Give your agent the docs it was never trained on.
Your agent already reads your codebase — CandleKeep gives it the rest. Upload PDFs, EPUBs, and markdown; your agent browses the table of contents, reads the exact pages, and cites them. No vector search, no hallucinated summaries.
Anthropic abandoned vector-RAG for agentic search in Claude Code — agents navigate books better than they retrieve chunks.
How it works
Add a book
Drag in a file or run ck items add
We structure it
Auto-paginated with an extracted table of contents
Your agent reads
It opens the exact pages and cites them by page
What you get
Upload anything
PDFs, EPUBs, and markdown — including scanned docs via OCR — structured into pages your agent can actually read.
Read by page
Your agent browses the shelf, checks the table of contents, and reads exact page ranges — the workflow a human researcher uses, not a similarity search.
Shelves
Group books into focused collections so your librarian knows exactly where to look for each domain.
Enrichment
Titles, authors, and tables of contents auto-filled, so every read returns a clean, citable source.
The difference a real book makes
Asked to audit a codebase for security holes, a generic agent flags what it remembers from training data and misses anything niche, recent, or specific to your stack.
The same agent, with a curated security book in its library, reads the relevant chapters and returns findings cited by page. In CandleKeep's own experiment, a curated 238-rule security book found 8× more critical vulnerabilities — same code, same model, different knowledge.