Docsy — Answers With Sources
A document workspace where Claude answers only from the files in the chat, every claim opens the exact passage it quotes, and questions are metered by plan through Stripe.
Chapters, description and comments are on YouTube.Watch on YouTube instead
One request, end to end.
Select a stage to read what it decides.
Read on the way in
A PDF, a Word file, text or Markdown is read the moment it is uploaded — .docx through mammoth — and a file that cannot be read is refused with no row written, so the library never holds a document the model cannot see.
Every document belongs to a workspace
A workspace is a Better Auth organization, and documents, chats and usage are all scoped by it — so the question of whose file this is gets answered once, not per query.
The documents are the context
The files attached to a chat go to Claude whole, with citations switched on and prompt caching on the document blocks, and the system prompt rules out anything the documents do not say.
Sources come from the API, not from memory
Citations arrive as structured events while the answer streams, and each one is turned into a marker that opens the exact quoted passage — with its page for a PDF, and the surrounding text for everything else.
A question is charged before it is answered
Each plan allows a number of questions a month, counted in a ledger kept apart from the chat history — so disconnecting mid-answer or deleting a chat gives nothing back — and the plan itself is always re-read from Stripe.
Three things you can point at afterwards.
Answers grounded in the reader's own files
Claude working only from the documents attached to the chat, streamed to the browser as it writes.
Citations anyone can check
Anthropic's Citations API carried through to the interface, so every claim opens the passage it rests on.
Usage limits that cannot be gamed
Monthly question limits per plan, a ledger that survives deleted chats, and entitlements granted only by Stripe's webhook.
Where to go from here.
Closest to this one by what it teaches, then by what it is built on.
The longest build on the channel: an AI app builder with workspaces, agent runs, subscription billing, and the auth boundary holding it all together.
A course marketplace in the shape of Udemy: instructors author courses that go live through an admin review, learners buy through Stripe and earn certificates anyone can verify, and every sale records the instructor's share.
A browser video editor for product videos: record or upload your app, let an AI director plan the camera, cuts and text, refine it on a multi-track timeline, then export an MP4 or share a link.