DeepSeek Harness 插件

dsh-agent-plugins-market

把任意 Claude Code / Codex / Cursor / Kimi / agent-plugins.org 的 git 市场仓库加为源,套件的技能、MCP 服务器、hooks 与斜杠命令在运行时注入 DSH 会话,并自带 Web 市场页。

跳到安装方式

来源信息

GitHub 仓库
Sivan757/dsh-agent-plugins-market
最近更新
2026年8月20日
分类
插件市场与管理
GitHub stars
0

安装

默认先复制一段 Prompt,让 Agent 读页面和仓库;需要自己装时再切到命令。

复制这段 Prompt,发给 DSH、Codex 或其他 Agent,让它先读页面和仓库。

请先不要安装。阅读这个 DeepSeek Harness 插件,说明它解决什么问题、会访问哪些文件、网络或密钥,以及如何安装和卸载。

插件页面:https://deepseekplugins.org/zh/plugins/Sivan757/dsh-agent-plugins-market
GitHub:https://github.com/Sivan757/dsh-agent-plugins-market
插件名:dsh-agent-plugins-market
作者:Sivan757
安装命令:dsh plugin --profile web add github:Sivan757/dsh-agent-plugins-market

确认前不要执行安装命令。

检查来源文件

安装前先看这个插件目录里的 README 和其他文件。

文件资源管理器3 个文件
README.md来源说明 · 只读预览

dsh-agent-plugins-market — Claude Code plugin marketplaces for DeepSeek Harness

English | [简体中文](README.zh.md) | Docs website

> Bring the Claude Code / Codex / Cursor plugin-marketplace ecosystem into DeepSeek Harness (DSH): install and inject agent plugins — skills, MCP servers, hooks and slash commands — from git marketplace repos, with a market page right inside the Web GUI.

dsh-agent-plugins-market is the standard way to run Claude Code, Codex, Cursor and Kimi plugin marketplaces inside DeepSeek Harness — zero conversion, zero file copying.

!npm !npm downloads !License !GitHub stars

![Agent Plugins Market screenshot](docs/screenshot.png)

![Suite detail (skills / MCP / commands preview)](docs/screenshot-detail.png)

Why?

DeepSeek Harness is a powerful agent harness — but its plugin ecosystem doesn't yet have the breadth of the Claude Code plugin marketplace world. There are hundreds of ready-made plugin marketplaces on GitHub (Claude Code .claude-plugin/marketplace.json, Codex .codex-plugin, Cursor, Kimi, agent-plugins.org v1.0.0 portable suites) full of skills, MCP servers, hooks and slash commands.

dsh-agent-plugins-market is the bridge: add any git marketplace repo as a source, install its suites, and their skills / MCP servers / hooks / commands are injected into your DSH sessions at runtime — no manual file copying, no conversion needed. Claude Code-authored skills work verbatim (${CLAUDE_PLUGIN_ROOT} is substituted automatically).

Quick start

> Requires DeepSeek Harness ≥ 0.1.0-rc.6 with a Web profile.

# install into a dsh profile (npm registry, recommended)
dsh plugin --profile <name> add dsh-agent-plugins-market

# or via pnpm inside the profile
pnpm add dsh-agent-plugins-market

Restart dsh, open Settings → Agent Plugins Market, add a marketplace repo as a source (e.g. https://github.com/anthropics/claude-plugins-official), and install suites with one click. Skills appear in the / slash menu; MCP tools appear as mcp__<suite>__<server>__<tool>; slash commands and /agent-* subagents register automatically.

<details> <summary><strong>More install options</strong></summary>

GitHub:

pnpm add github:Sivan757/dsh-agent-plugins-market
# or
dsh plugin --profile <name> add github:Sivan757/dsh-agent-plugins-market

Manual — add the package to the profile's dsh.profile.bundles (the package's cordis.patch.yml inserts the plugin row):

// ~/.dsh/profiles/<profile>/package.json
{
  "dependencies": { "dsh-agent-plugins-market": "^0.4.4" },
  "dsh": { "profile": { "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "dsh-agent-plugins-market"] } }
}

The built entry point (lib/, client/) is committed to the repository, so GitHub installs work without a prepare step.

</details>

Features

  • Plugin / suite management — configure git repository sources (markets), browse every discoverable plugin, install / uninstall / enable / disable / refresh per source or per plugin. Source ids are derived automatically from the repository manifest JSON — no manual input.
  • Market page in the Web GUI — source pills + search/actions, status tabs, a two-column card grid, and a plugin detail modal with previews for skills / MCP / hooks / commands / LSP.
  • Runtime injection

- Skills — a ctx.skills SkillProvider (project rank 250 / user rank 450); ${CLAUDE_PLUGIN_ROOT} is substituted so Claude Code-authored skills work verbatim, and appear in the / slash menu. - MCP servers — every valid mcp.json server of an enabled plugin mounts a live dsh-mcp-client child; tools appear as mcp__<plugin>__<server>__<tool>. - Hooks — a plugin's hooks/hooks.json mounts the dsh-hooks-claude-code bridge on the harness interception points (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStart, SubagentStop). - Commands / subagentscommands/*.md register as dsh slash commands; agents/*.md register as agent-<name> skills. - Context — the enabled-plugin catalog (user + project sections) is injected at session start; the agent_plugins tool queries it.

  • Runtime discovery — installed plugins are discovered from ~/.dsh/agent-plugins/.sources/<sourceId>/ (user dimension) and <project>/.dsh/agent-plugins/.sources/<sourceId>/ (project dimension). Local sources read the working tree directly, including uncommitted changes.

Supported plugin layouts

LayoutManifestNotes
agent-plugins.org v1plugin.jsonvendored 1.0.0 JSON Schema validation + spec §4 path rules
Claude Code market.claude-plugin/marketplace.json + .claude-plugin/plugin.jsonmarketplace plugins[].source relative paths
Universal.plugin/plugin.jsonmulti-client repos (e.g. vercel-plugin)
Cursor.cursor-plugin/plugin.jsondeclared skills paths
Kimi.kimi-plugin/plugin.jsoninline mcpServers
Codex.codex-plugin/plugin.json
Skill collection (manifest-less)none (synthetic)flat SKILL.md directory collections

One repo may carry several dialects at once (vercel/vercel-plugin ships all of them); the suite identity comes from the highest-precedence manifest while surfaces are scanned from the directories. mcp.json is validated strictly against the agent-plugins.org schema; .mcp.json is parsed leniently — top-level server-map shorthand, type: http / local / omitted type (stdio by command) normalization, ${CLAUDE_PLUGIN_ROOT} / ${CLAUDE_PLUGIN_DATA} / ${NAME:-default} placeholders, and unknown transports are tolerated per server. The marketplace manifest is authoritative for the suite set; manifest-less marketplace entries that carry skills and manifest-bearing container dirs not listed there are supplemented. Remote URL entries show as "remote reference" cards (metadata + source URL, not directly installable; add the repo as a source to install).

Configure marketplace sources

Sources persist in ~/.dsh/agent-plugins/state.json; cordis config seeds them (and re-adds missing ids on every boot):

- id: dsh-agent-plugins-market
  config:
    sources:
      - { id: agent-plugins, url: 'https://github.com/Sivan757/agent-plugins.git' }
      - { id: mattpocock-skills, url: 'https://github.com/mattpocock/skills.git' }
      - { id: claude-plugins-official, url: 'https://github.com/anthropics/claude-plugins-official' }
      - { id: ui-ux-pro-max, url: 'https://github.com/nextlevelbuilder/ui-ux-pro-max-skill.git' }
      - { id: my-local-plugin, url: '/Users/me/work/my-plugin', local: true }

A local: true source reads the directory in place (live working tree; never deleted on removal).

Comparison with other DSH ↔ Claude Code bridges

Capabilitydsh-agent-plugins-marketdsh-skills@claude2dsh/plugin@deepseek-ai/dsh-hooks-claude-code
Sourceany git marketplace repo (.claude-plugin, .codex-plugin, .cursor-plugin, .kimi-plugin, agent-plugins.org v1, manifest-less skills)~/.claude/skills dirs, project dirs, .skill packagesClaude Code sessions + skillsa Claude Code hooks.json config
Skills injection✅ + / slash menu✅ global skill library
MCP servers✅ live dsh-mcp-client mounts
Hooks✅ via dsh-hooks-claude-code bridge✅ (direct)
Slash commands / subagentscommands/*.md, agents/*.md
Market UI✅ full market page in the Web GUI✅ settings page
DirectionCC / Codex / Cursor ecosystem → DSHCC skills → DSHCC ↔ DSH session syncconfig → DSH

Want the opposite direction (dispatch work from Claude Code / Codex to DSH agents)? See dsh-crew.

FAQ

How do I install Claude Code plugins in DeepSeek Harness (DSH)?

Install this plugin in a dsh profile, then add any Claude Code marketplace repo as a source:

dsh plugin --profile <name> add dsh-agent-plugins-market

Then open Settings → Agent Plugins Market in the Web GUI, add the marketplace repo URL, and install suites with one click. Skills, MCP servers, hooks and slash commands are injected into your dsh sessions at runtime — no conversion, no file copying.

Does DeepSeek Harness support .claude-plugin/marketplace.json?

Yes — through this plugin. It reads .claude-plugin/marketplace.json + per-plugin .claude-plugin/plugin.json natively, plus .codex-plugin, .cursor-plugin, .kimi-plugin, .plugin (universal) and agent-plugins.org v1.0.0 plugin.json manifests (see the layout table above).

Can it inject MCP servers from a plugin suite?

Yes. Every valid mcp.json server of an enabled suite mounts a live dsh-mcp-client child, so MCP tools are callable by the DSH agent. mcp.json is validated strictly; .mcp.json is parsed leniently with placeholder support (${CLAUDE_PLUGIN_ROOT}, ${CLAUDE_PLUGIN_DATA}, ${NAME:-default}).

What about Claude Code hooks?

A suite's hooks/hooks.json is mounted through the official @deepseek-ai/dsh-hooks-claude-code bridge onto the harness interception points (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStart, SubagentStop). Only the mapped command-hook subset runs — see the bridge's README for the exact mapping.

Do Claude Code-authored skills work verbatim?

Yes. ${CLAUDE_PLUGIN_ROOT} is substituted automatically at runtime, so Claude Code ecosystem skills run unchanged and appear in the / slash menu.

dsh-agent-plugins-market vs manually copying skill files?

Manual copying breaks ${CLAUDE_PLUGIN_ROOT} paths, skips MCP servers/hooks/commands, and gives you no update path. This plugin installs whole suites from git sources with per-plugin enable/disable/refresh, and injects every surface (skills, MCP, hooks, commands, subagents) automatically.

Is it safe to install third-party suites?

The plugin never executes third-party code at install time: git sources clone through execFile (no shell), and third-party failures (broken manifests, invalid skills, escaping paths, unknown MCP transports, mount failures) are contained as per-plugin diagnostics. As with any third-party code, review a suite before enabling it.

Is it free and open source?

Yes — MIT licensed, published on npm, source on GitHub.

Requirements

  • ctx.skills (dsh-skill) is required.
  • Optional peers: @deepseek-ai/dsh-mcp-client (MCP injection), @deepseek-ai/dsh-hooks-claude-code (hooks bridge); missing capabilities degrade gracefully.
  • Web GUI ≥ 0.1.0-rc.6.

Security model

  • Git sources clone through git via execFile (no shell), depth 1, --ff-only pulls, 120s timeouts; local sources are read in place and never deleted.
  • Mutating HTTP routes accept same-origin POSTs only; bodies capped at 64 KiB.
  • Portable paths must start with ./ and resolve inside the plugin root (symlink escapes rejected); ${PLUGIN_ROOT} / ${PLUGIN_DATA} expand.
  • Third-party failures are always contained: broken manifests, invalid skills, escaping paths, unknown MCP transports, and mount failures are per-plugin diagnostics.
  • An error boundary wraps the whole market section and the detail modal: any preview render failure degrades to a notice instead of crashing the UI.

Known limitations

  • Project-dimension MCP servers are not mounted (dsh has no per-session tool scope); the project dimension covers skills and context.
  • Skill discovery has no file watcher: catalog changes apply after a manager action or a host restart.
  • Claude Code hooks support the mapped bridge subset; LSP is counted and previewed but not executed.

Development

pnpm install
pnpm run test        # vitest over fixture suites + multi-dialect parsing
pnpm run typecheck
pnpm run build       # tsc host + tsdown client + module-loader banner
pnpm pack

The docs website lives in [docs-site/](docs-site/) (Astro, deployed to GitHub Pages).

Vendored assets

The schemas/1.0.0/ JSON Schemas are vendored from agentplugins/agent-plugins-spec (spec 1.0.0 working draft); the spec forbids fetching schemas at load time.

Related projects

---

If this plugin helps you, please give it a ⭐ on GitHub — it helps more people find the Claude Code ecosystem inside DeepSeek Harness.