Data Science MCP
chat + plots + stats
Adds tools that let the model run shell commands, read/write arbitrary files, spawn background processes, and make outbound HTTP requests. A single prompt-injection is enough to cause destructive actions. Only enable inside a sandbox you trust (VM, container, throwaway user account).
exec()
Disables the sandbox around LLM-generated code in
generate_custom_plotly and
generate_custom_static_plot. When off (the default),
those tools block import, dunder attribute access,
and calls to eval/exec/open/etc.
Route each request through a supervisor that plans and delegates to specialist worker agents (one per tool category). No single model sees every tool. You'll see the plan and each worker's progress live in the chat.
chat + plots + stats