MCP Zettel

MCP开发工具

提供方 dhruvpatel1706

开发工具 未标注 无需鉴权 社区 / 第三方

An MCP server that provides Claude and other MCP clients with persistent memory through a Zettelkasten knowledge base of interconnected markdown notes. It enables LLMs to create, search, link, and reference atomic notes across sessions without requiring manual copy-pasting.

一句话选型结论

按资料字段整理

MCP Zettel 属于开发工具类 MCP 服务器,传输方式未标注,无需鉴权,没有现成安装命令,需按官方文档接入。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类开发工具 传输方式未标注 鉴权无需鉴权 部署方式本地部署
GitHub 星标 未标注 仓库热度,无公开仓库时显示未标注
工具数 未标注 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

开发工具
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量未标注暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

安装与配置

未标注 无需鉴权
上游未给出传输方式
  1. 1
    准备运行环境

    stdio 服务在本地起进程,需要先装好 Node.js 等运行时,再按安装命令拉取依赖。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后随便调一个只读工具,确认能返回结果,再接入正式流程。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
代码与仓库操作

在对话里直接查代码、提 issue、走 PR,省掉来回切工具。

本地命令与脚本

把构建、测试、格式化这类重复命令交给模型执行。

接口联调与测试

让模型调你的接口、跑一遍用例,快速定位问题。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 资料标注无需鉴权,装好就能连。

先想清楚的地方

  • stdio 方式要求本地装好运行时(如 Node.js),换机器要重新装一遍。
  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 资料没标注工具数量,实际能力要装好后确认。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在开发工具场景里让模型直接调用外部能力的人。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 完全不想在本地装环境的人。
  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
MCP Zettel当前 未标注 未标注 本地部署 无 2026-05-17
Git 838e2ec1 12 stdio · 本地进程 本地部署 有 2026-05-17
Markitdown MCP 2c653a5f 1 stdio · 本地进程 本地部署 有 2026-05-17
JavaGuide 未标注 stdio · 本地进程 本地部署 有 2026-04-11
awesome-claude-skills 未标注 stdio · 本地进程 本地部署 有 2026-04-26
Context7 831acf40 未标注 stdio · 本地进程 本地部署 有 2026-05-17

服务器档案

开发工具
字段来自项目仓库与公开索引,中文说明由本站标注
提供方dhruvpatel1706上游标注的维护方 dhruvpatel1706
分类开发工具本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式无需鉴权远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · 未标注本地安装时用到的包标识与运行时
部署类型本地部署上游给出的部署形态
最近核验2026-05-17上游最后核验时间

详细介绍

8.8 KB
来自项目仓库与上游目录的原文,未做删改

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```bash
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:```json
{
"mcpServers": {
"zettel": {
"command": "mcp-zettel-server"
}
}
}


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

### Claude Code```bash
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:```json
{
"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.```bash
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```bash
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:bash 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.

常见问题

7 问
答案由本站依据条目字段整理,不引用上游原文
MCP Zettel 怎么安装?

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

MCP Zettel 怎么在客户端里配置?

它的传输方式未标注,在客户端 MCP 配置里按官方文档填写启动命令和参数,保存后重启客户端生效。

MCP Zettel 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

MCP Zettel 需要 API Key 吗?

资料标注为无需认证,装好就能连;如果实际调用时报鉴权错误,按官方文档补上凭据。

MCP Zettel 能提供哪些能力?

资料显示能力标签包括 开发工具、命令行工具、自动化。具体工具清单以官方文档为准。

MCP Zettel 需要什么运行环境?

属于本地进程,需要在本机装好运行环境。

MCP Zettel 是本地还是远程?

它是本地进程(stdio),由客户端在本机拉起。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 2026-05-17。