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

English | 简体中文 | 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.

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. On legacy Web shells without the settings.section slot, the same UI falls back to a top-level Agent Plugins Market page entry. Skills appear in the / slash menu; MCP tools appear as mcp__<suite>__<server>__<tool>; slash commands and /agent-* subagents register automatically.

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

// ~/.dsh/profiles/<profile>/package.json
{
  "dependencies": { "dsh-agent-plugins-market": "^0.4.6" },
  "dsh": { "profile": { "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "dsh-agent-plugins-market"] } }
}

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. New shells render it as a localized Settings page; legacy shells without settings.section receive a guarded top-level page-mode fallback instead of a duplicate surface.
• 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 / subagents — commands/*.md register as dsh slash commands; agents/*.md register as agent-<name> skills.
  • Model context — skills flow through the host-native skill catalog and MCP tools register directly through dsh-mcp-client; suite inventory and source metadata remain available in the Web market page without adding a redundant model-facing inventory tool.
• Runtime discovery — installed plugins are discovered from configured source ids under ~/.dsh/agent-plugins/.sources/<sourceId>/ (user dimension) and <project>/.dsh/agent-plugins/.sources/<sourceId>/ (project dimension). Stale unmanaged user checkouts are ignored, while project checkouts remain state-authorized. Local sources read the working tree directly, including uncommitted changes.
• Native project layouts (zero-copy migration) — a repository's own .claude/ and .agents/ directories (skills, agents) are discovered in place as read-only project-native suites: no install, no copying, no state file. Repositories migrating from Claude Code or agent-plugins.org conventions work as-is; a project skill shadows an installed suite skill of the same name (project wins).

Supported plugin layouts

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

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 (project-dimension snapshots are cached for 5 seconds on the skill-list hot path).
• Claude Code hooks support the mapped bridge subset; LSP is counted and previewed but not executed.
• Native project layouts (.claude/, .agents/) inject skills and subagents; their commands/*.md are not registered as slash commands because the harness command registry is process-scoped, not session-cwd-scoped.

Development

pnpm install
pnpm run test        # vitest over fixture suites + multi-dialect parsing
pnpm run typecheck
pnpm run lint        # ESLint over the refactored source and tests
pnpm run format:check
pnpm run check:architecture
pnpm run check:refactor
pnpm run build       # tsc host + tsdown client + module-loader banner
pnpm pack

The internal modularization decision and staged migration are documented in docs/design/engineering-refactor-plan.md and docs/adr/0001-catalog-centered-modular-refactor.md.

The docs website lives in 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

• dsh-skills — centralize ~/.claude/skills and .skill packages into a global DSH skill library
• @claude2dsh/plugin — import Claude Code sessions & skills into DSH, sync sessions back
• dsh-crew — dispatch work from Claude Code / Codex to DSH agents
• deepseek-ai/deepseek-harness — the DeepSeek Harness itself (official)
• awesome-deepseek-harness-plugins — community DSH plugin directory

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