Shared across every LLM client
Install the Claude skill on your laptop, point Claude Desktop at
the MCP server, or paste a remote connector URL into the Claude
mobile app. Same knowledge, same bearer token, same search results.
Changing LLM runtime doesn't orphan your context.
Hybrid semantic + keyword search
Every query fuses dense vector similarity (pgvector HNSW) with
BM25 keyword ranking (ParadeDB) in a single SQL call. ~70 ms
end-to-end on a modest KB. Search fans out across every KB you
have access to, with per-result citations back to the source.
Git-push ingest
Every KB is a bare git repo behind SSH. git clone,
edit locally, git push — the server re-indexes
changed files incrementally. Bulk-import an Obsidian vault, a
recipe folder, or a research archive in seconds.
Share with your people
Per-KB membership with owner / editor / reader / search roles.
One family recipe book, one household travel plan, one team
research archive — all readable by the LLMs of every
member, each on whichever client they prefer.
Programmable by design
HTTP API is the canonical contract; the skill, MCP server, CLI,
and admin UI are thin clients over it. Mint, rotate, and revoke
tokens programmatically. Every capability a human can do in the
admin UI is an endpoint.
Self-hosted and yours
Runs on your own infrastructure — a Proxmox LXC, a VM, a
laptop. Embeddings go through heivol-modelproxy so no LLM
vendor ever sees your raw data. No training on your content.
Ever.