandboxed HTML preview), live subagent topology graphs and Codex-style side chats.

• side chat (beta): threads that inherit full parent context, one-click promotion to top-level
• task panel: background tasks with exit codes, live output, force-kill

Markets & Registries

19. awesome-dsh-plugin/awesome-dsh-plugin - ⭐22.4k

The curated list of DSH plugins installable via dsh plugin add (each declares a dsh.bundle manifest). The fun part: the list doubles as a plugin market index - add its npm-served dsh-plugin-catalog/index.json in the Discover tab and the host searches it directly.

• an awesome list that is also an index: mirrored on GitHub Pages and dshmarket.com
• open contribution: PRs welcome

24. dsh-market/dsh-market - ⭐10.8k

The plugin market embedded inside DSH as a tab. Browse 4,200+ community plugins with category filters, ratings and sorting; one-click install, mostly live-reload without restart. One line: dsh plugin --profile web add dshmarket.

• host-aware compatibility: reads engines.dsh and lockstep peer deps, filters only true mismatches
• giscus comments: every card opens a GitHub Discussions thread

28. anbeime/skill - ⭐9.2k

Self-described most-complete, fastest-updating skill store. GitHub Actions crawl 6 upstreams (OpenAI Skills, VoltAgent, Awesome DSH Plugin, etc.) daily, holding 182 official skills plus 69 Chinese local skills - 251 total. Caveat: a CC-BY-4.0 badge in the README but no license file in the repo - check before redistributing.

• automated pipeline: python main.py --once or --daemon
• validation tooling: SKILL.md spec validator included

42. iflytek/skillhub - ⭐5.4k

iFlytek's enterprise-grade open-source agent skill registry. Built for on-premise deployment behind your firewall: publish skill packages to namespaces, search, CLI install, governed by RBAC and audit logs. Docker and Kubernetes supported.

• private skill sharing: public-registry polish, on-prem constraints
• curated starter collection: vetted community skills ship by default

Design & Creative

3. nexu-io/open-design - ⭐106.0k

"The 1st collaborative AI design agent workspace." Start with a brief; your coding agent or OpenDesign Cloud creates a prototype that follows your design system; the team shares feedback and refines it in the same project. Apache 2.0 - skills, templates, design systems and plugins are all inspectable and adaptable. DSH is a first-class native runtime.

• brief → prototype → collaboration in one workspace
• 100K stars, celebrated with an OpenDesign Go credits event

15. alchaincyf/huashu-design - ⭐28.7k

"Type. Hit enter. A finished design lands in your lap." An HTML-native design skill that ships launch animations, clickable app prototypes, editable decks and print-grade infographics in 3–30 minutes. Fully MIT since 2026-05-14, including commercial use.

• brand awareness: feed it your logo, palette and UI screenshots and it reads the brand's temperament
• ~90-style library with samples, to keep output out of AI-slop territory

34. ZSeven-W/openpencil - ⭐7.1k

"The world's first open-source AI-native vector design tool." Describe a UI in natural language and it appears on an infinite canvas with streaming animation. The highlight is concurrent agent teams - an orchestrator decomposes a page into spatial sub-tasks that multiple agents draw in parallel (hero, features, footer).

• Design-as-Code: the canvas is the artifact
• built-in MCP server: one-click install into Claude Code, Codex, OpenCode, Kiro, Copilot

39. JimLiu/baoyu-design - ⭐6.1k

Brings claude.ai's artifact types - editable slide decks, design canvases, design systems, data dashboards, motion films - to local agents. No website, no subscription, no upload step: your machine's agent does the work and every artifact stays in your repo. The five editors also ship as DSH plugins on npm.

• motion-director / dataviz: film direction and chart guidance as separate skills
• self-contained outputs: each artifact opens in its own local editor

Memory & Context

8. volcengine/OpenViking - ⭐41.6k

Volcengine's (ByteDance) self-evolving context database. Knowledge, memory and skills all live in a viking:// virtual filesystem that agents navigate with ls/tree/read/grep - and humans can open and inspect. A 3-layer read path (L0 summary → L1 outline → L2 full text) saves tokens.

• anti-blackbox: not an embedding blob but a browsable filesystem
• ov compile: compiles sessions into wikis, knowledge graphs and reports

22. EverMind-AI/EverOS - ⭐13.8k

A local-first memory runtime that stores conversations, files and agent trajectories as canonical Markdown, syncing local SQLite and LanceDB indexes for fast retrieval. No MongoDB, Elasticsearch or Redis - just Markdown + SQLite + LanceDB; hand-edit the .md files and a cascade watcher syncs the index.

• dual track: user episodes/profile and agent cases/skills as separate first-class surfaces
• reflection: offline consolidation of episode clusters between sessions

23. MemTensor/MemOS - ⭐12.3k

A memory operating system for LLMs and agents ("MemOS 2.0 Stardust"). A unified memory API that is graph-structured and inspectable rather than a black-box embedding store; multi-modal memory; knowledge bases as composable "memory cubes." DSH plugins (local and cloud) shipped in August 2026.

• benchmarks: LoCoMo 88.83, LongMemEval 89.20; claims the lead in a 14-product memory eval
• 4 entry points: Cloud API / docker compose / OpenClaw / DSH-Hermes local plugin

30. plastic-labs/honcho - ⭐7.9k

Plastic Labs' memory infrastructure for stateful agents. Store → Reason → Query → Inject: messages go in, background reasoning maintains "peer" representations of users, agents, groups and projects over time. Reasoning-first memory rather than chunk matching is the differentiator. Managed, CLI-local and self-hosted options.

• peer-centric model: tracks what one peer knows about another
• ops tooling: honcho doctor and friends for deployment checks

50. bowenliang123/dsh-context - ⭐4.2k

Context insight and management for DSH. A first-class sidebar panel with 7-day token/cost KPIs, activity heatmap, composition donut and session cards; a /context command; and a Fleet tab showing the agent family (parent and each sub-agent) as live cards. Starred and recommended by two DSH team members.

• context transparency: composition, compaction and evolution visualized
• fleet view: multi-agent sessions tracked card by card

Skills & Distillation

9. alchaincyf/nuwa-skill - ⭐37.5k

Type a name; get that person's way of thinking distilled into an agent skill. "Why stop at colleagues?" - Jobs, Musk, Munger and Feynman go through a research → extraction → verification pipeline. Not quote replay: it extracts 5 layers including expression DNA, decision heuristics, anti-patterns and known limits.

• 5-layer distillation: down to what the person would not do
• 50+ compatible runtimes: Claude Code, Codex, Cursor, DSH and more

13. freestylefly/awesome-gpt-image-2 - ⭐35.5k

"Prompt as Code" - an industrial prompt engine and template library for GPT Image 2/2.5: 500+ reverse-engineered cases, 20+ templates, plus a 2.5 spotlight comparing Sunburst and Flare on identical prompts. Every case publishes its full prompt and generation record.

• reproducible gallery: prompt, settings and result as one set
• side-by-side comparisons: model differences you can drag to verify

17. titanwings/distilly - ⭐26.4k

The ancestor of nuwa-skill, formerly "Colleague Skill." Distills a person's experience, judgment, voice and ways of working into a reusable Person Profile from messages, documents, interviews and public sources. For the colleague who quit, the family drifting apart, the author you will never meet - while explicitly not claiming to clone the person behind the profile.

• source-grounded: observable evidence only
• wide targets: colleagues, partners, family, fictional characters, yourself

43. QingYunA/answer-me-with-html - ⭐5.3k

Answers hard questions with a readable one-page HTML instead of a wall of text. The model writes only the text (21% of tokens); a bundled CLI generates the SVG diagrams (47%), CSS (15%) and tags (17%) in ~50ms - about 1/8 the tokens of hand-written HTML. The published token measurements make it even better.

• the honest way to save tokens: give the model only what it is good at
• single-file offline: the page is the artifact

Productivity & Documents

4. tt-a1i/archify - ⭐93.5k

Turns ideas, plans or codebases into explorable interactive HTML diagrams. Start with one sentence ("diagram the web request flow") and refine with follow-ups ("highlight the cache-miss path"). Point it at a repo and it produces evidence-based architecture maps pinned to commits. 130K+ cumulative installs via skills.sh.

• repo-less start: from travel itineraries to system architectures
• explainer mode: conclusion-first one-pagers with tables, charts and citations

7. reactive-resume/reactive-resume - ⭐45.3k

A free, open-source resume builder: live PDF preview, PDF/Word/MD/JSON export, Docker self-hosting. v6 adds Write/Design/Check modes, 18 ATS-safe templates and a public ATS checker. No tracking, no ads, no paid tier.

• application tracking built in: List, Board, Insights and Calendar views
• your-data-yours philosophy: AI features only through a provider you connect

10. DayuanJiang/next-ai-draw-io - ⭐36.7k

Draw and edit real draw.io diagrams by chatting. Built-in AWS/Azure/GCP/Kubernetes icon libraries; upload a screenshot and the AI redraws it; every AI change lands as a version card you can compare, restore or undo. Output is standard .drawio, so hand-editing still works. Web, desktop and MCP server.

• selective edits: pick shapes and change only those
• version cards: compare, restore, undo for every AI change

16. Molunerfinn/PicGo - ⭐27.5k

The 2017-vintage elder: upload an image once, get the link copied automatically. Qiniu, Tencent COS, UPYUN, Aliyun OSS, GitHub, SM.MS and Imgur out of the box; S3, R2 and MinIO via plugins. It now directs new image hosts to the plugin system instead of adding them to core.

• editor integrations: Obsidian, Typora, VS Code
• plugin ecosystem: build your own image-host plugins

29. YaoApp/yao - ⭐8.3k

Born 2021. "All your agents and workspaces in one place" - a self-hosted agent management platform on your own devices. Chats become kanban tasks, workspaces stay isolated, and outputs accumulate into documents that become your knowledge base. DSH integrated; desktop, Android (beta), browser and an open API (SSE/WebSocket).

• multi-device: same board from anywhere
• agent-facing guide: a standard fact sheet so AI agents install it correctly

31. lingfengQAQ/webnovel-writer - ⭐7.8k

A long-form webnovel writing system with one goal: keep the AI remembering settings, landing foreshadowing and holding the outline across hundreds of chapters. A complete pipeline from init, volume planning and chapter writing to review, memory sedimentation, status queries and a read-only dashboard. master is the Claude Code v6; the v8 branch is a DSH-based writing workbench.

• serialization consistency system, not a generate-and-forget generator
• memory sedimentation: settings accumulate as chapters pass

32. Devin-AXIS/iPolloWork - ⭐7.4k

An enterprise-grade, local-first agent workbench. Connects Codex, DeepSeek Harness and OpenCode on one control surface; unifies plugins, skills, tasks, schedules and editable outputs across code, docs, decks, design and video.

• runtime-neutral: explicit compatibility boundaries preserve each ecosystem's strengths
• outcome-centric teamwork: describe the outcome, agents plan and execute, your team reviews and approves

Agent Tools

1. morluto/rea - ⭐431.5k

The #1 of this map: an MCP server that reverse engineers anything with agents, from app behavior down to native binaries. With Hopper/Ghidra/IDA integrations it analyzes binaries, JS/Electron apps, .NET assemblies and websites - locally - and attaches evidence and limitations to every conclusion. npx rea-agents setup registers it across agents in one line.

• feature cloning: dissect a feature you like, source-free, into your own version
• evidence-backed: every conclusion ships with proof and limits

26. Tencent/BrowserSkill - ⭐10.0k

Tencent's tool that connects agents to the Chrome or Edge you already use, leveraging your existing logins. Work happens in a separate Agent Window; tabs are explicitly borrowed and returned. Any shell-capable agent (Codex, Claude Code, Cursor...) attaches via the same bsk CLI; DSH gets a dedicated plugin with native browser tools.

• debugging evidence: actions linked to requests/responses/console; silent MP4 recording
• shared control: pause, let a human fix things, the agent reconciles and continues

33. superdesigndev/treg - ⭐7.2k

"OpenRouter, but for agent tools." One base URL and one token unlock a curated catalog of thousands of endpoints - SEO, social, enrichment, scraping, image/video generation - priced per call from a cent. It carries the accounts behind subscription walls (Semrush at $139/mo and friends) and bills fractions of cents.

• credential custody: no provider signups; team keys never leave the server
• ask for the task, not the tool

47. liustack/modlens - ⭐4.8k

"Gives a text-only model sight, and just paste the image." DeepSeek flagships and GLM-5.3 are text-only; ModLens reads images pasted straight into chat and returns structured JSON evidence - OCR, layout, semantics. No save-to-file-and-pass-a-path dance.

• bolt-on vision: multimodal behavior for text-only models
• ModSearch: a free web-search plugin from the same author

Utilities & Tuning

20. HunYiSage/dsh-purge - ⭐18.3k

A DSH "jailbreak" tool supporting per-model prompt swapping. The README self-describes as a cybersecurity red-team tool: the open-source build is local-lab only (loopback/RFC1918 targets), game cheats are prohibited and enforced at the prompt and session gates. Still, its essence is guardrail removal - keep it in isolated environments, keep it away from production.

• declared boundaries: local-lab-only and no-cheats, enforced in code
• community handover: maintained by a team with original attribution preserved

35. yjh051108/dsh-routing-suite - ⭐7.0k

A three-piece suite: a runtime injector (hot reload, unload, route self-healing, no restart) + task-aware reasoning-mode routing presets (P1–P23 measured) + a graded task protocol (brainstorm → spec → check-ins → wrap-up → final verification, with a red-team gate and audit endpoint). The research tone - "models get lazy; we measured it and built countermeasures" - is refreshing.

• measurement first: experiment data published per preset
• audit endpoints: protocol violations are tracked

44. T-Auto/dsh-ops - ⭐5.2k

A token-efficiency plugin: adds ops-series tools and hides some official ones, loading scenario-specific tools (PDF/Word reading, bulk file edits) dynamically without bloating the system prompt. For native mode it returns compact numbered output to break error loops; for PTC mode, structured JSON tools and an 18-API SDK.

• low-intrusion design: flip the switch off and the original tools return
• published benchmarks: per-harness scores and overhead comparisons

48. kingcharleslzy-ai/dsh-purge - ⭐4.5k

The community-maintained edition of dsh-purge (original by YuJunZhiXue; MIT and attribution preserved). 84 host patches, 25-layer state self-checks, a full-page settings sidebar. Update channels offer optional mirrors, but tokens never go to the mirror and installs verify the official source. Bundles a Qwen Studio integration for local image generation (MLX, ComfyUI).

• compat management: per-version verification records (DSH Desktop 0.2.2-rc.1 etc.)
• backup & rollback: local backups before updates

Fun & Widgets

37. MeteorNOX/DeepSeek-Balance-Whale-Widget - ⭐6.9k

A whale widget living in the bottom-right of the DSH web UI: API balance, today's usage and per-turn consumption in bubbles you can fully customize - click sequences, modular layout, weighted appearance, random phrases and images. One-line dsh plugin install, no session token needed.

• pure fun that is also useful: balance monitoring as a character
• custom bubbles: a widget that became a platform

40. Fei2411/DS-Bot - ⭐5.6k

A Windows installer (~212MB) for a team of long-lived bots. Double-click to install - no admin rights, no Node, no DSH needed; a DeepSeek account or API key is enough. Supports main bots, bot-to-bot messages and group chats; macOS, Linux and DSH plugin paths exist too.

• dependency-free install: download, double-click, done
• bot society: bots talk to each other, group chats included

41. PC2005-cloud/dsh-pet - ⭐5.6k

A desktop pet living inside DSH: idle breathing, random actions (dozing, solving a Rubik's cube, eating hotpot), screen roaming, bouncy clicks, drag-and-throw physics and a right-click menu - 100+ transparent animations chained seamlessly. Reacts to DSH session events and shows balance-tier animations and speech bubbles.

• three-piece architecture: prompt recipe → asset generation chain → plugin, so anyone can clone and make their own pet
• desktop escape: leaves the browser as a transparent topmost window

45. elysia395/dsh-wallpaper-engine - ⭐5.1k

Brings local Wallpaper Engine scene, web and video wallpapers into the DSH chat interface. The bundled open-source WebWallGL engine renders particles, skeletal animation, SceneScript, audio reactions and mouse interaction in real time, degrading to MP4 or static frames on failure. Registers no tools and injects no prompt text - zero token overhead.

• the zero-token aesthetic: backgrounds stay backgrounds
• liquid-glass sub-package: individually disable-able

49. crafter-station/petdex - ⭐4.3k

A public gallery of animated companions for coding agents like Codex. Three parts: a web gallery (petdex.dev), a CLI (npx petdex install boba), and a desktop app that floats a pet reacting to your coding agent's activity in real time. Every pet is a folder; every folder is a Pokédex entry.

• gallery → CLI → app: submission, review and installation in one pipeline
• agent-reactive: the pet responds while you code

Models & RAG

6. Ebony-Vinyl/dsh-our-free-model - ⭐61.0k

Install it and frontier models - DeepSeek V4.1 Flash, Kimi K3 - become free and unlimited, with no login, signup or API key. The README is transparent about sources: the free lane comes from OpenCode's Zen gateway and Kilo AI's public gateway. Zero-dependency pure plugin mounting; upgrades verify SHA-256, swap atomically and roll back automatically.

• source transparency: it never hides where the free rides come from
• but be careful: free lanes can close anytime, and upstream ToS risk is yours

11. Tencent/WeKnora - ⭐36.5k

Tencent's LLM knowledge platform. Enterprise document understanding, semantic retrieval and reasoning unified so RAG lookup, multi-step agent tasks and a wiki all work on the same knowledge bases. Skills install from ClawHub/SkillHub/Git/ZIP into session-persistent Docker/E2B sandboxes; a BrowserSkill extension drives your own Chrome.

• three modes, one base: RAG, agent and wiki share knowledge
• session-persistent sandboxes: a living execution environment alongside chat

Learning

18. walkinglabs/learn-harness-engineering - ⭐23.2k

A project-based course on the environment, state management, verification and control mechanisms that make AI coding agents reliable - available in 15 languages including Korean. An August 2026 section dissects frontier harness designs (Pi; Claude Code's four-layer memory, five-level compaction, hooks and sub-agent isolation; Codex) through a five-subsystem framework.

• Korean edition exists: the entry barrier is low
• reverse-engineering approach: learn by tearing down finished harnesses

Trends

1. Speed itself. Within two weeks of the dsh launch (8/13), two desktop clients, a market, a TUI, pets, wallpapers and a context viewer appeared. 14 of the Top 50 were born in August.
2. A Chinese-community-led ecosystem; Korean coverage is a blank spot. Nearly half the READMEs are Chinese (many with English). That gap is part of why this post exists.
3. "Distillation" as a new genre. The nuwa-skill/distilly line - turning a person's thinking into skills - is a category I have not seen in other agent ecosystems.
4. Big tech funds infrastructure, individuals fund UI and fun. Tencent (knowledge platform, browser tool), Volcengine (context DB), iFlytek (enterprise skill registry); individuals do clients, widgets and distillation.
5. Legacy projects join via the topic. PicGo (2017), Reactive Resume (2020), Yao (2021), honcho (2023) - the topic itself became a distribution channel.
6. Memory/context is already a red ocean. Five entries in the Top 50 alone (OpenViking, EverOS, MemOS, honcho, dsh-context). Their differentiators all converge on the same idea: transparency - Markdown, filesystems, graphs.

Before You Install

• The jailbreak pair (two dsh-purge repos): local-lab-only or not, they are guardrail-removal tools. Keep them in isolation; there is no reason to run them against real services.
• The free-model plugin (dsh-our-free-model): clean engineering, but free lanes close without notice and upstream ToS risk lands on you.
• Popular repos without licenses (anbeime/skill etc.): README badges sometimes differ from actual license files. Open LICENSE before redistributing.

Appendix: all 50

Open the table (rank · repo · stars · language · license · created)

#	Repo	Stars	Lang	License	Created
1	morluto/rea	431,542	TypeScript	MIT	2026-04
2	deepseek-ai/deepseek-harness	272,851	TypeScript	MIT	2026-08
3	nexu-io/open-design	105,980	TypeScript	Apache-2.0	2026-04
4	tt-a1i/archify	93,538	JavaScript	Apache-2.0	2026-04
5	ruvnet/ruflo	80,897	TypeScript	MIT	2025-06
6	Ebony-Vinyl/dsh-our-free-model	61,047	JavaScript	MIT	2026-09
7	reactive-resume/reactive-resume	45,343	TypeScript	MIT	2020-03
8	volcengine/OpenViking	41,567	Python	AGPL-3.0	2026-01
9	alchaincyf/nuwa-skill	37,524	Python	MIT	2026-04
10	DayuanJiang/next-ai-draw-io	36,690	TypeScript	Apache-2.0	2025-03
11	Tencent/WeKnora	36,514	Go	custom	2025-07
12	esengine/DeepSeek-Reasonix	36,081	Go	MIT	2026-04
13	freestylefly/awesome-gpt-image-2	35,471	JavaScript	MIT	2026-04
14	anywhere-labs/dsh-desktop	34,899	TypeScript	MIT	2026-08
15	alchaincyf/huashu-design	28,707	HTML	MIT	2026-04
16	Molunerfinn/PicGo	27,462	TypeScript	MIT	2017-12
17	titanwings/distilly	26,380	Python	MIT	2026-03
18	walkinglabs/learn-harness-engineering	23,166	TypeScript	MIT	2026-03
19	awesome-dsh-plugin/awesome-dsh-plugin	22,434	JavaScript	CC0-1.0	2026-08
20	HunYiSage/dsh-purge	18,287	JavaScript	Apache-2.0	2026-08
21	dataelement/dsh-desktop	17,285	JavaScript	MIT	2026-08
22	EverMind-AI/EverOS	13,810	Python	Apache-2.0	2025-10
23	MemTensor/MemOS	12,259	TypeScript	Apache-2.0	2025-07
24	dsh-market/dsh-market	10,813	TypeScript	MIT	2026-08
25	zhu1090093659/dsh-web	10,081	TypeScript	Apache-2.0	2026-08
26	Tencent/BrowserSkill	10,028	TypeScript	MIT	2026-06
27	ccch1mneyyy/dsh-Go	9,949	Dart	AGPL-3.0	2026-10
28	anbeime/skill	9,162	Python	none	2026-02
29	YaoApp/yao	8,265	Go	custom	2021-09
30	plastic-labs/honcho	7,945	Python	AGPL-3.0	2023-09
31	lingfengQAQ/webnovel-writer	7,834	Python	GPL-3.0	2026-01
32	Devin-AXIS/iPolloWork	7,379	TypeScript	custom	2025-08
33	superdesigndev/treg	7,153	Python	custom	2026-07
34	ZSeven-W/openpencil	7,067	Rust	MIT	2026-02
35	yjh051108/dsh-routing-suite	7,000	JavaScript	MIT	2026-08
36	ccch1mneyyy/dsh-TUI	6,919	TypeScript	MIT	2026-08
37	MeteorNOX/DeepSeek-Balance-Whale-Widget	6,878	JavaScript	MIT	2026-08
38	Q00/ouroboros	6,318	Python	MIT	2026-01
39	JimLiu/baoyu-design	6,101	JavaScript	custom	2026-06
40	Fei2411/DS-Bot	5,614	JavaScript	Apache-2.0	2026-10
41	PC2005-cloud/dsh-pet	5,575	TypeScript	MIT	2026-08
42	iflytek/skillhub	5,388	Java	Apache-2.0	2026-03
43	QingYunA/answer-me-with-html	5,309	JavaScript	MIT	2026-10
44	T-Auto/dsh-ops	5,234	JavaScript	MIT	2026-10
45	elysia395/dsh-wallpaper-engine	5,115	JavaScript	MIT	2026-08
46	omdsh-dev/DSH-better-sidebar	4,897	TypeScript	MIT	2026-08
47	liustack/modlens	4,766	TypeScript	MIT	2026-02
48	kingcharleslzy-ai/dsh-purge	4,497	JavaScript	MIT	2026-10
49	crafter-station/petdex	4,324	TypeScript	MIT	2026-05
50	bowenliang123/dsh-context	4,164	TypeScript	Apache-2.0	2026-08

Closing

Restating the method: the list comes from the GitHub Search API (topic:dsh-plugin, sort=stars desc, fetched 2026-11-05); each summary is a translation/condensation of the repo's raw README. Sources mix Chinese and English, so some nuance is inevitably lost. Star counts drift daily - this is a snapshot of a moment, not the truth of a ranking.

Reading all 50, the essence of this topic is not "plugins" - it is the attempt to assemble anything on top of a harness: desktops, phones, pets, wallpapers, even a person's way of thinking. I am curious how this map will shift when I re-run it next quarter.