Kallimachos
☰

Subway · phone · claude.ai

example

What did we decide on auth in that March architecture discussion?

Signed session cookies (HttpOnly · SameSite). JWT was rejected because it cannot be revoked after issue.

  • wiki/decisions/auth.md
  • raw/conversations/sessions/2026-03-14-arch.md

A Claude that knows your knowledge — on any device.

Your Obsidian notes and Claude Code conversations become one searchable graph. Every answer carries its sources.

Self-host for free →AGPL-3.0 · one Docker line
Start on cloud →$15/mo billed yearly
376
documents
8,179
entities
12,355
relations
100MB
graph

Measured on the author's vault · 2026-08-29 · 1,129 notes (47MB) → 100MB of index and graph in 80 files

Claude Code memory — conversations, not just notes

Every answer carries its sources

Each entity and relation response includes the documents (docs) and external references (refs) behind it. When it is wrong, you can point at where.

src/kal_mcp.py

Korean actually searches

multilingual-e5-small embeddings layered with a 2·3-gram full-text index. Embeddings alone collapse on Korean; this does not.

src/kal_search.py

Claude Code sessions join the graph

Session logs are distilled into the same graph as your notes. The answer to "what did we decide last time" is usually inside a conversation.

src/distill_sessions.py

Not search. Questions.

Obsidian search matches strings. This answers "why did we decide it that way back then" — because your notes and conversations are woven into entities and relations.

How it works →

Your vault becomes a galaxy

Drawn from synthetic data — the real picture comes from your vault. An Obsidian plugin renders it inside your notes.

Self-hosted RAG Obsidian users pay nothing. Cloud changes where the graph lives.

cloud · persistence

A graph that outlives a restart

  • A graph you built once stays beyond one machine
  • The same answer from any device
  • Change laptops without re-extracting
  1. laptop ─┐
  2. phone ─┼─▶ graph ─▶ same answer
  3. claude.ai ┘

cloud · index

The index stays awake while your machine sleeps

  • An always-on remote MCP endpoint
  • Straight from claude.ai, Claude Code, ChatGPT and Gemini
  • stdio MCP in Claude Code stays free
  1. 03:14 laptop sleeping
  2. 03:14 mcp listening
  3. 03:15 phone → query → answer + sources

free · self-host

Everything, in one Docker line

  • Unlimited vault
  • Web UI and galaxy view included
  • AGPL-3.0 — all of the code is open
  1. $ docker compose up -d
  2. api running
  3. web running → http://localhost:5173

architecture

Extraction runs on your machine. The server never calls an LLM.

  • Entity extraction runs locally on your own Claude subscription
  • Only what the index needs goes to cloud — your vault folder is not mirrored
  • So server cost is storage alone, and the price is flat
  1. Obsidian vault + Claude session logs
  2. │ extract (your machine · your subscription)
  3. ▼ graph 100MB
  4. │ push
  5. ▼ cloud storage
  6. │
  7. ▼ remote MCP

privacy — said up front

What goes to the cloud is not metadata. It is the knowledge itself.

Entity names, relations, the chunk text used for search. The server must read them to search them, so end-to-end encryption is impossible. Better to say it now than have you find out later.

  • If privacy is why you use Obsidian, use the free self-host. Same features.
  • The server never runs an LLM over your notes. Extraction happens on your machine, on your subscription.
  • The server code will be AGPL too — before the first external user. Not "we can't read it", but "here is exactly what we do".
Read more →

Common questions

What is the difference between free and cloud?

Same features. What differs is where the graph lives. Self-host keeps it on that machine; cloud keeps it beyond one machine, so claude.ai and your phone get the same answers.

Why $19 a month?

It is where comparable open-source + hosting tools sit (Plausible · Umami · n8n · Ghost). Billed yearly it is $15 a month.

Do my notes go to the server?

On cloud, yes — entity names, relations and the chunk text used for search. The server has to read them to search them, so end-to-end encryption is impossible. If privacy is the point, self-host. Same features.

Does the server run an LLM over my notes?

No. Entity extraction runs on your machine on your own Claude subscription; the server only stores and searches the result.

Our legal team blocks AGPL

The same code under a different licence. Email us — the price is set when you ask.

Does Korean search work?

Yes. multilingual-e5-small embeddings layered with a 2·3-gram full-text index, measured on a Korean vault.

A Claude that knows your knowledge — on any device.

Start free. Move to cloud when you want it on your phone.

Self-host for free →Start on cloud →