We read every repository's package.json and file list, then applied the rules from the official DSH publish spec. No language model guessed anything — every verdict stores the exact fields it was based on, so you can re-derive it yourself.
Percentages are computed over the 284 entries checked, not the full index. 6,962 entries have no package.json we could find — they are reported as not checked, which is not the same as fine.
Categories come from the DSH source tree, not from our imagination
Agent Core
The product spine: append-only session log, system-prompt assembly, tool registry, Agent interface and the default loop driver. No provider slots here — you extend it by listening to its events.
598 entries
Bundles
The distribution format: an npm package declaring dsh.bundle with a cordis.patch.yml that inserts or overrides plugin rows. Installed with `dsh plugin --profile <name> add <pkg>`. Three ship in-box: base, web-app, headless.
184 entries
Out-of-process SDK
The protocol stack for driving a Harness runtime from another process: wire protocol, TypeScript client, stdio JSON-RPC server. Callers supply the runtime executable and its cordis.yml.
175 entries
Model Providers
The provider-neutral LLM seam. Adapters register on ctx.llm; only two ship in-box (llm-deepseek, llm-pi-ai). Swapping a model means writing one adapter — the loop and compaction never change.
99 entries
MCP Servers
MCP client bridge: connects external Model Context Protocol servers and registers their tools on ctx.tools under server-qualified names. One plugin row per server; stdio and streamable-http transports; HMR hot-swaps.
85 entries
Workspaces
Persistent workspaces: user directories with titles and ordered session membership, built over the domain storage facility.
66 entries
Browser UI Plugins
The browser half of the dsh web GUI: shell boot, browser-host communication, shared UI services and 30+ ui-* feature plugins composed through the slot system.
53 entries
Typert Type Graph
Generates runtime artifacts from source types so host and browser share one typed RPC contract: generator, registry, loader, protocol.
46 entries
Skills
Discovers reusable agent instructions and exposes them through a provider-neutral catalog and loader. The filesystem provider scans five ranked roots for SKILL.md or flat Markdown; remote and embedded providers are explicitly allowed.
36 entries
Web Search & Fetch
Provider-neutral web search and fetch. Four providers ship in-box (Exa, Perplexity, native DeepSeek search, HTTP fetch); tool-web owns the stable model-facing names.
29 entries
A profile is what a user boots — an ordered set of bundles with install commands you can copy
Delegate to sub-agents instead of one long thread.
3 bundles
Get off the terminal.
2 bundles
A DSH profile tuned for building and reviewing web UI.
3 bundles
Search, read, and keep what you found.
3 bundles
Same work, fewer tokens.
2 bundles
The head of the ecosystem — with its install verdict attached
deepseek-ai
DeepSeek Harness: Everything is a Plugin.
nexu-io
🎨 The open-source Claude Design alternative. 🖥️ Local-first desktop app. 🖼️ Your coding agent becomes the design engine: prototypes, landing pages, dashboards, slides, images & video — real files, HTML/PDF/PPTX/MP4 export. 🤖 Claude Code / Codex / Cursor / DeepSeek Harness / OpenCode / Hermes & 20+ CLIs via BYOK.
CherryHQ
AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs
titanwings
将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? It's giving rebirth era. Welcome to Digital Life 1.0. 🫶
tt-a1i
Agent skill for beautiful, verifiable architecture, workflow, sequence, data-flow, and lifecycle diagrams—self-contained HTML with motion and crisp export.
USArmyResearchLab
Dshell is a network forensic analysis framework.
anywhere-labs
为 DeepSeek Harness (DSH) 生态打造的现代化桌面端体验
Devin-AXIS
A next-generation, source-available AI workspace with a self-evolving agent runtime for editable code, design, presentations, websites, and video—a Codex alternative that integrates DeepSeek Harness for subagent delegation, combining iPolloWork’s complete AI workbench with DSH’s specialized agents and both plugin ecosystems in one workflow.
68 entries execute code on your machine at install time, outside the agent sandbox
deepseek-ai
DeepSeek Harness: Everything is a Plugin.
nexu-io
🎨 The open-source Claude Design alternative. 🖥️ Local-first desktop app. 🖼️ Your coding agent becomes the design engine: prototypes, landing pages, dashboards, slides, images & video — real files, HTML/PDF/PPTX/MP4 export. 🤖 Claude Code / Codex / Cursor / DeepSeek Harness / OpenCode / Hermes & 20+ CLIs via BYOK.
CherryHQ
AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs
anywhere-labs
为 DeepSeek Harness (DSH) 生态打造的现代化桌面端体验
Recently indexed repositories
deepseek-ai
DeepSeek Harness: Everything is a Plugin.
nexu-io
🎨 The open-source Claude Design alternative. 🖥️ Local-first desktop app. 🖼️ Your coding agent becomes the design engine: prototypes, landing pages, dashboards, slides, images & video — real files, HTML/PDF/PPTX/MP4 export. 🤖 Claude Code / Codex / Cursor / DeepSeek Harness / OpenCode / Hermes & 20+ CLIs via BYOK.
CherryHQ
AI productivity studio with smart chat, autonomous agents, and 300+ assistants. Unified access to frontier LLMs
titanwings
将冰冷的离别化为温暖的 Skill，欢迎加入数字生命1.0！Transforming cold farewells into warm skills? It's giving rebirth era. Welcome to Digital Life 1.0. 🫶
tt-a1i
Agent skill for beautiful, verifiable architecture, workflow, sequence, data-flow, and lifecycle diagrams—self-contained HTML with motion and crisp export.
USArmyResearchLab
Dshell is a network forensic analysis framework.
anywhere-labs
为 DeepSeek Harness (DSH) 生态打造的现代化桌面端体验
Devin-AXIS
A next-generation, source-available AI workspace with a self-evolving agent runtime for editable code, design, presentations, websites, and video—a Codex alternative that integrates DeepSeek Harness for subagent delegation, combining iPolloWork’s complete AI workbench with DSH’s specialized agents and both plugin ecosystems in one workflow.
DSH Market indexes every repository in the DeepSeek Harness ecosystem and labels each one with what it actually is — bundle, skill, agent preset, MCP server or plain Cordis plugin — and whether it can be installed at all.
Every repository carrying the dsh-plugin topic, filterable by install verdict, language, licence and install method.
Skills are directories with a SKILL.md — a different install mechanism from bundles, listed separately.
The official mcp-client bridge takes a one-line config, so existing MCP servers run under DSH unchanged.
Browse MCP servers →Profiles bundle a scenario into one bootable set, with the install order, conflicts and prerequisites spelled out.
Browse profiles →What the verdicts mean, and how to check them yourself.
A directory of DeepSeek Harness plugins that leads with one question nobody else answers: can this thing actually be installed? Every entry is checked against the official publish spec and labelled with the result, instead of being listed and left for you to find out the hard way.
No. It is an independent index built from public GitHub data and the official DeepSeek Harness documentation. Nothing here is endorsed by DeepSeek, and no listing is paid.
Five different things get called "plugin" here: bundles (an npm package plus cordis.patch.yml, installed with dsh plugin add), skills (a directory holding SKILL.md), agent presets (agent.cordis.yml), MCP servers (one line of config), and plain Cordis plugins. They install in completely different ways, so every entry is labelled with which one it actually is.
The repository declares no dsh.bundle in its package.json. In the official docs' own words, dsh plugin add warns and activates no layer — the install looks like it worked and nothing happens.
Every verdict comes from reading the repository's package.json and file list, then applying the rules in the official publish spec. No language model is asked to judge anything, and each entry stores the exact fields its verdict was based on.
It installs from git, it needs a build step, and it declares no prepare script. The package manager clones it, never builds it, and the loader then looks for output that was never produced.
preinstall, install, postinstall and prepare run on your machine at install time — outside the agent sandbox, before any permission prompt exists. 68 entries here do that. It is not an accusation, it is something to know before you type the command.
No, and the wording is deliberate. It means no package.json could be found, so no rule could be applied. 6,962 entries are in that state and they are excluded from every percentage on this page.
The 284 entries actually checked, not the 7,210 indexed. Anything unchecked is reported separately rather than quietly counted as passing.
Entries are collected from the GitHub API and re-checked in batches. Each detail page shows when that specific entry was last checked, so you can judge staleness per entry instead of trusting one site-wide claim.
Every verdict on this site stores the fields it was derived from. Disagree with one and you can re-check it against the spec yourself in about a minute.
