Background

Qiskit and the Runtime changes more frequently than LLMs are trained...this issue is a proposal to fetch relevant, up-to-date documentation across the Qiskit stack -- from Qiskit SDK, Runtime, Addons, Functions, and beyond

Description

Qiskit users are increasingly turning to LLMs, but there's a giant mismatch in utility between the latest Frontier models capabilities and user needs.

Frontier models are able to understand use-cases and stitch common use-cases to typical libraries (pySCF, CPLEX, networkx etc.), but are unable to generate up-to-date Qiskit code. This MCP aims to support LLMs with context from the latest up-to-date Qiskit documentation, in several fashions:

1. Fetching high-level documentation sections to understand latest tools available, and their use-cases (think: "what, why")
2. Fetching tool-specific documentation + tutorials, to ensure the latest functional implementation (think: "how")
3. And static analysis that is able to check code generated is accurate and up-to-date against Qiskit docs / tutorials / API-refs, before returning to users

In an ideal world, the frontier models are doing the larger form contextualising, and leveraging the MCP for all things Qiskit for correctness in the flow

Definition of done

Users wants to run a maximally independent set problem, with some networkx graph

• Leverage an external model (e.g. Claude Code) with this MCP installed
• It should be able to suggest qiskit-addon-opt-mapper or the two Qiskit Functions for optimization, and generate code to transform graph to whatever format is needed
• It should be able to use the two QAOA tutorials to suggest hardware efficient implementations
• All code generated is leveraging the latest Qiskit without errors