Not sure if anyone here has attempted to use/ modify Plotly cloud’s embedded agent MCP toolset. This is a voyage of discovery for me, is what I’m after possible? Demo is running on mtt-tracker.plotly.cloud, and the /_mcp stream is available. But agent(s) only able to see one board/ dataset at a time.
I was able to get this working with multiple dataframes on import! I need to take a step back and think more about the function handle approach.
I got this working with multiple dataframes on import too. I think I need to step back and rethink the function handle approach a bit.
Thanks for calling this out! What you described is correct for the moment since the embedded chat feature is still in Beta. We’re actively working on allowing the embedded chat to access your custom MCP endpoints within the app. Right now it will only have access to the data, but it will have MCP access in the coming weeks!
I work at aicost.tools, so my first thought here is what each question to the embedded agent will cost once it can call MCP tools. With several dataframes plus a list of tool definitions in context, every chat turn sends a lot more tokens than the question itself, and that adds up quickly on a public demo. @Robert_Claus, when MCP access lands, will app authors be able to limit which tools or tables the agent sees per board, or see token usage per session?
Yes, this is needed. I have moved out of Plotly and using custom MCP defs + guardrails (and recently jev helpers) to limit this. Seems like a general MCP problem but needs attention from Plotly here.