A Model Context Protocol server that lets investors trade from Claude or ChatGPT.
StratX is Motilal Oswal’s in-house algorithmic F&O trading platform: a marketplace of exchange-approved back-tested strategies with no-code and Python builders on top.
Reaching it means opening the app. The Model Context Protocol server exposes StratX to any MCP-capable client, so an investor can build a strategy, place it and square it off from Claude or ChatGPT, inside the assistant they already have open.
I was techno-functional lead. I ran benchmarking, tool and schema design, and the legal, audit and compliance scoping. The front end is React, the back end Python, on AWS.
An API written for developers assumes a caller who has read the documentation and holds the context in their head. A tool exposed over MCP has neither. The model sees a name, a description and a parameter list, and decides from those alone whether the tool fits what the user asked for.
That changes what good design means.
Tool names and descriptions are the interface, not documentation about it, so they have to say precisely what the tool does and what it does not do. Tools are drawn narrowly, because one general endpoint with a mode flag gives the model a decision it will sometimes get wrong. Errors return what the caller should do next rather than a status code, since the model reads the error and retries. Nothing is left implicit in session state, because the model cannot be relied on to remember what it set three turns ago.
Benchmarking was how we found the cases where the model picked the wrong tool, and most of the fixes were to names and descriptions rather than to code.
An investor placing a real trade from a third-party assistant raises questions a normal feature does not. Which system is the record of the instruction, what gets logged and where, and what the firm can attest to afterwards if the client disputes it.
I scoped that with legal and audit before the tools were finalised, because the answers constrained what the tools were allowed to do.