mcp-zettel

Give Claude (or any MCP client) persistent memory in the form of a Zettelkasten.

An MCP server that exposes a personal knowledge base of atomic, interconnected markdown notes to Claude Desktop, Claude Code, Cursor, or any MCP-compatible client. Plug it in and the assistant can create notes, cross-reference them with [[wiki-links]], search by keyword or tag, and walk backlinks — letting it build and use durable context across sessions without you copy-pasting anything.

---

Why

Every chat you start with an LLM begins with zero context about what you've already decided, written, or learned. The Zettelkasten method (small atomic notes + explicit links between them) is a strong fit for LLM-accessible memory: chunks are naturally small, links make relevance explicit, and the storage is plain markdown on your own disk.

This MCP server exposes that knowledge base to any LLM client via MCP tools, so the model can:

• Create a new note when you share a decision or insight worth keeping
• Search for notes on a topic before answering ("what did I decide about X?")
• Link notes bidirectionally to build up a graph ("this contradicts [[a3f2c9]]")
• Walk backlinks to find everything connected to a concept

You keep plain markdown files on disk. The model gets structured access to them.

How it looks in a client

After connecting the server, the LLM can do things like this (your client will show actual tool calls):

> What did I conclude about RAG chunk sizes?
[searches notes with query "rag chunk size"]
[reads 2 matching notes]
Based on your notes a3f2c9 ("RAG chunk sizing") and b7e412 ("Sentence-boundary
splitting"), you concluded: 800 chars with ~15% overlap, sentence-aligned.
You flagged that pure character chunking ([[2f00a1]]) hurt recall on your
arxiv set and moved away from it.

Exposed MCP tools

Tool	Purpose
create_note(title, body, tags)	Create a new atomic note. Use [[other_id]] in body to link.
read_note(note_id)	Fetch a single note.
update_note(note_id, title?, body?, tags?)	Update any fields, leaving others unchanged.
delete_note(note_id)	Remove a note permanently.
list_notes(tags?)	List every note; optional tag filter is intersection.
search_notes(query, tags?, limit?)	Keyword search — titles and tags weigh more than body.
search_notes_semantic(query, tags?, limit?)	v0.2. Embedding-based search for conceptual queries. Uses fastembed on-device (no API call).
link_notes(from_id, to_id, label?)	Append a [[to_id]] wiki-link to from_id's body.
get_backlinks(note_id)	Every note whose body references this one.
linked_notes(note_id)	The ids this note links to (outbound).
suggest_links(text, exclude_ids?, limit?)	v0.6. Given arbitrary text (e.g. what you're about to save as a new note), return the most-likely-to-belong-linked existing notes. Hybrid-fuses keyword + semantic rankings via RRF, so you don't have to pick which search to use.

Plus MCP resources:

• zettel://all — one-line index of every note
• zettel://{note_id} — full rendered note
• zettel://graph — v0.4. Mermaid diagram of every note + [[wiki-link]] in the vault, rendered inline by any markdown+mermaid client (Claude Desktop, Obsidian, mdBook…).
• zettel://graph/tag/{tag} — v0.4. Same diagram, but restricted to notes carrying {tag} plus their direct neighbors — useful once the full graph gets too noisy to read.

Two search tools, not one

Keyword search is what you want when you know the term. It's cheap, the ranking is predictable, and exact matches always beat similar-sounding ones. Semantic search wins when the query wording doesn't match the note wording — asking for "rate limiting" when the note calls it "throttling", or "why my cache is cold" when the note is about "TTL tuning". The LLM can call whichever makes sense; the tool descriptions tell it which is which.

The embedding model is BAAI/bge-small-en-v1.5 by default (384-dim, ~130 MB, CPU-only). Override with MCP_ZETTEL_EMBEDDING_MODEL. The index rebuilds lazily on the first semantic query after any write, so there's a short wait the first time — after that it stays in memory for the life of the server process.

Install

git clone https://github.com/dhruvpatel1706/mcp-zettel.git
cd mcp-zettel
pip install -e .

Python 3.10+.

Wire it up

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on your OS, and add:

{
  "mcpServers": {
    "zettel": {
      "command": "mcp-zettel-server"
    }
  }
}

Restart Claude Desktop. The zettel tools are now available to the model.

Claude Code

claude mcp add zettel -- mcp-zettel-server

Cursor / Continue / any stdio MCP client

Point the client at mcp-zettel-server as a command; the server speaks MCP over stdio.

Custom storage location

Set MCP_ZETTEL_ROOT to override the default ~/.mcp-zettel:

{
  "mcpServers": {
    "zettel": {
      "command": "mcp-zettel-server",
      "env": { "MCP_ZETTEL_ROOT": "/Users/you/vault" }
    }
  }
}

Use the CLI directly (no MCP client needed)

The same store is accessible via a plain CLI — useful for power users who want to inspect, edit, or seed the knowledge base outside of an LLM session.

mcp-zettel create "RAG chunk sizing" \
  --body "Settled on 800 chars, 15% overlap, sentence-aligned. See [[b7e412]]." \
  --tag rag --tag decisions

mcp-zettel list --tag rag
mcp-zettel search "sentence boundary"
mcp-zettel show a3f2c9
mcp-zettel backlinks a3f2c9

On-disk layout

~/.mcp-zettel/
└── notes/
    ├── a3f2c9.md          ← one markdown file per note
    ├── b7e412.md          ← YAML frontmatter: title, tags, created_at, updated_at
    └── ...                ← body is plain markdown; [[id]] is a wiki-link

Every note is a single file. That means: easy backups (git), easy grep, no lock-in. If you ever stop using this server, you still have a directory of markdown files.

Design choices

• Files, not a database. One note per markdown file means you can edit in any editor, back up with git, and inspect without tooling. The store is thin glue on top.
• Short hex IDs, not slugified titles. [[a3f2c9]] is stable — rename the title and all inbound links still resolve. Also shorter than a filename-based slug.
• Bidirectional links derived, not stored. Backlinks are computed on read by scanning every note's body for [[target_id]]. No separate index to keep consistent. Trivial at the scale this is designed for (≤ low thousands of notes).
• Title/tag weighted search. Title hits count 3×, tag hits 2×, body hits 1×. Matches the intuition that if a note's title mentions "retrieval" it's more about retrieval than a note that says the word once in the middle of the body.
• FastMCP, not low-level MCP. The MCP Python SDK's decorator-based FastMCP surface means tools are just Python functions with Pydantic-typed args — no manual JSON Schema authoring.

Development

pip install -e ".[dev]"
pytest
black --check src tests
isort --check-only --profile black src tests
flake8 src tests --max-line-length=100 --ignore=E501,W503,E203

CI runs on Python 3.10 / 3.11 / 3.12.

Inspect the server interactively with the MCP inspector:

npx @modelcontextprotocol/inspector mcp-zettel-server

Prompt templates (v0.3)

MCP clients that support prompt menus (Claude Desktop, Cursor) get four server-side templates that encode the "right way" to do common zettelkasten moves without you having to re-type the instructions:

Prompt	What it does
distill_conversation(conversation, max_notes?)	Take a chat transcript, extract discrete insights worth saving as atomic notes. The model proposes titles/bodies/tags; you approve, it calls create_note.
find_linkable_notes(concept, limit?)	Before writing a new note, surface existing ones that might want to link to/from it — via search_notes_semantic.
daily_note(prompt_date?)	Drop a daily journal template (worked-on / learned / blockers / notes-created-today).
summarize_by_tag(tag, style?)	Summarize everything under a tag. Style = bullets / essay / outline.

These are just string-returning functions registered with @mcp.prompt(). Keeping the wording server-side means the same "distill" prompt behaves consistently whether you invoke it from Claude Desktop, Claude Code, or Cursor.

Roadmap

• [x] v0.2 — embedding-backed semantic search alongside keyword search
• [x] v0.3 — @mcp.prompt() templates for common note operations
• [x] v0.4 — graph-view resource (zettel://graph) returning a mermaid diagram of links
• [ ] v0.5 — remote Streamable HTTP transport for multi-device access

License

MIT. See LICENSE.