Built because the context never followed.

I work across several AI tools every day. Each one is capable and each one starts from nothing. The same architecture, the same conventions, the same decisions we already settled and the same things I never want suggested again — typed out, in full, in every new chat and in every new tool.

I tried what everyone tries. A markdown file I pasted in. A page I linked. A custom-instructions block I kept re-editing. They all break the same way: nothing keeps them current, nothing scopes them, and nothing records where a claim in them came from. Meanwhile the knowledge I generated ended up owned by whichever vendor’s memory feature it happened to land in.

So SuperThinking is deliberately not another assistant. It is the layer underneath: you connect a source once, it keeps structured context current, and any tool that speaks the Model Context Protocol can read the slice you permitted — with provenance attached and the whole thing exportable and deletable by you. The full definition is in the docs.

What runs today: source connections, a read-only MCP endpoint any compatible client can query, hand-written context items with recorded consent, and export. What does not: OAuth, writing over MCP, semantic retrieval, per-tool permission slices, and self-serve deletion of everything a source produced.

The capability ledger is the itemised version, and the roadmap is the order. There are no user counts, uptime figures or certifications on this site, because none of those would be true yet.

Independent and self-funded. SuperThinking is one slice of a broader AI product practice; the other two are below, and they are different problems entirely.

If the problem is one you have, or you have connected a source and something is wrong, write to me directly at hugo@superthinking.ai.