原生 DSH 插件，把 DeepSeek Harness 变成对话式数据分析师：连上数据库，用自然语言提问，得到基于 SQL 的结论与 HTML 报告。支持 9+ 种数据库引擎，带只读保护。通过 `dsh plugin --profile web add @yejiming/dsh-data-agent` 安装。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-08-23
许可证
MIT
01它能帮你完成什么？
Analyze data by conversation: connect a database and ask business questions in plain language
DSH views the schema, writes and runs SQL against real data, iterates, and returns clear data conclusions and business insights; analysis reports save as offline HTML
Analysts and engineers who run DeepSeek Harness and want to query MySQL/PostgreSQL/SQLite/Oracle/ClickHouse/SQL Server etc. by talking to the agent
02如何接入 DeepSeek Harness？
前置条件
- DeepSeek Harness with the `dsh plugin` command
- Database client installed for your engine (mysql/psql/sqlcmd etc.) and network access to the target DB
安装步骤
- 01
Run `dsh plugin --profile web add @yejiming/dsh-data-agent` (or from GitHub: `dsh plugin --profile web add github:omdsh-dev/dsh-data-agent`)
$ dsh plugin --profile web add @yejiming/dsh-data-agent
- 02
Start the web UI with `dsh --profile web`, create a session, pick the '数据模式' (Data Mode) preset, then connect a database in the workbench
$ dsh --profile web
验证接入成功
作者未说明
回滚
- Uninstall with `dsh plugin --profile web remove @yejiming/dsh-data-agent` (also `remove` for desktop and dsh-tui profiles); clear the preset with `rm -rf "$DSH_HOME/.agent-presets/data-agent"` if needed.
03DSH 适配与能力边界
Native DSH plugin that connects databases and lets DeepSeek Harness analyze data by conversation (auto SQL, business insights, HTML reports, read-only protection).
Conversational data analysis
Natural-language business questions→DSH writes and executes SQL, iterates on real results, returns conclusions and business insights
When not in read-only mode the agent can run update/management statements at your requestAI data governance (catalog)
Database schema, columns, comments, relations→AI-proposed business meanings per table/column, confirmed by the user; later queries read them via catalog-search/catalog-get/metric-get
Catalog persists only desensitized source summaries, metadata, versions and human curated meanings — no passwords, credentials, query results or sample rowsHTML analysis reports (render-analysis)
Analysis results→Versioned offline HTML reports (metric/line/bar/pie/scatter/table) saved under analysis-reports/ in the working dir
Reports are atomically saved to the session working dir; the full dashboard is self-contained (inline data/styles/SVG) and opens offline18-tool data protocol
Same 8-tool protocol across web/desktop/dsh-tui/headless→sql-query, sql-write, sql-cmd, render-analysis, catalog-search, catalog-get, metric-get plus str_replace_editor for files
Host/community tools like describe_image and ssh_* do not enter Data Mode
04适合谁？何时不该用？
适合
- Analysts and engineers who run DeepSeek Harness and want to query MySQL/PostgreSQL/SQLite/Oracle/ClickHouse/SQL Server etc. by talking to the agent
不适合
- The plugin and its ecosystem adapters run in-process inside DSH — not an OS, process or realm sandbox. Ecosystem permission only negotiates admission; it does not replace DB account privileges, network isolation or host security policy.
- For production databases, use a read-only DB account and enable read-only mode in the connection form; otherwise the agent can run update/management statements.
- Each engine needs its own CLI client (mysql/psql/sqlcmd/clickhouse HTTP); SQL Server needs Microsoft ODBC sqlcmd 18.x and only supports SQL Login (no Windows/Entra/DSN/named instances).
05兼容性、维护与安全提示
- The plugin and its ecosystem adapters run in-process inside DSH — not an OS, process or realm sandbox. Ecosystem permission only negotiates admission; it does not replace DB account privileges, network isolation or host security policy.
- For production databases, use a read-only DB account and enable read-only mode in the connection form; otherwise the agent can run update/management statements.
- Each engine needs its own CLI client (mysql/psql/sqlcmd/clickhouse HTTP); SQL Server needs Microsoft ODBC sqlcmd 18.x and only supports SQL Login (no Windows/Entra/DSN/named instances).
MIT · 由 facts pack 抓取（README 分支 main）
06常见问题
怎么安装 dsh-data-agent？
执行 `dsh plugin --profile web add @yejiming/dsh-data-agent`（或 `github:omdsh-dev/dsh-data-agent`），然后 `dsh --profile web`。新建会话选择「数据模式」预设，并在工作台连接数据库。
支持哪些数据库？
MySQL、PostgreSQL、SQLite、Oracle、Hive、Impala、ClickHouse、Apache Doris 与 SQL Server。各引擎需各自的 CLI 客户端；SQL Server 需要 Microsoft ODBC sqlcmd 18.x，且仅支持 SQL Login。
我的数据库安全吗？
插件在 DSH 进程内运行（非沙箱）。生产环境建议使用只读数据库账号，并在连接表单开启只读模式；否则智能体会按你的要求执行更新语句。
分析报告长什么样？
render-analysis 工具会把带版本号的、自包含的 HTML 报告（metric/line/bar/pie/scatter/table）写入会话工作目录的 analysis-reports/，可离线打开并随会话日志持久化。
07相关的 DSH 工作流
weknora
作者 tencent
开源的 LLM 知识库平台：把原始文档转化为可检索的 RAG 知识库、自主推理智能体和自动维护的 Wiki 系统，支持多租户部署。
honcho
作者 plastic-labs
用于构建有状态 AI 智能体的记忆库，支持跨会话连续学习与上下文工程。
mirage
作者 strukto-ai
全球首个面向 AI 智能体的统一虚拟文件系统。
reme
作者 agentscope-ai
ReMe：面向 AI Agent 的记忆管理套件——帮 Agent「记住我、优化我」，支持 RAG 与长期记忆。
08数据与来源
dsh-data-agent是DeepSeek Harness（DSH）的数据分析插件。
dsh plugin --profile web add @yejiming/dsh-data-agent
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-23。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。