我用DeepSeek V4+Claude Code+Ob (English)

Generated: 2026-06-24 17:49:23

---

I spent three days and finally found that AI knowledge base solution that gets you Audi performance for the price of an Alto

Ever feel this way?

More and more information, but less and less brainpower to handle it. All those great articles you read every day, book notes, project docs—scattered everywhere like a pile of junk you never sorted. And when you need something, you dig for ages—and still can’t find it!

I’ve been tinkering with personal knowledge bases off and on for five years now.

Five years. That’s longer than some people’s love lives.

I started with Notion—data in the cloud, spinning wheel for half an hour every time I opened it, and this nagging fear: what if the server goes down one day and all my notes go with it?

Then I switched to Yuque. Well, we all know what happened there.

Evernote? It got slower and slower. Searching for something took half a minute. I lost my temper and just uninstalled it.

What I really wanted was simple: data on my own hard drive, AI that could help me dig things up, and don’t make it too expensive.

Just those three requirements.

Guess what? No single solution could deliver all three at once.

Either it was outrageously expensive—Notion AI at 10 bucks a month, which barely lets you write anything before it runs out, just fixing a few typos and it’s gone. Or the AI was so dumb it made you want to pull your hair out—some local knowledge bases still rely on keyword matching, which is basically using year-2000 technology to manage 2025-level information.

It wasn’t until I recently started using the DeepSeek API that I felt this was finally practical.

And not just practical—I mean the kind of satisfaction you get when you pay the price of an Alto and drive away in an Audi.

---

Why I chose this combo

Let me start with my reasoning.

Obsidian—I’ve been using it for over two years. Local Markdown files, all on my own machine. The plugin ecosystem is incredibly powerful. How powerful? If you can dream it, there’s probably a plugin for it. But there’s one fatal flaw—searching is basically Ctrl+F. Trying to find a note that says “microservice decomposition last March” among two thousand notes? You’ll be scrolling forever. Tags, you say? I’ve gone through three different tagging systems, and they all ended up a mess.

Claude Code is a command-line tool from Anthropic. It can read files, search content, and do cross-directory analysis. It works great, but using the Claude API directly is just too expensive—like burning money.

DeepSeek’s API caught my eye—cost per million tokens is less than a tenth of Claude’s, and its text understanding is close to their own closed-source models. What does that mean? For the price of a cup of milk tea, you can run analyses on hundreds of thousands of words.

String these three together, and the logic is clear: Obsidian as the warehouse, Claude Code as the hands and feet, DeepSeek API as the brain. Data stays local, AI capabilities come online, and the price is low.

Just looking at the setup, I felt it could work. But I didn’t just theorize—I actually built it, ran it for three days, and only then came back to write this.

---

Four steps from zero to running (don’t worry, it’s not complicated)

Step 1: Install Obsidian

Download from the official site, double-click, install—free.

My advice: Don’t put your Vault on the system drive. I put mine at D:\my-obsidian-vault. If the system crashes and you need to reinstall, your notes won’t be lost. Imagine writing faithfully for three years and then one blue screen wipes everything out—that’s devastating.

Step 2: Install the Claudian plugin

This plugin lets Claude Code work directly inside Obsidian, so you don’t have to switch to the command line.

Claudian isn’t on the official marketplace yet, so you need to install it via BRAT. Open Obsidian Settings, go to third-party plugins, disable safe mode, search for BRAT, install it, and enable it. In BRAT, click “Add beta plugin”, enter https://github.com/YishenTu/claudian, choose latest version, and click Add.

It takes about 5 minutes total.

One pitfall: BRAT sometimes reports network errors. Don’t panic—try again. I only succeeded on the third try. I almost threw my laptop across the room, but I told myself good things take time.

Step 3: Install Claude Code and configure the DeepSeek API

First, install Node.js. Mac users can run brew install node, Windows users go to nodejs.org and download the LTS version. After installation, run in your terminal:

node -v
npm -v

If you see version numbers, you’re good.

Then install Claude Code globally:

npm install -g @anthropic-ai/claude-code

Now the key part: create a .claude folder in your user directory, create a settings.json file inside, and write the following content. Note that the DeepSeek API is fundamentally compatible with the OpenAI format, but Claude Code needs to call the Anthropic interface. I used a local proxy bridge, so the actual configuration points to a local service instead of directly using DeepSeek’s address. You can also refer to the comments section for other methods, or use a third-party translation layer to call DeepSeek via Claude Code.

{
 "env": {
 "ANTHROPIC_AUTH_TOKEN": "your DeepSeek API Key",
 "ANTHROPIC_BASE_URL": "http://localhost:8080/anthropic",
 "ANTHROPIC_MODEL": "deepseek-chat",
 "API_TIMEOUT_MS": "300000",
 "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
 }
}

Get your API Key from the DeepSeek website—one yuan buys you tens of thousands of tokens.

I set the timeout to 300 seconds—long document operations might be slow, and too short a timeout can cause disconnections. I had it at 60 seconds before, and it kept dropping out during long note sessions, which drove me crazy.

Step 4: Configure Obsidian

Open Obsidian Settings → About → Advanced → Enable Command-line Interface Mode.

Then in the Claudian plugin