MCP implementation

[radumarias]

Evolving this experiment.

Have a minimalistic MCP server that handles these

• manage a folder where code for charts will be created
• prepares a basic structure based on charts lib used
• exposes this to LLM with instructions of what lib used and programming lang
• starts a http server to serve the code in that folder
• llm will directly change the files in there
• then will call a tool from MCP to refresh the server to pick-up new code and refresh the viewer page

Implementation

Create MCP server

Use any coding agent to create a basic MCP server with typescript

Tools in MCP server

• get_supported_charts_libs: returns unique libs info array derived from names of charts libs which we support. for now just use https://www.amcharts.com/
  example: [ { "id": "amcharts", "name": "amcharts", "instructions": "how to use the lib and any useful info" } ]
• initialize_wk(chart_lib_id): will create the workspace folder and code template based on chart lib
• get_wk_path: path of workspace folder
• refresh: will reload the code on the server. Maybe restart the server or whatever is needed for the give framework.
  [todo]: think of a mechanism to refresh the viewer page, maybe with websockets

Tests

• use any AI agent like Claude Code or any other agent
• add postgres MPC to get data
• add our charts MCP, config with code folder and http port
• in AI agent ask for some data from postgres MCP
• then ask to draw it with charts MCP
• ask several changes
• make sure all are rendered ok