Vega-Lite

Implements the save_data MCP tool that accepts tabular data (CSV, JSON, or structured records) and persists it in a module-level dictionary keyed by user-provided names. The server maintains session-scoped data in memory without external database dependencies, enabling LLMs to store intermediate datasets during multi-step visualization workflows. Data is retrieved by name in subsequent tool calls, creating a stateful context bridge between conversational turns.

Implements the visualize_data MCP tool that accepts a Vega-Lite JSON specification template and a reference to a previously saved dataset by name, then merges the data into the spec's data.values field and returns the complete visualization specification. The tool performs JSON schema composition, allowing LLMs to define chart structure (axes, encodings, marks) separately from data, enabling reusable visualization templates and data-driven chart generation without requiring LLMs to construct full Vega-Lite specs from scratch.

Supports two output modes controlled by the --output_type command-line argument: PNG rendering (via Vega-Lite's built-in renderer) for visual output suitable for display in UI clients, and text mode for terminal/log-based environments. The server initializes with the chosen output type at startup and applies it uniformly to all visualize_data calls, enabling deployment flexibility across headless servers, desktop clients, and web interfaces without code changes.

Implements the MCP server specification using the mcp Python framework (v1.0.0+), communicating with MCP clients via stdio streams using JSON-RPC 2.0 message format. The server.py module registers handlers for list_tools and call_tool via @server decorators, which are invoked by the MCP client to discover available tools and execute them. This architecture enables seamless integration with Claude Desktop and other MCP-compatible clients without requiring HTTP servers or custom protocol implementation.

The list_tools handler advertises available tools (save_data and visualize_data) to MCP clients with full schema definitions including parameter names, types, descriptions, and required fields. This allows clients to present tool options to users and validate inputs before invocation. The schema definitions are embedded in the tool metadata returned by list_tools, enabling LLMs to understand tool capabilities and construct appropriate invocations without external documentation.

The main(output_type) async function in server.py initializes the MCP server and binds it to stdio streams for communication with the MCP client. It uses asyncio.run() to execute the async initialization, setting up the server's event loop and stream handlers. The entry point in __init__.py parses the --output_type command-line argument and invokes main(), creating a complete initialization pipeline from CLI invocation to active MCP server ready to receive tool calls.

The visualize_data tool accepts a Vega-Lite specification template (JSON object with chart structure, encodings, marks, etc.) and merges a previously saved dataset into the spec's data.values field. This composition approach allows the LLM to define chart structure separately from data, then bind them at visualization time. The tool performs shallow JSON merging, inserting the data array into the spec without modifying other fields, enabling template reuse across different datasets.

Implements a naming system where datasets saved via save_data are stored in a module-level dictionary keyed by user-provided names, and visualize_data retrieves them by name. This design allows LLMs to refer to datasets symbolically (e.g., 'sales_data', 'monthly_metrics') rather than passing large data objects between tool calls, reducing message size and improving readability of tool invocation sequences. The naming system is implicit and unvalidated — any string is accepted as a dataset name.