Tabellum

dev.dudko由 dudko.dev 發布v0.1.2更新於 Oct 25, 2026

MCP bridge to the Tabellum database client: schema, reads, SQL drafts; writes approved in the app.

發布方自己的服務：PDF to Markdown

已驗證STDIO僅桌面資料庫開發者工具

服務詳細資訊

識別碼: dev.dudko/tabellum
傳輸: STDIO
執行環境: Desktop only
信任: Verified
發布於: Sep 26, 2026

安全提示

MCP 服務會接收對話中的工具呼叫請求，並可能執行外部操作。啟用前請先了解該服務能做什麼，並只授予它所需的權限。

註冊中心已驗證此服務的命名空間屬於 dudko.dev。它以軟體包形式執行，而不是託管端點，因此沒有可與該網域比對的端點。

SourceWeft 的檢查只涵蓋條目是否完整、端點是否回應；不審查伺服器的程式碼，也不審查其工具的行為。

在 README 中新增徽章

向讀者展示該 MCP 服務已收錄於 SourceWeft，徽章連結到本頁。

[![Listed on SourceWeft](https://sourceweft.com/badge/mcp/dev.dudko%2Ftabellum.svg)](https://sourceweft.com/mcp/dev.dudko%2Ftabellum?utm_source=readme&utm_medium=badge)

概覽

AI 產生的概覽

將 AI 助理連接到 Tabellum 桌面資料庫用戶端，可檢視結構、執行唯讀查詢、起草 SQL，寫入操作需在應用程式內確認。

可以做什麼: 檢視連線、結構、物件、使用者與角色，以及按連線開啟分享後的樣本資料列
執行唯讀查詢並檢視執行計畫，開啟查詢分頁並寫入 SQL 草稿
提出寫入、DDL 與管理語句，等待你在應用程式中點擊核准
將結構與另一連線或已儲存的快照比對，並產生補齊差異的遷移語句
列出、讀取並提出儀表板與結構設計
適用情境: 適合已在 Tabellum 桌面用戶端管理 PostgreSQL、MySQL/MariaDB、SQL Server 或 BigQuery 的使用者，希望助理協助探索結構、起草查詢與遷移建議，同時由自己掌控所有會變更資料的操作。
執行需求: 需在 macOS、Windows 或 Linux 上安裝並執行 Tabellum 桌面應用程式；npm 套件需 Node.js 24 以上，透過 npx 啟動。轉送程式會從使用者家目錄下的設定檔讀取應用程式的本機連接埠與每次安裝產生的權杖，無需手動設定金鑰。

安裝前請注意

助理可讀取結構資訊、使用者與角色，以及在按連線開啟資料列分享時讀取樣本資料列。寫入、DDL 與管理語句僅為提案，需在應用程式中核准，10 分鐘後逾時失效。呼叫會在應用程式的代理日誌中留下紀錄；重新產生權杖後需重新啟動 MCP 伺服器。

安裝

在 SourceWeft 中

開啟儀表板中的 Tabellum，將其新增到工作區。
為需要使用其工具的對話啟用該服務。

Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照儲存庫中的啟動說明。

README

Tabellum — MCP setup

Let your AI assistant work with your databases through Tabellum, a desktop database client for PostgreSQL, MySQL/MariaDB, SQL Server and BigQuery. The MCP server is hosted by the running app on your machine; this package is a small stdio shim that forwards to it. Queries the assistant writes appear live in the editor, and nothing that changes data runs without your click in the app.

Server: tabellum
Transport: stdio (local process)
Package: @dudko.dev/tabellum-mcp — Node.js 24+
Auth: a per-install token the app generates; the shim reads it for you from ~/.tabellum/mcp.json.

What you can do

With the Tabellum app open, the assistant gets the same tools as the app's own agent — the app serves them, so the list follows the app version. Reads run at once; anything that writes waits for your click in the app.

Tool	What it does
Look around
`list_connections`	Connection names, drivers, read-only flags. No hosts, no users.
`describe_schema`	Every schema of a connection with its table, view, function, procedure and sequence names.
`describe_object`	Columns, indexes, constraints and the `CREATE` of one object.
`list_users`	Users and roles with their flags, where the database lists them.
`sample_rows`	Up to N rows of a table — only where the connection shares rows.
Query
`run_read`	Runs a statement only if the engine classifies it as a read; anything else answers `needs_approval`.
`explain`	The estimated plan of a read; nothing runs.
`propose_write`	A write, DDL or admin statement for your approval in the app (10-minute timeout).
`open_editor_tab`	Open a query tab bound to a connection.
`draft_to_editor`	Type SQL into a tab while you watch, without running it.
Compare schemas
`compare_schemas`	Diff a connection against another or a saved snapshot; with `migration: true`, the statements that close the gap. Nothing runs.
`save_schema_snapshot` / `list_schema_snapshots`	Save a connection's schema to the workspace; list the snapshots.
Dashboards
`list_dashboards` / `get_dashboard`	The workspace's dashboards; one dashboard's file and panels.
`propose_dashboard`	A new or changed dashboard — you see the diff and apply it.
Schema designs
`list_designs` / `get_design`	The workspace's schema designs; one design's file and objects.
`propose_design`	A new or changed design — you see what is added, changed and removed. Nothing runs against a database.
`describe_feature`	The file format and rules of `designer`, `dashboards` or `schema_compare`, read before proposing one.

Resources, for clients that prefer them: tabellum://connections, tabellum://schema/<connection>, tabellum://snapshots, tabellum://dashboards, tabellum://designs. Every call is attributed in the app — a badge on the tab and a line in the agent log.

Install in Claude Code (plugin)

/plugin marketplace add dudko-dev/tabellum-mcp-setup/plugin install tabellum@tabellum/reload-plugins

/plugin marketplace add dudko-dev/tabellum-mcp-setup/plugin install tabellum@tabellum/reload-plugins

Start the Tabellum app first (macOS, Windows or Linux). The shim reads the app's local port and token from ~/.tabellum/mcp.json, written while the app runs; with the app closed every call answers "Tabellum is not running".

Prefer not to use the marketplace? Add the server directly:
claude mcp add tabellum -- npx -y @dudko.dev/tabellum-mcp
claude mcp add tabellum -- npx -y @dudko.dev/tabellum-mcp

Connect from other clients

Claude Desktop

claude_desktop_config.json (Settings → Developer → Edit Config):

json

{  "mcpServers": {    "tabellum": {      "command": "npx",      "args": [        "-y",        "@dudko.dev/tabellum-mcp"      ]    }  }}

{  "mcpServers": {    "tabellum": {      "command": "npx",      "args": [        "-y",        "@dudko.dev/tabellum-mcp"      ]    }  }}

Cursor

~/.cursor/mcp.json or .cursor/mcp.json:

json

{  "mcpServers": {    "tabellum": {      "command": "npx",      "args": [        "-y",        "@dudko.dev/tabellum-mcp"      ]    }  }}

{  "mcpServers": {    "tabellum": {      "command": "npx",      "args": [        "-y",        "@dudko.dev/tabellum-mcp"      ]    }  }}

VS Code (GitHub Copilot / MCP) and other clients

json

{  "servers": {    "tabellum": {      "type": "stdio",      "command": "npx",      "args": [        "-y",        "@dudko.dev/tabellum-mcp"      ]    }  }}

{  "servers": {    "tabellum": {      "type": "stdio",      "command": "npx",      "args": [        "-y",        "@dudko.dev/tabellum-mcp"      ]    }  }}

Safety

The engine classifies, you approve. Every statement is classified by the app's parser; anything that is not a read goes through propose_write and waits for your click. A label claimed by the model never overrides it.
Loopback only. The app listens on 127.0.0.1 with a per-install bearer token (Settings → Agents shows it and can regenerate it).
Row data is opt-in per connection. sample_rows returns rows only where you enabled sharing for that connection.

Troubleshooting

"Tabellum is not running" — open the app, then retry.
Unauthorized after regenerating the token — restart the MCP server in your client so the shim re-reads ~/.tabellum/mcp.json.
A write seems stuck — it is waiting for approval in the app; after 10 minutes it returns declined: timeout.