Recently, I came across a tweet claiming that you can run Claude Code locally for free, no API or cloud needed. I was really curious and followed the process through.


First, let's talk about the steps—they're not complicated:

**Step 1: Install Ollama**

This is a local model runtime engine that works on both Mac and Windows. After installation, it just runs quietly in the background. To verify it's working properly, visit localhost:11434.

**Step 2: Download the Model**

If your machine has good specs, you can try qwen3-coder:30b; for average specs, gemma:2b or qwen2.5-coder:7b will also work. Just run one command in the terminal to download it.

**Step 3: Connect to the Local "Brain"**

Set two environment variables to direct Claude's requests to the local Ollama service:


```bash

export ANTHROPIC_BASE_URL="http://localhost:11434"

export ANTHROPIC_AUTH_TOKEN="ollama"

```

**Step 4: Start Running**

Go to your project directory and type `claude` to start chatting.

It's true that you don't need an internet connection and don't have to pay for API calls. Privacy is also maximized—all code runs on your own machine.

---

**But several issues mentioned in the comments do exist:**

A user named RemiDev asked: Can only Qwen models be used? The answer is yes, for now. There are many optional models in the Ollama ecosystem, but compatibility with Claude Code's interface is currently limited.

Alex Kyprianou put it more bluntly: "Qwen sucks compared to opus". While this is a bit extreme, there is indeed a gap in reasoning ability between local models and the official Claude version. John Almighty even said Ollama's "reasoning and logic is at 0".

Another practical issue— not all computers can run it. Models with 30b parameters have high requirements for memory and GPU. If you force a regular thin-and-light laptop to run it, chances are the fan will spin wildly and it will lag so much you'll question life.

Also, calling this a "new discovery" is indeed an overstatement. The combination of Ollama + local models has been playable since 2025. Tech savvie added in the comments that similar results can be achieved with LiteLLM.

---

**My Judgment:**

Free, no internet required, and privacy protection—these three points are indeed attractive to some people. But if you're looking for Claude's native reasoning quality, running a scaled-down version locally may disappoint you. It's more reasonable to treat it as an interesting learning toy or an alternative in specific scenarios (such as offline environments or situations with extremely high code privacy requirements).

If you want to experience AI programming at no cost, it's worth a try. Don't set your expectations too high—just treat it as a game.