One library. Every agent reads from it.
One library, every runtime. CandleKeep is the neutral knowledge substrate your agents carry across Claude Code, Cursor, Codex, and Claude Desktop — without locking your company's brain to any single model vendor.
Books over RAG: five sub-agents reading five different books in parallel cost less than a single RAG pipeline with re-ranking.
How it works
Install once
curl … | sh, then ck setup
Connect your agents
Plugins for Claude Code & Cowork, MCP for the rest
Read anywhere
The same library across every runtime
What you get
Claude Code & Cowork plugins
Install once with ck setup; a librarian and reader live inside your agent, auto-triggering when it plans.
ck CLI
The control plane for your library: list, read, upload, and manage books from any terminal.
MCP server
An OAuth-gated endpoint any MCP-compatible runtime can read from, scoped to your library and the marketplace.
Multi-agent research
A librarian scouts your library and the marketplace, routes page ranges to parallel readers, and synthesizes cited findings.
Embed via API/MCP
Use CandleKeep as the knowledge backend inside your own product; your users' agents read from your curated library without ever seeing CandleKeep.
Your company's knowledge shouldn't be captive to one vendor's model
Every model vendor is racing to own company context — but they build it captive: your knowledge makes their model smart and stays inside their product. Switch models, rebuild your brain.
CandleKeep is the neutral substrate between your knowledge and every runtime. One curated library — delivered via plugin, CLI, or MCP — to every agent your team runs, on whatever model you choose. The brain belongs to the company, not the vendor.