mcp-server-chart
- Type
- MCP server
- Scanned
- Nov 12, 2026
- Version
- e2a8fb7c185e
- Files read
- 102
This MCP server provides 27 tools for generating charts, maps, diagrams, and spreadsheets from structured data through AntV's cloud rendering service. The biggest risk is that every tool uploads all supplied data to an external Alipay/AntV server by default while declaring itself read-only, which lets an AI agent send confidential data off the machine without asking for confirmation, and its optional HTTP and SSE modes run without authentication or origin restrictions.
What it can reach
- Network
- Sends all chart data and configuration via HTTP POST to https://antv-studio.alipay.com/api/gpt-vis (or a custom VIS_REQUEST_SERVER URL), and optionally listens on a local HTTP port (default 1122) when run in SSE or streamable mode.
- Secrets
- Reads the optional SERVICE_ID environment variable (along with VIS_REQUEST_SERVER and DISABLED_TOOLS) and sends SERVICE_ID to the rendering server when generating map charts.
Findings
-
MediumChart tools upload all input data to a remote cloud server while claiming readOnlyHint
This package does not render charts locally; every tool call sends all chart values, labels, and table rows via HTTP POST to https://antv-studio.alipay.com/api/gpt-vis by default, returning a remotely hosted image URL. Although the README's environment variable section mentions this default service, none of the 27 tool descriptions or manifest.json warn that data leaves the computer, and every tool sets readOnlyHint: true. Because MCP clients often auto-approve readOnlyHint tools without asking the user, a prompt-injected agent or an agent charting sensitive local files can upload confidential data to the third-party service without confirmation.
Fix: Set readOnlyHint to false and openWorldHint to true on all chart tools so MCP clients ask for confirmation before sending data externally, and state clearly in tool descriptions and manifest.json that data is uploaded to the remote rendering service.
-
LowHTTP and SSE transports have no authentication and allow any website via wildcard CORS
When started with --transport streamable or --transport sse, the Express servers require no authentication token and perform no Origin or Host header checks. In streamable mode, cors({ origin: "*" }) lets any website the user visits in a browser send cross-origin POST requests to http://localhost:1122/mcp and read the responses, and docker-compose.yaml exposes port 1122 on 0.0.0.0 to the local network. A malicious webpage or network peer can invoke tools directly, abusing a private VIS_REQUEST_SERVER instance or the user's SERVICE_ID secret.
Fix: Remove wildcard CORS, validate Origin and Host headers to prevent cross-origin browser and DNS-rebinding requests, require authentication on HTTP/SSE endpoints, and bind Docker port mappings to 127.0.0.1 by default.
-
LowDISABLED_TOOLS only hides tools from listing and unstripped arguments can overwrite chart type
The DISABLED_TOOLS environment variable only filters the list returned by ListToolsRequestSchema; CallToolRequestSchema and callTool() never check getDisabledTools(), allowing a prompt-injected agent or direct MCP client to call disabled tools by name. In addition, callTool() validates input with compiledSchema.safeParse(args) but discards result.data and spreads raw args after type in generateChartUrl(), allowing extra properties in tool arguments to overwrite the chart type and be forwarded to VIS_REQUEST_SERVER.
Fix: Enforce getDisabledTools() inside callTool(), pass the parsed result.data object from Zod instead of raw args, and place type after ...options in generateChartUrl() so callers cannot overwrite it.
