DSH Plugin Ecosystem - AI-powered plugin discovery platform for DeepSeek Harness

What is the DSH Plugin Ecosystem

The DSH Plugin Ecosystem is a centralized discovery and management platform for plugins built for DeepSeek Harness. It addresses a persistent pain point for agent developers and infrastructure teams: the fragmentation of plugin sourcing across disparate GitHub repositories. Before this ecosystem existed, developers had to manually track individual repositories, verify manifest validity by inspecting package.json files, and manage installation across multiple channels—a process prone to silent breakage and version drift.

The platform solves this by acting as a single, AI-ranked discovery layer. It indexes plugins from verified GitHub repositories tagged with the dsh-plugin topic, then applies keyword pre-filtering combined with AI relevance ranking based on user intent. This means that when an operator searches for a "memory plugin" or an "approval workflow," the results are ranked by semantic relevance to that intent rather than merely matching raw keywords.

As of the latest data snapshot (8/21/2026), the ecosystem lists 376 verified plugins, all synced from GitHub on a six-hour cadence. Each plugin's metadata is refreshed automatically, and listings are validated for manifest integrity before they appear. The provenance model is fully transparent: every plugin points back to its source GitHub repository, so operators can trace any listing to its canonical origin.

Importantly, the ecosystem adopts a safety-first listing posture. Being listed is a discovery signal, not an endorsement. The platform gives developers a structured way to find, compare, and evaluate plugins, but the responsibility for reviewing source code, permissions, and release artifacts remains with the operator.

Core Capabilities of the DSH Plugin Ecosystem

The ecosystem's functional capabilities go beyond simple directory listing. At its core, it implements a layered discovery pipeline that combines structured categorization with AI-driven relevance ranking.

AI-based relevance ranking is the platform's differentiating engine. Unlike static directory listings that rely on alphabetical or popularity ordering, the ecosystem uses AI to rank plugins based on user intent. When a developer issues a search query, a keyword pre-filtering stage narrows the candidate set, and then an AI relevance model re-orders results to surface plugins that best match the semantic meaning of the query. This is particularly valuable for a catalog as large as 376 plugins, where naive search would bury relevant results under name-matching noise.

Structured plugin categorization organizes the catalog into a functional taxonomy that mirrors real deployment concerns:

• Agent Frameworks — runtime infrastructure such as sandbaseai/sandbase-harness, a local-first AI agent runtime with sandboxed sessions, MCP tools, memory, credentials, audit/replay, and a built-in console. It supports OpenAI, Anthropic, MiniMax, DeepSeek V4, and OpenAI-compatible models on your own infrastructure.

• Agent Skills — reusable agent capabilities, exemplified by sandbaseai/sandbase-skills, which ships 88 open-source Agent Skills covering research, social intelligence, marketing, and business workflows. It is compatible with Codex, Claude Code, Cursor, Gemini CLI, and DeepSeek Harness.

• Memory & Audit — such as PerryLink/dsh-memento, which implements bounded, layered, approval-gated, auditable cross-session memory using ctx.memory with a SQLite provider, a memory tool, and frozen snapshot injection.

• Safety & Approval — like PerryLink/dsh-auto-review, a second-model AI auto-reviewer for approval requests. A read-only reviewer subagent returns structured allow/deny verdicts with reasons, fail-closed by default, and fully auditable.

• UI & Skins — visual customization such as d-dev0101/open-sea-skin, a WebGPU ocean skin distributed as a Chrome/Edge extension with a static installer and native integration.

• Utilities — operational tooling like lxzy-7/dsh-plugin-guard, an install safety net providing pre-install snapshots, one-click/automatic rollback, guarded boot, and incident reports that trigger agent analysis.

• Cross-Platform — channel integrations including tencent-connect/dsh-qqbot for QQ and tkwkeven/dsh-lark-all for Feishu/Lark, which provides a WebSocket bridge, parallel task sessions, and media integration.

Sync mechanism ensures the catalog stays current. GitHub metadata is refreshed every six hours, so new releases, updated descriptions, and version changes propagate automatically. However, installed plugins do not update themselves—operators must run an explicit plugin update command to pull changes from the source.

Who Uses the DSH Plugin Ecosystem

The ecosystem serves distinct operator profiles, each with concrete technical requirements that map to specific plugin categories.

Agent developers building custom AI assistants use sandbaseai/sandbase-harness as a local-first runtime foundation. The plugin provides sandboxed sessions, MCP (Model Context Protocol) tool integration, persistent memory, credential management, and an audit/replay facility—all running on their own infrastructure. For this audience, the "local-first" property is decisive: it removes the dependency on external inference services and keeps model selection open across OpenAI, Anthropic, MiniMax, DeepSeek V4, and any OpenAI-compatible endpoint.

Team workflow automation teams deploy sandbaseai/sandbase-skills to inject 88 ready-made Agent Skills into their pipeline. These cover research workflows (web search, document synthesis), social intelligence (profile analysis, engagement drafting), marketing content generation, and business operations. Because the skills are model-agnostic and compatible with Codex, Claude Code, Cursor, and Gemini CLI, teams can standardize on one skill library while running different agent runtimes in different departments.

Compliance-sensitive teams—such as those in finance, healthcare, or regulated sectors—adopt PerryLink/dsh-memento for cross-session memory that is bounded and audit-gated. The plugin's layered design, built on ctx.memory with a SQLite provider, means every memory write is approval-gated and recorded. The frozen snapshot injection mechanism ensures that memory state at a given point in time can be reconstructed for audit or rollback purposes.

Security-conscious operators use PerryLink/dsh-auto-review to add a second-model review layer ahead of every approval request. The reviewer subagent is read-only and returns structured allow/deny verdicts with explicit reasons. Its fail-closed default means that if the reviewer cannot reach a definitive verdict, the request is denied—eliminating silent approvals. The full audit trail makes this suitable for environments where every agent decision must be reconstructable.

Channel integration teams bridging agent outputs into messaging platforms rely on the Cross-Platform category. tkwkeven/dsh-lark-all provides a Feishu/Lark all-in-one channel plugin with a WebSocket bridge for real-time message flow, parallel task session isolation, and media integration. tencent-connect/dsh-qqbot similarly connects DeepSeek Harness to the QQ ecosystem.

Technical Architecture and Installation

The DSH plugin system is built on a well-defined technical foundation: installable plugins are npm-compatible packages that point to a Cordis configuration layer. Understanding this structure is essential for both consuming and authoring plugins.

Plugin package structure. At the repository root, every valid plugin must include a package.json with a dsh.bundle.patch declaration. This declaration references a cordis.patch.yml file, which describes how the plugin integrates into the Cordis configuration layer. The repository must also contain runnable code or compiled artifacts. Optional artifacts include a SKILL.md file for agent guidance, client bundles, and documentation.

Installation workflow. The CLI-based installation is straightforward. Prerequisites are Node.js and pnpm installed on the target machine. The install command is:

npx @deepseek-ai/dsh plugin --profile web add github:owner/repo

Replace owner/repo with the plugin's GitHub repository path. For production deployments, you should pin the source to a specific commit:

npx @deepseek-ai/dsh plugin --profile web add github:owner/repo#commitSHA

The #commitSHA suffix locks the installation to an immutable commit, preventing unexpected code changes from a future push from silently entering your environment on the next install.

Publication pipeline. To publish a plugin to the ecosystem, tag the repository with the dsh-plugin GitHub topic. Repositories are synced every six hours and verified for manifest validity before appearing in the listings. This means a new publication can take up to six hours to become visible.

Creator guidance. Plugin authors should follow the official DeepSeek Harness Plugin Creation Guide for the canonical structure, supported declarations, and best practices around the dsh.bundle.patch manifest.

DSH Plugin Ecosystem vs. Manual GitHub Tracking

The baseline alternative to the ecosystem is manual tracking: maintaining a spreadsheet of GitHub repositories, checking each one periodically for releases, and manually verifying package.json manifests. This section quantifies the difference.

Discovery efficiency. The ecosystem indexes 376 verified plugins in one searchable catalog. Manual tracking requires browsing GitHub search results across arbitrary repositories, triaging unrelated projects, and maintaining a personal registry of candidates. With AI ranking, the platform surfaces relevant plugins by intent—a search for "approval workflow" returns semantically matched results rather than name-only matches. Manual keyword browsing on GitHub cannot replicate this ranking quality.

Verification. Every plugin in the ecosystem is validated for manifest validity on each six-hour sync. Manual tracking has no such gate: a repository can push a breaking change to its package.json or cordis.patch.yml at any time, and the failure is only discovered when installation breaks. The ecosystem's automated validation catches manifest issues before they reach your environment.

Update cadence. The platform refreshes GitHub metadata every six hours. Tracking 376 repositories manually at that cadence is impractical—even a weekly check across 20 repositories is a significant recurring time investment. Automation removes this toil entirely.

Safety posture. The ecosystem is explicit that listings are discovery signals requiring operator review. Manual tracking has no guidance layer at all—there is no taxonomy, no verification signal, and no structured way to compare options.

The trade-offs are real, and an operator evaluating the platform should understand them clearly.

Frequently Asked Questions