Opyt vs Claude on its own

Opyt runs inside Claude. The question is not which one, but whether Claude alone can see your sources.

Start with the concession, because it is a large one. You do need Claude, or Cursor, or another client. Opyt is an MCP server with no interface of its own, and it is worthless without one. For anything public and indexed, your assistant's built-in search already finds it and Opyt adds nothing at all.

The problem is that the sources Opyt is built on are not publicly fetchable. Fetching a post on x.com returns HTTP 402 Payment Required; X's robots.txt admits Google and almost nobody else; and an assistant's web fetch makes a plain request that does not log in or fill in forms. Your bookmarks, likes, follows and Substack subscriptions are private to your account by definition, so no crawler will ever reach them, however good it gets.

The second problem outlasts the first. Even where a source is reachable, an assistant that fetches it keeps nothing. The next session starts from zero, re-fetches and re-pays. Nothing accumulates, so there is nothing to count, no convergence between questions asked months apart, and no way to ask what you have actually been reading.

At a glance

Where Claude on its own is better

• Anything public and indexed. If the answer is on the open web, built-in search finds it and Opyt contributes nothing.
• Deep research on a topic you have never followed. Opyt returns nothing there by construction. The store holds only what you already read, and none of it is about a field you never followed. Deep research is included in the plan you already pay for.
• Projects and a long context window handle “here are twenty documents, answer this” with no install and no store.
• Agentic browsing can drive your logged-in session, so it can reach X. The limit is throughput rather than capability: it renders and reads pages one at a time, on request, and a bookmark backlog or a full author archive is not a browsing session.
• Nothing to install. Opyt needs Python 3.10+, a clone, a virtualenv and one API key before it answers anything.

Where Opyt is better

• It reads what no crawler can. The seed is 691 bookmarks, 99 liked authors, 492 follows and 19 subscriptions, all private to the account, and a plain fetch of an x.com post returns HTTP 402 (checked 2026-08-29).
• It keeps what it fetched: 2,830 documents at roughly 40 KB each, one SQLite file, 115.9 MB when exported whole (measured 2026-08-26). An assistant re-fetches the same page every session that needs it.
• Whole archives. Confirming one author pulls 3,412 posts, 38 repos and 96 essays, and the X half is free, since it reads the session already logged into your browser.
• Retrieval is one embedding call plus a local scan: about $0.000001 and 0.4 s to turn the question into a vector (measured 2026-08-30), then BM25 and cosine over 8,095 stored 4,096-dimension vectors, measured at 229 ms and 84 MB resident. An assistant answers by fetching and reading, once per source, every time it is asked.
• It is read-only by construction. Agentic browsing drives your live session with the ability to act. Opyt never posts, follows or likes.

Which one

Opyt is worthless without a client like Claude, so this was never a fork. The narrow question is whether Claude alone can reach your bookmarks, your subscriptions and a person's whole archive. It cannot, and what it does reach, it does not keep.

install Opyt free, MIT, five minutes