# 10 Free GitHub Repos That Turn Claude + Obsidian Into a Real Second Brain
**作者**: kai
**日期**: 2026-08-07T21:21:03.000Z
**来源**: [https://x.com/0xkkai/status/2085838657068347401](https://x.com/0xkkai/status/2085838657068347401)
---
In the past four months, GitHub has quietly filled up with repositories that solve the same problem from different angles: how to make Claude actually remember what you know, using Obsidian as the storage layer.
Most of them are wrappers. Some of them are architectures. Two of them are worth building your workflow around today. This is the honest breakdown - what each one actually does, what it doesn't, and which one to install first depending on who you are.
> Before the list, three architectures you need to understand. Otherwise every repo will look the same, and you'll waste a weekend installing four things that do the same thing.
## The Three Architectures Behind Every Repo On This List
1. The Karpathy LLM Wiki Pattern
In April 2026, Andrej Karpathy published an 800-word gist on GitHub called LLM-WIKI. It described a three-folder architecture: raw/ for source documents, wiki/ for compiled linked pages Claude actually reads from, and instructions/ for the rules that keep the wiki from drifting into chaos.
The move that made it viral: Claude reads any given raw file once in its entire life, extracts what matters, writes it into the wiki, and never touches the raw file again.
> Why this matters: Every subsequent query pulls from the compiled wiki, which cuts token spend 70-90% on repeat lookups and gives you a knowledge base that actually compounds instead of decays.
Repos in this family: #1, #4, #5, #10.
## 2. The Skills Architecture
Anthropic quietly shipped a folder system in July where Claude auto-loads specialized instructions based on what you ask. You drop a folder into your skills directory containing a SKILL.md that describes when to trigger and what to do. Claude reads only the description at startup - the full skill loads only when the trigger fires.
For an Obsidian vault, this means you can define workflows like "when I mention a paper, activate paper-reader skill" or "when I open the vault, load my note-organization convention" without polluting your main context window.
Repos in this family: #3, #9.
## 3. The MCP Bridge
Model Context Protocol is the standard Anthropic released for connecting Claude to external tools and data. An MCP server acts as a live bridge - Claude sends structured requests, the server queries your vault, returns results in real time.
MCP repos come in two flavors: read-only (safe, Claude can only pull data) and read-write (powerful, Claude can create/edit notes).
> Rule of thumb: Read-only is what most researchers should start with. Read-write is what enables the "vault builds itself while you sleep" workflow.
Repos in this family: #2, #6, #7, #8.
## The 10 Repos, Ranked and Analyzed
## 1. AgriciDaniel/claude-obsidian
Architecture: Karpathy Wiki Pattern (purest implementation)
What it does: You drop any source document - pdf, markdown note, web clip, transcript - into a raw/ folder. Claude reads it once, compiles it into 8-15 linked wiki pages, and files it into the graph. Next time you query, Claude speaks only from the wiki with sources already cited.
Setup difficulty: Medium. You install as a Claude Code plugin, point it at an empty vault, then start dropping files. Initial vault takes 20-30 minutes to configure the way you want. After that, ingest is a single command.
Best for: Anyone new to the Karpathy pattern who wants the cleanest reference implementation. This is the closest thing to a canonical setup and is what most creators building content around this pattern are running.
Actual limitations: Opinionated folder structure - if you already have a vault organized differently, migration is painful. No built-in semantic search - retrieval relies on Claude opening the wiki and following links. Not designed for team use.
> Verdict: Start here if you're building your first knowledge vault. Skip if you already have 500+ notes in a different structure.
## 2. eugeniughelbur/obsidian-second-brain
Architecture: MCP (read-write) + Skills + custom commands
What it does: Provides persistent memory for Claude Code and six other CLI agents. Ships with 45 named commands including hybrid semantic search (BM25 + embeddings), self-rewriting notes that update themselves when you contradict them, key-less web research (uses public search), and scheduled agents that maintain the vault while you sleep.
Setup difficulty: High initially, low forever after. First-time setup requires installing local embedding models and configuring the scheduler. Once running, it's the most hands-off system on this list.
Best for: Power users with a vault of 1,000+ notes who want to stop babysitting it.
> The killer feature: Scheduled agents. You can tell Claude to lint your wiki nightly, deduplicate similar notes weekly, or resurface stale content monthly, with no intervention
Actual limitations: Complexity tax. Reading the docs takes 90 minutes. Local embeddings require ~4GB RAM. Some commands assume Claude Code specifically and won't work with Claude Desktop or the web app.
> Verdict: The most powerful setup on this list. Reach for it once your vault has outgrown your ability to maintain it manually.
## 3. kepano/obsidian-skills
Architecture: Skills (Anthropic-native pattern)
What it does: Ships pre-built agent skills authored by Kepano (the creator of Obsidian). Works with Claude Code, Codex, and Open Code. Each skill is a folder with a SKILL.md describing when to trigger and what to do - no manual context management needed.
Setup difficulty: Low. Clone the repo into your Claude skills folder, done. Skills auto-register on next Claude startup.
Best for: Anyone who wants "official-ish" defaults. Because Kepano authored these, the conventions align with how Obsidian actually thinks about knowledge management - folders, tags, backlinks. If you're already fluent in the Obsidian ecosystem, this feels native.
Actual limitations: Skills-only, no MCP. You can't do live queries - Claude reads the vault through the file system, so real-time vault state changes require refreshes. Also: fewer skills shipped than the ecosystem clones like #9.
> Verdict: The safest bet for anyone deep in the Obsidian mindset already. The trust factor here is high - you're running code from the person who built the thing.
## 4. qhuang20/obsidian-skills
Architecture: Skills + Karpathy Wiki Pattern
What it does: A Claude Code plugin that ships with a dedicated llm-wiki skill implementing the Karpathy pattern directly, plus several utility skills for note management. Trigger the wiki skill by mentioning a source; Claude does the ingest and linking silently.
Setup difficulty: Low-medium. Standard Claude Code plugin install. The wiki skill needs one initial config file describing your vault's naming conventions.
Best for: Users who want the Karpathy pattern without building it from scratch. Faster path to a working wiki than repo #1 if you're comfortable in Claude Code specifically.
Actual limitations: Less battle-tested than #1. The utility skills are convenient but overlap with #3 and #5, so you'll want to pick a lane.
> Verdict: Solid if you already live in Claude Code. Consider #1 for a cleaner reference and #5 for a purer Karpathy focus.
## 5. ekadetov/llm-wiki
Architecture: Karpathy Wiki Pattern (pure play)
What it does: Nothing but the Karpathy pattern, implemented as a Claude Code plugin. No extra bells. Ingest, query, save, lint - the six commands from the original gist, tightened up.
Setup difficulty: Very low. Install, point at vault, run /init. Working ingest in under 5 minutes.
Best for: Purists. If you want the pattern implemented exactly as Karpathy described it and nothing else muddying the surface area, this is it.
Actual limitations: No semantic search, no scheduled agents, no MCP bridge. Deliberately minimal, which is either a feature or a dealbreaker depending on your appetite.
> Verdict: Best minimum-viable Karpathy pattern implementation. Great starting point, may outgrow it.
## 6. iansinnott/obsidian-claude-code-mcp
Architecture: MCP (read-write via Obsidian plugin)
What it does: An Obsidian plugin that runs an MCP server inside your Obsidian instance. When Claude Code queries the server, it can read files, create new notes, modify existing ones, follow backlinks, and update the graph - all through Obsidian's own API, so nothing bypasses Obsidian's file safety.
Setup difficulty: Medium. Install as an Obsidian community plugin, enable the MCP server, add the endpoint to Claude Code config. Requires Obsidian to be running for the bridge to work.
Best for: Users who want Claude to interact with a live Obsidian instance - respecting plugins, templates, dataview queries, all of it.
> What makes this one different: It's the only repo on this list that goes through Obsidian's API rather than the raw file system.
Actual limitations: Requires Obsidian to be open. Bridge latency is higher than file-system MCPs (~200ms per query vs. ~20ms). No support for mobile Obsidian.
> Verdict: The right choice if your vault depends heavily on Obsidian plugins (Dataview, Templater, Excalidraw). Overkill for plain markdown vaults.
## 7. noesskeetit/second-brain-mcp
Architecture: MCP (read-only + semantic search)
What it does: Offline, plugin-free MCP server. Reads vault as plain files from disk, builds a local semantic index, and exposes semantic search rather than path-based file access. Works with Claude Code, Cursor, Zed, or any MCP-compliant client.
Setup difficulty: Low. One config command, one Claude config edit, done. Semantic index builds automatically on first run.
Best for: Multi-tool users. If you switch between Claude Code, Cursor, and Zed depending on task, this gives you a consistent semantic memory across all of them without duplicating setup.
Actual limitations: Read-only. Claude can query and retrieve but cannot write back - so this pairs well with a separate write-oriented tool like #4 or the Karpathy skills.
> Verdict: Best MCP for semantic retrieval across multiple AI tools. Pair with a write-capable tool for the full workflow.
## 8. CoMfUcIoS/second-brain-mcp
Architecture: MCP (strictly read-only)
What it does: Intentionally read-only MCP server. Claude can query the vault every way - search, follow links, extract metadata - but it cannot write, delete, or modify anything. Ever.
Setup difficulty: Low. Same profile as #7.
Best for: Security-conscious users. If you have 5+ years of notes and the idea of Claude accidentally rewriting one gives you anxiety, this removes the risk entirely.
Actual limitations: Read-only means you lose the automation loop (Claude can't file new notes for you). Every "save this insight" moment requires you to manually copy from chat into Obsidian.
> Verdict: The right choice for archivists. Wrong choice for automation workflows.
## 9. sunnyhasija/obsidian-plugin-skills
Architecture: Skills (Obsidian-specific expansions of #3)
What it does: Expands on Kepano's official skills with community-contributed additions - daily journal skill, meeting-notes skill, book-highlight extractor, spaced-review skill. All Obsidian-native, no MCP.
Setup difficulty: Low.
Best for: Users who liked #3 but wanted more variety out of the box.
Actual limitations: Community-maintained means quality varies skill by skill. Some are polished, some are drafts. Read each SKILL.md before enabling.
> Verdict: Good complement to #3. Cherry-pick specific skills rather than enabling everything.
## 10. C-Bjorn/MegaMem
Architecture: MCP + Temporal Knowledge Graph (powered by Graphiti)
What it does: Syncs your Obsidian vault into a temporal knowledge graph and exposes it to Claude via MCP. Ships with 12 graph tools + 11 vault file tools. Unlike every other MCP on this list, MegaMem tracks when concepts changed - so Claude can answer "what did I believe about X in April vs. now" instead of just "what do I believe about X".
Setup difficulty: Medium-high. Requires a running graph store (Graphiti backend) plus the MCP config. First-time setup ~40 minutes.
Best for: Anyone whose thinking evolves over time - analysts, researchers, journal writers. If you regularly change your mind about topics you've already documented, this is the only repo that captures the drift.
Actual limitations: Heaviest infrastructure footprint on the list. Overkill if your vault is mostly evergreen references. Requires more memory than pure-file MCPs.
> Verdict: The only repo here that treats your knowledge as changing over time. Reach for it when temporal reasoning matters.
## Which One to Install by User Type
> If you're new to this and just want a working vault: Install #1 AgriciDaniel/claude-obsidian. Follow the quickstart, ingest one file, watch the wiki appear. If you like it, you now have a working reference implementation.
> If you already have 500+ notes and want automation: Install #2 eugeniughelbur/obsidian-second-brain. Absorb the setup cost once, then never manually maintain your vault again.
> If you use Claude Code as your daily driver: Install #4 qhuang20/obsidian-skills. Skills-native, Claude Code-native, fastest time-to-working-state.
> If you're a researcher who values retrieval over automation: Install #7 noesskeetit/second-brain-mcp for read-only semantic search across your existing vault. Add write capability later with #4 if needed.
> If your vault is precious and you don't want automated writes: Install #8 CoMfUcIoS/second-brain-mcp. Strictly read-only. Nothing can be corrupted.
> If your Obsidian setup lives on plugins (Dataview, Templater): Install #6 iansinnott/obsidian-claude-code-mcp. It's the only one that goes through the Obsidian API and respects your plugin stack.
## 5-Minute Quickstart With Repo #1
```
# 1. Clone the repo
git clone https://github.com/AgriciDaniel/claude-obsidian.git
cd claude-obsidian
# 2. Install as a Claude Code plugin
claude plugin install .
# 3. Point it at a fresh Obsidian vault (empty folder works)
export OBSIDIAN_VAULT="$HOME/Documents/second-brain"
mkdir -p "$OBSIDIAN_VAULT"/{raw,wiki,instructions}
# 4. Drop a source file into raw/
cp ~/Downloads/some-paper.pdf "$OBSIDIAN_VAULT/raw/"
# 5. In Claude Code, run:
# /ingest raw/some-paper.pdf
```
Within a minute Claude reads the file, writes 8-15 linked wiki pages, and files them into your vault. Open Obsidian, switch to graph view, and you'll see the first nodes of what will become your second brain.
> The compounding moment: Every subsequent ingest expands the graph without ever re-reading the original source. That's the entire point.
## What's Not On This List, and Why
Smart Connections / Copilot / QuickAdd plugins. Solid Obsidian AI plugins, but they don't implement the wiki pattern or use MCP. They put AI features on top of your existing workflow rather than restructuring how the vault compiles knowledge. Different tool category.
Notion / Mem / Reflect. Not Obsidian, not open source, not on plain markdown. If you care about owning your data as portable files, they're not options.
RAG-only wrappers. Any repo that just wraps embeddings + retrieval without the wiki pattern is retrieving from your mess, not building order out of it.
> That's the whole insight from Karpathy - raw is source code, wiki is compiled product. RAG without compilation is grep with vibes.
## The Honest Take
You need two of these, not ten.
Pick one from the Karpathy Wiki family (#1, #4, or #5) to be the compiler that turns raw sources into structured knowledge. Pick one from the MCP family (#2, #7, or #8) to give Claude runtime access to that knowledge across every AI tool you use.
That's it. Two repos, one compiler, one bridge. Everything else on this list is either a variation of the two or a specialty tool for a specific edge case.
> The mistake most people make is installing five of these and losing weekends to config debt. Start with two. Build the vault. Add the third only when you feel the specific pain the third one solves.
## GitHub links
1. https://github.com/AgriciDaniel/claude-obsidian
2. https://github.com/eugeniughelbur/obsidian-second-brain
3. https://github.com/kepano/obsidian-skills
4. https://github.com/qhuang20/obsidian-skills
5. https://github.com/ekadetov/llm-wiki
6. https://github.com/iansinnott/obsidian-claude-code-mcp
7. https://github.com/noesskeetit/second-brain-mcp
8. https://github.com/CoMfUcIoS/second-brain-mcp
9. https://github.com/sunnyhasija/obsidian-plugin-skills
10. https://github.com/C-Bjorn/MegaMem
> i test setups like this on myself and share what actually works. follow @0xkkai if you want to see the next one.
## 相关链接
- [kai](https://x.com/0xkkai)
- [@0xkkai](https://x.com/0xkkai)
- [115K](https://x.com/0xkkai/status/2085838657068347401/analytics)
- [https://github.com/AgriciDaniel/claude-obsidian](https://github.com/AgriciDaniel/claude-obsidian)
- [https://github.com/eugeniughelbur/obsidian-second-brain](https://github.com/eugeniughelbur/obsidian-second-brain)
- [https://github.com/kepano/obsidian-skills](https://github.com/kepano/obsidian-skills)
- [https://github.com/qhuang20/obsidian-skills](https://github.com/qhuang20/obsidian-skills)
- [https://github.com/ekadetov/llm-wiki](https://github.com/ekadetov/llm-wiki)
- [https://github.com/iansinnott/obsidian-claude-code-mcp](https://github.com/iansinnott/obsidian-claude-code-mcp)
- [https://github.com/noesskeetit/second-brain-mcp](https://github.com/noesskeetit/second-brain-mcp)
- [https://github.com/CoMfUcIoS/second-brain-mcp](https://github.com/CoMfUcIoS/second-brain-mcp)
- [https://github.com/sunnyhasija/obsidian-plugin-skills](https://github.com/sunnyhasija/obsidian-plugin-skills)
- [https://github.com/C-Bjorn/MegaMem](https://github.com/C-Bjorn/MegaMem)
- [@0xkkai](https://x.com/0xkkai)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [5:21 AM · Aug 8, 2026](https://x.com/0xkkai/status/2085838657068347401)
- [115.7K Views](https://x.com/0xkkai/status/2085838657068347401/analytics)
- [View quotes](https://x.com/0xkkai/status/2085838657068347401/quotes)
---
*导出时间: 2026/8/9 20:51:23*
---
## 中文翻译
# 10个免费的开源GitHub库：将Claude + Obsidian打造成真正的第二大脑
**作者**: kai
**日期**: 2026-08-07T21:21:03.000Z
**来源**: [https://x.com/0xkkai/status/2085838657068347401](https://x.com/0xkkai/status/2085838657068347401)
---
在过去的四个月里，GitHub上悄然涌现了许多代码库，它们从不同角度解决着同一个问题：如何以Obsidian为存储层，让Claude真正记住你所知道的知识。
其中大部分只是简单的封装，有些是架构设计，而有两个绝对值得你今天就围绕它们建立工作流。这是一份客观真实的剖析——每个库的实际作用、局限性，以及根据你的身份，你应该先安装哪一个。
> 在列出清单之前，你需要了解三种架构。否则，你看每个库都会觉得差不多，最终会浪费整个周末去安装四个功能相同的工具。
## 本榜单中每个代码库背后的三种架构
1. Karpathy的LLM Wiki模式
2026年4月，安德烈·卡帕西（Andrej Karpathy）在GitHub上发布了一篇800字的 gist，名为 LLM-WIKI。其中描述了一种三文件夹架构：`raw/` 用于存放源文档，`wiki/` 用于存放Claude实际读取的已编译链接页面，而 `instructions/` 则用于存放防止wiki陷入混乱的规则。
让这个理念迅速火爆的举措是：Claude在其整个生命周期内只读取任何给定的原始文件一次，提取出关键信息，将其写入wiki，然后再也不碰那个原始文件。
> 为什么这很重要：随后的每次查询都会从编译好的wiki中提取信息，这将重复查询的Token消耗削减了70-90%，并为你提供了一个真正在积累增值而不是衰退的知识库。
属于该类别的代码库：#1、#4、#5、#10。
## 2. 技能架构
Anthropic在7月份悄然推出了一套文件夹系统，Claude可以根据你的提问自动加载特定指令。你可以将一个包含 `SKILL.md` 的文件夹放入你的技能目录中，该文件描述了何时触发以及执行什么操作。Claude在启动时只会读取描述——完整的技能只有在触发时才会加载。
对于Obsidian的知识库来说，这意味着你可以定义诸如“当我提到一篇论文时，激活paper-reader技能”或“当我打开知识库时，加载我的笔记组织规范”之类的工作流，而不会污染你的主上下文窗口。
属于该类别的代码库：#3、#9。
## 3. MCP 桥接
模型上下文协议（MCP）是Anthropic发布的用于将Claude连接到外部工具和数据的标准。MCP服务器充当着一座实时桥梁——Claude发送结构化请求，服务器查询你的知识库，并实时返回结果。
MCP代码库分为两种类型：只读（安全，Claude只能拉取数据）和读写（强大，Claude可以创建/编辑笔记）。
> 经验法则：大多数研究人员应该从只读模式开始。而读写模式则是实现“知识库在你睡觉时自我构建”工作流的基础。
属于该类别的代码库：#2、#6、#7、#8。
## 10个代码库：排名与深度分析
## 1. AgriciDaniel/claude-obsidian
架构：Karpathy Wiki模式（最纯粹的实现）
功能：你可以将任何源文档——PDF、Markdown笔记、网页剪藏、文字稿——拖入 `raw/` 文件夹。Claude读取一次，将其编译成8-15个相互链接的wiki页面，并将其归档到图表中。下次查询时，Claude只会从wiki中提取信息，并已附注信息来源。
安装难度：中等。你需要将其作为Claude Code插件安装，指向一个空的知识库，然后开始拖入文件。初始知识库需要20-30分钟才能配置成你想要的样子。此后，导入数据只需一条命令。
最适合：刚接触Karpathy模式并希望获得最清晰参考实现的人。这是最接近规范设置的方案，也是大多数围绕该模式制作内容的创作者正在使用的工具。
实际局限：文件夹结构过于主观——如果你已经有一个采用不同方式组织的库，迁移过程将非常痛苦。没有内置语义搜索——检索完全依赖于Claude打开wiki并跟随链接进行。并非为团队使用而设计。
> 结论：如果你正在构建你的第一个知识库，请从这里开始。如果你已经有500+篇结构不同的笔记，请跳过它。
## 2. eugeniughelbur/obsidian-second-brain
架构：MCP（读写）+ 技能 + 自定义命令
功能：为Claude Code和其他六种命令行代理提供持久记忆。自带45个命名命令，包括混合语义搜索（BM25 + 嵌入）、当你提出矛盾信息时能自我更新的笔记、无需密钥的网络研究（使用公开搜索），以及在你睡觉时维护知识库的定时代理。
安装难度：初次较高，之后永久极低。首次设置需要安装本地嵌入模型并配置调度程序。一旦运行起来，它是这个列表中最免维护的系统。
最适合：拥有1000+篇笔记且不想再当“保姆”去手动维护的高级用户。
> 杀手级功能：定时代理。你可以让Claude每晚对你的wiki进行代码检查，每周对相似笔记进行去重，或者每月重新浮现陈旧内容，全程无需任何干预。
实际局限：复杂性成本。阅读文档需要90分钟。本地嵌入模型需要约4GB内存。某些命令专门针对Claude Code设计，无法在Claude桌面版或网页版中运行。
> 结论：这是榜单上最强大的配置。当你的知识库庞大到你已无法手动维护时，就可以使用它了。
## 3. kepano/obsidian-skills
架构：技能（Anthropic原生模式）
功能：提供由Kepano（Obsidian创始人）编写的预构建代理技能。兼容Claude Code、Codex和Open Code。每个技能都是一个包含 `SKILL.md` 的文件夹，用于描述何时触发以及执行什么操作——无需手动进行上下文管理。
安装难度：低。将代码库克隆到你的Claude技能文件夹中即可。技能会在Claude下次启动时自动注册。
最适合：想要“准官方”默认设置的人。因为这些技能由Kepano编写，所以其约定与Obsidian实际的知识管理理念（文件夹、标签、反向链接）高度契合。如果你已经对Obsidian生态系统非常熟悉，会觉得它非常原生化。
实际局限：仅限技能，无MCP支持。你无法进行实时查询——Claude通过文件系统读取知识库，因此实时的库状态变更需要刷新。此外，提供的技能数量少于像#9这样的社区克隆版。
> 结论：对于已经深谙Obsidian理念的人来说，这是最稳妥的选择。它的信任度极高——你运行的是由该软件构建者本人编写的代码。
## 4. qhuang20/obsidian-skills
架构：技能 + Karpathy Wiki模式
功能：这是一个Claude Code插件，自带直接实现Karpathy模式的专用 `llm-wiki` 技能，以及几个用于笔记管理的实用技能。通过提及信息源来触发wiki技能；Claude会在后台默默完成信息提取和链接工作。
安装难度：中低。标准的Claude Code插件安装方式。该wiki技能需要一个描述你知识库命名约定的初始配置文件。
最适合：想要直接使用Karpathy模式而不想从零开始搭建的用户。如果你特别习惯在Claude Code中操作，它比#1号代码库能更快地构建出可用的wiki。
实际局限：没有#1经过那么多实战测试。实用技能虽然方便，但与#3和#5有重叠，因此你需要做出取舍。
> 结论：如果你已经离不开Claude Code，这是一个非常可靠的工具。如果你想要更纯粹的参考实现可以考虑#1，想要更纯粹的Karpathy模式可以看#5。
## 5. ekadetov/llm-wiki
架构：Karpathy Wiki模式（纯粹专注版）
功能：除了Karpathy模式外什么都没有，完全作为一个Claude Code插件实现。没有花里胡哨的附加功能。导入、查询、保存、代码检查——正是原版 gist 中的六个命令，且经过了优化。
安装难度：极低。安装，指向知识库，运行 `/init`。5分钟内即可实现数据导入。
最适合：纯粹主义者。如果你希望完全按照Karpathy所描述的方式来实现该模式，且不希望任何其他功能把界面搞得复杂化，这就是为你准备的。
实际局限：没有语义搜索，没有定时代理，没有MCP桥接。刻意追求极简，这既可以是它的一个优点，也可能直接劝退你，完全取决于你的个人需求。
> 结论：最佳的极简版Karpathy模式实现。作为起点非常好，但随着时间推移你可能会觉得不够用。
## 6. iansinnott/obsidian-claude-code-mcp
架构：MCP（通过Obsidian插件进行读写）
功能：这是一个在你的Obsidian实例中运行MCP服务器的插件。当Claude Code查询该服务器时，它可以读取文件、创建新笔记、修改现有笔记、追踪反向链接并更新图谱——所有这些都是通过Obsidian自带的API完成的，因此不会绕过Obsidian的文件安全保障机制。
安装难度：中等。作为Obsidian社区插件安装，启用MCP服务器，将端点添加到Claude Code配置中。桥接工作的前提是必须保持Obsidian处于运行状态。
最适合：希望Claude与实时Obsidian实例进行交互的用户——能够完全兼容你的插件、模板、Dataview查询等所有内容。
> 与众不同之处：它是本榜单中唯一通过Obsidian的API而不是原始文件系统进行操作的代码库。
实际局限：需要保持Obsidian处于打开状态。桥接延迟高于基于文件系统的MCP（每次查询约200ms，而文件系统约为20ms）。不支持Obsidian移动端。
> 结论：如果你的知识库严重依赖Obsidian插件（如Dataview、Templater、Excalidraw），这是正确的选择。但对于仅包含纯Markdown文本的库来说则有些大材小用。
## 7. noesskeetit/second-brain-mcp
架构：MCP（只读 + 语义搜索）
功能：离线、无插件的MCP服务器。从磁盘将知识库作为纯文件读取，构建本地语义索引，并提供基于语义的搜索而非基于路径的文件访问。兼容Claude Code、Cursor、Zed或任何支持MCP的客户端。
安装难度：低。只需一条配置命令，一次Claude配置文件编辑，即可完成。语义索引在首次运行时自动构建。
最适合：多工具用户。如果你根据任务在Claude Code、Cursor和Zed之间切换，这个工具能让你在所有工具中拥有一致的语义记忆，而无需重复设置。
实际局限：只读。Claude可以查询和检索，但无法写回——因此它与像#4这样专门负责写入的工具或Karpathy技能搭配使用效果最佳。
> 结论：跨多种AI工具进行语义检索的最佳MCP。配合一个具备写入能力的工具即可组成完整工作流。
## 8. CoMfUcIoS/second-brain-mcp
架构：MCP（严格只读）
功能：刻意设置为只读的MCP服务器。Claude可以通过各种方式查询知识库——搜索、追踪链接、提取元数据——但它永远无法执行写入、删除或修改任何内容的操作。
安装难度：低。与#7相同。
最适合：注重安全性的用户。如果你有超过5年的笔记，且一想到Claude不小心重写了其中的某篇就会感到焦虑，那么这款工具能完全消除你的风险。
实际局限：只读意味着你失去了自动化闭环（Claude无法为你归档新笔记）。每个“保存这个灵感”的瞬间，都需要你手动从聊天窗口复制到Obsidian中。
> 结论：档案管理者的正确选择。但对于追求自动化工作流的人来说则是错误的选择。
## 9. sunnyhasija/obsidian-plugin-skills
架构：技能（对#3的Obsidian专属扩展）
功能：在Kepano官方技能的基础上进行了社区扩充——包含每日日志技能、会议记录技能、书籍高亮提取器、间隔复习技能。全部为Obsidian原生，无MCP。
安装难度：低。
最适合：喜欢#3但希望开箱即用拥有更多样化功能的用户。
实际局限：由社区维护意味着各个技能的质量参差不齐。有些非常完善，有些则只是草稿。启用前请务必阅读每个 `SKILL.md`。
> 结论：作为对#3的补充非常棒。建议精挑细选特定的技能，而不是一股脑全部启用。
## 10. C-Bjorn/MegaMem
架构：MCP + 时间知识图谱（由Graphiti提供支持）
功能：将你的Obsidian知识库同步到一个时间知识图谱中，并通过MCP暴露给Claude。自带12种图谱工具 + 11种知识库文件工具。与榜单上的其他MCP不同，MegaMem能追踪概念发生变化的时间——因此Claude不仅能回答“我现在对X的看法是什么”，还能回答“相比四月份，我对X的看法发生了什么变化”。
安装难度：中高。需要运行一个图谱存储库（Graphiti后端）以及进行MCP配置。首次设置大约需要40分钟。
最适合：思维会随时间演变的人——分析师、研究人员、日志作者。如果你经常对已经记录过想法的主题改变看法，这是唯一能捕捉到你认知变化的代码库。
实际局限：在所有工具中基础设施占用最为庞大。如果你的知识库主要是一些永久不变的参考资料，那就大材小用了。比纯文件MCP需要更多的内存。
> 结论：这是榜单上唯一将你的知识视为随时间推移而变化的工具。当你需要进行基于时间的推理时，请选择它。
## 根据用户类型决定安装哪一个
> 如果你是新手，只想要一个能用的知识库：安装 #1 AgriciDaniel/claude-obsidian。按照快速入门指南导入一份文件，看着wiki自动生成。如果你觉得不错，那么你就已经拥有了一个可用的参考实现。
> 如果你已经有500+篇笔记并希望实现自动化：安装 #2 eugeniughelbur/obsidian-second-brain。咬牙承受一次安装成本，之后永远无需再手动维护你的知识库。
> 如果你日常主要使用Claude Code：安装 #4 qhuang20/obsidian-skills。原生支持技能，原生支持Claude Code，能以最快速度进入工作状态。
> 如果你是更看重检索能力而非自动化的研究人员：安装 #7 noesskeetit/second-brain-mcp 对现有知识库进行只读语义搜索。如果以后需要写入功能，可以再添加 #4。
> 如果你的知识库极其宝贵，不想让AI自动写入：安装 #8 CoMfUcIoS/second-brain-mcp。严格只读，绝对不会有文件被损坏。
> 如果你的Obsidian重度依赖插件（如Dataview、Templater）：安装 #6 iansinnott/obsidian-claude-code-mcp。它是唯一通过Obsidian API进行操作并兼容你插件栈的工具。
## 使用 #1 代码库的5分钟快速入门
```
# 1. Clone the repo
git clone https://github.com/AgriciDaniel/claude-obsidian.git
cd claude-obsidian
# 2. Install as a Claude Code plugin
claude plugin install .
# 3. Point it at a fresh Obsidian vault (empty folder works)
export OBSIDIAN_VAULT="$HOME/Documents/second-brain"
mkdir -p "$OBSIDIAN_VAULT"/{raw,wiki,instructions}
# 4. Drop a source file into raw/
cp ~/Downloads/some-paper.pdf "$OBSIDIAN_VAULT/raw/"
# 5. In Claude Code, run:
# /ingest raw/some-paper.pdf
```
在一分钟内，Claude就会读取文件，编写8-15个相互链接的wiki页面，并将它们归档到你的知识库中。打开Obsidian，切换到图表视图，你就会看到将成为你“第二大脑”的最初节点。
> 知识复利时刻：随后的每一次导入都会扩展你的图表，且永远不需要重新阅读原始信息源。这就是这套系统的全部核心要义。
## 哪些工具不在榜单上，以及为什么
Smart Connections / Copilot / QuickAdd 插件。这些都是非常出色的Obsidian AI插件，但它们没有实现wiki模式，也没有使用MCP。它们只是在你现有的工作流之上叠加AI功能，而不是重塑知识库编译知识的方式。属于不同的工具类别。
Notion / Mem / Reflect。不是Obsidian，不开源，也不基于纯Markdown。如果你在乎以可移植文件的形式掌控自己的数据，它们就不在你的选择范围内。
纯RAG封装工具。任何仅仅是封装了“嵌入+检索”而没有采用wiki模式的代码库，都只是在你的一团乱麻中进行检索，而不是在帮你建立秩序。
> 这就是Karpathy的全部洞见所在——原始文件是源代码，wiki是编译后的产品。没有经过编译的RAG，不过是凭感觉在用grep搜索。
## 最真诚的建议
你只需要这其中的两个，而不是十个。
从Karpathy Wiki家族（#1、#4 或 #5）中挑选一个，作为将原始信息源转化为结构化知识的编译器。从MCP家族（#2、#7 或 #8）中挑选一个，让Claude在你使用的所有AI工具中都能在运行时访问这些知识。
就是这样。两个代码库，一个编译器，一座桥梁。这份名单上的其他所有工具，要么是这两者的变体，要么是针对特定边缘情况的专业工具。
> 大多数人都会犯的一个错误是，安装了五个这样的工具，然后把周末的时间都耗费在处理配置债务上。从两个开始。建立好你的知识库。只有当你确切感受到某个工具能解决你的特定痛点时，再添加第三个。
## GitHub 链接
1. https://github.com/AgriciDaniel/claude-obsidian
2. https://github.com/eugeniughelbur/obsidian-second-brain
3. https://github.com/kepano/obsidian-skills
4. https://github.com/qhuang20/obsidian-skills
5. https://github.com/ekadetov/llm-wiki
6. https://github.com/iansinnott/obsidian-claude-code-mcp
7. https://github.com/noesskeetit/second-brain-mcp
8. https://github.com/CoMfUcIoS/second-brain-mcp
9. https://github.com/sunnyhasija/obsidian-plugin-skills
10. https://github.com/C-Bjorn/MegaMem
> 我亲自对这类设置进行测试，并分享真正有效的经验。如果你想看下一个测试，请关注 @0xkkai。
## 相关链接
- [kai](https://x.com/0xkkai)
- [@0xkkai](https://x.com/0xkkai)
- [115K](https://x.com/0xkkai/status/2085838657068347401/analytics)
- [https://github.com/AgriciDaniel/claude-obsidian](https://github.com/AgriciDaniel/claude-obsidian)
- [https://github.com/eugeniughelbur/obsidian-second-brain](https://github.com/eugeniughelbur/obsidian-second-brain)
- [https://github.com/kepano/obsidian-skills](https://github.com/kepano/obsidian-skills)
- [https://github.com/qhuang20/obsidian-skills](https://github.com/qhuang20/obsidian-skills)
- [https://github.com/ekadetov/llm-wiki](https://github.com/ekadetov/llm-wiki)
- [https://github.com/iansinnott/obsidian-claude-code-mcp](https://github.com/iansinnott/obsidian-claude-code-mcp)
- [https://github.com/noesskeetit/second-brain-mcp](https://github.com/noesskeetit/second-brain-mcp)
- [https://github.com/CoMfUcIoS/second-brain-mcp](https://github.com/CoMfUcIoS/second-brain-mcp)
- [https://github.com/sunnyhasija/obsidian-plugin-skills](https://github.com/sunnyhasija/obsidian-plugin-skills)
- [https://github.com/C-Bjorn/MegaMem](https://github.com/C-Bjorn/MegaMem)
- [@0xkkai](https://x.com/0xkkai)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [5:21 AM · Aug 8, 2026](https://x.com/0xkkai/status/2085838657068347401)
- [115.7K Views](https://x.com/0xkkai/status/2085838657068347401/analytics)
- [View quotes](https://x.com/0xkkai/status/2085838657068347401/quotes)
---
*导出时间: 2026/8/9 20:51:23* 
