表格魔术

Installation

Visual Query Builder

_{Full reference on tabularis.dev →}

• Drag-and-Drop: Build queries visually with ReactFlow.
• Visual JOINs: Connect tables to create relationships.
• Advanced Logic: WHERE/HAVING filters, aggregates (COUNT, SUM, AVG), sorting, and limits.
• Real-time SQL: Instant code generation.

Configuration Storage

_{Full reference on tabularis.dev →}

Configuration is stored in ~/.config/tabularis/ (Linux), ~/Library/Application Support/tabularis/ (macOS), or %APPDATA%\tabularis\ (Windows): connection profiles, saved queries, app settings (config.json), custom themes, and per-connection editor preferences. You can move this folder from Settings > Storage (or with the TABULARIS_DATA_DIR environment variable), for example to an iCloud Drive or Dropbox folder to sync connections across machines; installed plugins always stay local. Tabs and queries are restored when you reopen a connection. The wiki covers the full file layout and every config.json option, including custom AI model overrides.

<details> <summary>How Follow System works on Linux</summary>

<br />

On Linux, Follow System reads the XDG desktop settings portal's org.freedesktop.appearance/color-scheme preference and follows its live updates, including GNOME's dark-mode toggle with the standard Adwaita GTK theme. The resolved light/dark theme is applied explicitly to GTK window decorations and the webview. A portal value of 0 (no preference) resolves to light, so switching back to the desktop default cannot reuse the app's previously forced dark theme. If the portal is unavailable or returns an unsupported value, Tabularis falls back to the native window theme, then the browser media query. The default window capability grants core:window:allow-set-theme to allow native theme changes. macOS and Windows use native theme notifications. See the portal specification.

Portal command reads reuse one cached D-Bus connection, while the live watcher keeps its own connection. Failed reads discard the cached connection so the next request can reconnect; failed connection attempts are not cached. The existing two-second command timeout also bounds concurrent requests waiting for the cache.

</details>

AI Features (Optional)

_{Full reference on tabularis.dev →}

Optional Text-to-SQL and query explanation powered by OpenAI, Anthropic, MiniMax, OpenRouter, Ollama (local models, no API key, full privacy), and any OpenAI-compatible API (Groq, Perplexity, Azure OpenAI, LocalAI, ...). Model lists are fetched from your provider and cached locally; custom models can be configured per provider.

Plugin System

_{Full reference on tabularis.dev →}

Tabularis is hackable with an external plugin system. Plugins are standalone executables that communicate with the app over JSON-RPC 2.0 via stdin/stdout, and can be written in any language.

• Install Plugins: Browse and install community drivers from Settings → Available Plugins, no restart required.
• Manage Drivers: View all registered drivers (built-in and plugins) in Settings → Installed Drivers and uninstall plugins with one click.
• Any Database: Add support for DuckDB, MongoDB, or any other database by writing or installing a plugin.
• Plugin Registry: Official plugins are listed in plugins/registry.json.
• Developer Guide: See plugins/PLUGIN_GUIDE.md to build your own driver in any language.
• Declarative themes (development): Settings → Appearance → Manage themes supports local packages, previews and personal/VS Code imports without executable plugin activation. See the theme author guide for the separate tabularis-theme CLI, packaging and release gates; public runtime/tooling rollout is not implied by this development feature.

MCP Server: AI Agent Integration

_{Full reference on tabularis.dev →}

Tabularis includes a built-in MCP (Model Context Protocol) server that lets AI agents read your database schema and execute queries directly from their chat interface.

tabularis --mcp

One-click setup for Claude Desktop, Cursor, and Windsurf: open Settings → MCP Server Integration, click Install Config next to your client, and restart it. Manual configuration is covered in the wiki.

Available tools

Once connected, your AI agent can:

Every tool accepts an optional output_format argument. JSON is the default, preserving the existing response format for all clients. You can choose JSON or TOON as the default under Settings → MCP Server Integration; a tool call's output_format argument overrides that preference. TOON is especially compact for tabular query results passed to an LLM. MCP transport remains JSON-RPC in either mode; only the text inside the tool result changes.

Example prompts

"Show me all tables in my production database and describe the orders table"

"Write and run a query to find the top 10 customers by total order value this month"

"Check if there are any missing indexes on the users table"