Dashboard Agent MCP
Enables agents to create interactive dashboards from natural language queries, returning both text results and portable dashboard resources for MCP Apps-compatible hosts.
README
Dashboard Agent MCP
An MCP Apps server for the Dashboard Graph Agent. Other agents call one MCP tool and receive both:
- a concise text result for the model; and
- a portable interactive dashboard rendered from a
ui://MCP App resource.
The consuming agent does not need the dashboard renderer. A compatible MCP host loads the renderer from this server and displays it in a sandboxed iframe.
Architecture
MCP host -> create_dashboard -> Dashboard Agent LangGraph API
<- structuredContent + ui://dashboard-agent/dashboard-app.html
-> embedded React/Recharts dashboard
The MCP server does not query Graph or DuckDB directly. Planning, joins, graph traversal, DuckDB fallback, localization, and source lineage remain owned by the existing Dashboard Agent backend.
The browser UI is packaged into one HTML resource at build time. MCP Apps-compatible hosts retrieve that resource from this server, so callers do not install React, Recharts, or any dashboard components.
Requirements
- Node.js 20 or newer
- A running Dashboard Agent LangGraph server, unless fixture mode is enabled
Setup
npm install
Copy-Item .env.example .env
npm run build
npm start
The MCP endpoint is http://127.0.0.1:3001/mcp and the health endpoint is http://127.0.0.1:3001/health.
When the MCP client runs in Docker Desktop, add host.docker.internal to
MCP_ALLOWED_HOSTS and connect the client to the host-published port.
For a local server without the Dashboard Agent backend:
$env:DASHBOARD_AGENT_FIXTURE = "true"
npm start
Development
npm run dev
This runs the MCP server and the standalone dashboard view. The UI dev server uses fixture data so chart layout can be reviewed without an MCP host.
Open the standalone preview at http://127.0.0.1:5173/dashboard-app.html.
Codex project
This repository includes AGENTS.md and .codex/config.toml and can be opened directly as a Codex project at D:\git\dashboard-agent-mcp. The local Codex configuration also registers that path as trusted.
MCP client configuration
Streamable HTTP
Point a compatible MCP client at:
http://127.0.0.1:3001/mcp
stdio
{
"mcpServers": {
"dashboard-agent": {
"command": "node",
"args": ["D:/git/dashboard-agent-mcp/dist/server/server/main.js", "--stdio"],
"env": {
"DASHBOARD_AGENT_URL": "http://127.0.0.1:2024",
"DASHBOARD_AGENT_ASSISTANT_ID": "dashboard_agent"
}
}
}
}
Tool
create_dashboard
Input:
{
"question": "Show users by institute split by learning status",
"locale": "en",
"timezone": "Asia/Bangkok",
"maxCharts": 6
}
The tool is read-only and has no external side effects. It returns a DashboardDocument in structuredContent. The MCP App receives the same document and renders the charts.
The result also contains a server-rendered PNG preview of the primary chart. This core MCP image fallback displays in clients that support tools but not the MCP Apps UI extension, without requiring a client-specific chart renderer. HTTP deployments also publish a temporary image URL and return a ready-to-copy Markdown image line. This lets text-only models place the generated graph in their final response instead of needing to understand binary MCP image content. The complete text result is already formatted as a clean final answer so the calling model can return it verbatim without exposing handoff instructions.
Set MCP_PUBLIC_BASE_URL to the URL the user's browser can reach. With a local
SSH tunnel on port 3001, use http://127.0.0.1:3001. Preview retention is
configurable through DASHBOARD_PREVIEW_TTL_MS and
DASHBOARD_PREVIEW_MAX_ENTRIES.
Dashboard command
Type /dashboard followed by the analytical question as a normal chat message.
The MCP tool description privately tells the model to route that message to
create_dashboard; no MCP prompt is published, so clients do not expand hidden
instructions into the visible composer. Because MCP has no portable hidden slash
command mechanism, /dashboard is intentionally not added to client prompt
autocomplete.
Compatibility
Interactive rendering requires an MCP Apps-compatible host. A host that only implements core MCP still receives the text summary and structured dashboard data, but it cannot display the embedded iframe without MCP Apps support.
Commands
npm run dev Run MCP server and UI dev server
npm run dev:server Run only the MCP server in watch mode
npm run dev:ui Run only the standalone UI
npm run build Type-check and build server + single-file MCP App
npm test Run unit tests
npm start Start the production MCP server
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。