Made withOllamaClip

Claude Code + Ollama = Free Unlimited Coding AI

E

Eric Tech

Creator on the Ollama channel

Watch on YouTube ↗

381K viewsFeb 16, 202614:10 runtime

Learn how to use Claude Code with Ollama to run local large language models for free, avoiding expensive subscription plans. This guide walks you through setting up Ollama, comparing recommended coding models like Qwen and Deepseek Coder based on your hardware, and configuring…

Chapters

#ollama#clip#aillm

Using Local and Cloud AI Models with Claude Code via Ollama

This document outlines how to integrate Ollama with Claude Code to run Large Language Models (LLMs) for coding assistance, offering a free alternative to paid plans. It covers setup, model selection, practical strategies, and performance comparisons.

Getting Started

Install Ollama: Download Ollama from the official website and install it using the provided terminal command for your operating system.
Verify Installation: Check if Ollama is installed correctly by running ollama --version in your terminal.
List Models: View your installed models at any time with the ollama list command.

Model Selection and Configuration

Claude Code's launcher simplifies model selection and installation.

Launching: In your project directory, run launch claude. The tool will suggest models based on your machine's VRAM.
Choosing a Model: Select a model that fits your hardware. For a 16GB VRAM machine, a model like Qwen 8B (requiring ~11GB) is a suitable choice.
Switching Models: If you have multiple models installed, you can easily switch between them using the ollama launch claude --config command, which presents a selection menu.

Note
You can also manually download any model from the Ollama library using ollama pull <model-name>.

Strategy for Effective Local AI Coding

Local models are generally less powerful than premium models like Claude Opus. To maximize their effectiveness, adopt a component-driven development approach.

Isolate Tasks: Use tools like Storybook to break down your application into individual, standalone components.
Reduce Context: By focusing on one component at a time, you provide the AI with a much smaller and more manageable context, leading to better and more accurate results.

Practical Workflow & Model Comparison

The video demonstrates redesigning an email template component to compare model performance.

Local Model (GLM-4 9B):
- Process: Used the frontend_design skill to generate a wireframe and then code.
- Result: The output contained significant inaccuracies (e.g., placing the word "verify" as a logo). This highlights the accuracy trade-off with smaller local models.
Cloud Model (GLM-4.7 via Ollama):
- Process: Connected to the cloud-hosted GLM-4.7 model and repeated the task.
- Result: Produced a much cleaner, more accurate design that respected the existing project structure. It was able to successfully refine the design based on iterative, natural language feedback with screenshots.

Important
The GLM-4.7 model is not truly local; it runs in the cloud but is accessed via the Ollama interface. It offers significantly better performance but relies on an external service.

Key Takeaways

Hybrid Approach: For best results, use a hybrid approach. Use small, fast local models for simple tasks and switch to a more powerful cloud-connected model like GLM-4.7 for complex generation or design tasks.
GLM-4 Free Tier: The GLM-4 model offers a free tier with a limit of 250,000 input tokens per hour, which is generally sufficient for a single coding session in Claude Code.
Component-Driven Development: Breaking down problems is the key to successfully using less powerful local LLMs.

Introduction: Free Unlimited Coding AI with Ollama

Upgrading to the Mac plan with Claude Code is quite expensive. And that is why in this video I'm going to show you how you can use Ollama with Claude Code to run local models inside of your local machine. We're going to explore all the recommended models and compare which model is the best based on your computer specs to building applications using Claude Code with local coding large language model. And furthermore, we're going to walk through step by step on how to set this up on your local machine, how you can be able to configure this, switch to different models, and find out which model is best for you. So with that being said, if you're interested, let's get into it.

Installing and Verifying Ollama

All right, so to get started, first thing first, we're going to navigate to Ollama, which is a place where we can download and be able to use any large language model here, open source, onto our local machine. So what we're going to do here is we're going to click on download. And based on your operating system here, I'm just going to choose Mac because I'm using Mac, but feel free to choose whichever operating system you're using for your local machine. But I'm just going to copy this command right here and paste it into my terminal and install this onto my local machine.

Awesome.

Choosing and Installing a Model for Claude Code

So once we have the Ollama here installed onto our local machine, the next thing we're going to do here is try to run it and see if it's actually successfully installed. So if I were to do ollama --version, I can be able to see the current version that I'm using. So what I can also do here is I can also do ollama list to be able to see all the list of models that I have currently on my local machine. So if I were to run this, you can see currently I don't have any models. So in our case, let's take a look at how to install one for our Claude Code.

So what we're going to do here is we're going to go to models right here instead of Ollama. And right here you can see that these are the list of models that we installed. But specifically we're going to install one for Claude Code. So all I had to do here is just click on the landing page. So if we were to scroll down to the landing page here you can see there's a coding section and for that there's a Claude Code. So if I were to click on this it will basically shows you the recommended models for Claude Code. For example Qwen Coder and GPT-OSS.

Launching and Configuring Local Models

So what I can do here is that let's say if I were to cd to my project called bookkeeping applications and all I had to do here is just do launch claude and simply you can see that it gives you a recommended models that you can choose from. For example, GO-4.7 Flash for 25 GB of VRAM, but currently my machine is using MacBook Pro. So mine is only 16 GB RAM for the VRAM. So what I will go with is probably just the Qwen 3 8 billion parameters, which is only 11 GB for the VRAM. So in this case I'm just going to choose that option for now.

And of course, you can also choose some other options based on your machine. And you can actually be able to learn more about the specs by clicking on the models in Ollama and learn more about which model is best fit for you. So, for example, like GPT-OSS, there's like the 65 GB RAM that you can choose from. So, whichever option is best for you, you can go with that. Right?

So, what we can do now is I'm just going to choose this option right here, Qwen 3, a billion parameters. And I'm just going to say yes, download this. And it's going to pull the manifest and try to set this up on your local machine. But obviously if you don't want to download this way, you want to download this manually with the model that you like. For example, like GPT-OSS 20 billion parameters, what you can do is that you just open your terminal and just do ollama pull for this models and it will simply just download it, right? And then once you've done download that, you're just going to do ollama list and it will basically show all the models that you have in your coder machines.

Switching Between Multiple Models

But in our case here, back to the installation for the Qwen 3 coder. Here you can see that it's asking us if you want to access the workspace. So quick safety check, is the current project created or one trusted like you own the code, well-known open source project. So I will say yes I trust this project and right here you can see we have our Claude Code model here is switch to the Qwen 3 3B parameters here. So if I were to say hi, let me just run this command right here and let's see what it generates. And here you can see this is what it generates. It says, "Hello, how can I assist you today?"

Okay. And if you're curious about how you can be able to get the status line right here, you can check out this video right here, which I talk about how you can get your adding your status line bar, which shows the current context window that you have with your current conversation with any large language model inside of Claude Code. So feel free to check that video out to learn more. And of course, let's say we have multiple models here in our local machine. How can we be able to switch it between those models in our Claude Code? For example, all you have to do here is just ollama launch claude and you just do --config. And here you can be able to have the option to select which model you want. So I'm just going to choose one here that I have already installed which you can see here we can choose GLM4 and also we have our Qwen 2.5 coder, right? So these are all the options that we have. All right.

Strategy: Using Storybook for Effective Local AI Coding

So that's pretty much how we can be able to, you know, use Ollama here connected with Claude Code using any local large language model here for your Claude Code coding sessions, right? But now what I want to do is basically talk about how we can use it because let's be honest, all those open source models here are no better than the Opus 4. So how can we be able to make sure that when we use our local models with Claude Code with the best performance with the best efficiency? And that comes with how we're going to develop our application. For example, if you're going to build the entire front end by just saying, "Oh, hey, build this front end applications." Well, that's going to be a lot of components, that's going to be a lot of pages thrown into the context. But what if we can pitch people to use like Storybook to be able to actually break those components down into its own standalone component and then just perfect it, right? We're only perfecting one component at a time. So that means the less context where AI had to figure it out. So in this case, that would be perfect for, you know, using local model here to do that.

So if you wanted to learn more about Storybook, you can actually be able to navigate to Storybook. It is actually npm, very simple way, simply just run this command here onto your project and you can see that all the components, all the pages, right, broken down into this own single page and all we have to do here is just perfect one single component, one single page. So we have those building blocks when we try to, you know, use like Opus or any other like big large language model here to building, you know, blocks together, right?

So for example, if I were to just run the npm run storybook and here it's going to start my story book which you can see here.

Example: Redesigning a Component with a Local Model

First of all, I'm just going to navigate to this story book that I created which contains like all my email templates, the error pages that we have, right? As well as the base tables, so different tables that we have in our applications.

So in this case, what I want to do now is basically showing you exactly okay, let's say if I wanted to use the local models here, right? Take it one step at a time. This is actually one single component and literally what we can do is basically take it one component at a time. For example, this email verification page, right? Where user just click on it, it will basically verify themself. And what we can do here is that I'm just going to copy this link and come back to Claude Code for example. Let's say if I'm using GLM4 with 9B parameters and I'm just going to paste this one right here. Or even better, we can even use the frontend_design skill from Claude skill. And if you haven't checked out my Claude skill video on how to use it inside of Claude Code, then feel free to check it out. But what you can do here is that you can use the frontend_design skill which was, you know, part of the Claude features. And then this what we're going to do first, I'm just going to turn on the plan mode and basically paste the link and then use the frontend_design skill for slash command. And now what I'm going to do here is I'm just going to say, "Based on the current email verification email template, I want you to trigger the frontend_design skill here to redesign the entire email. Have a cleaner look of our email template for this particular email verification. So I want you to do that and please help me to complete this. And before you generate the code, I want you to create a terminal wireframe to show me exactly what is it going to look for the email template before we do the implementation. Please do that. Thanks." And that's basically my prompt. And I'm just going to run this and try to see what it does here.

Analyzing Local Model Limitations and Seeking Better Options

All right. Then here you can see that we have our results, which here you can see that we have our terminal wireframe. Now I don't think this is really accurate but here you can see we have our button here, the logo right, and also text company. So it makes our email template much more shorter and much more concise. But now if I were to look at the code that is going to generate in plan mode, you can see that this is the entire code, right? We have our title but what's missing here is sometimes the accuracy for local models here isn't really accurate, right?

You can see that it generates a HTML page but it doesn't show. So you can see that this is logo but what we put in the logo is verify, right? It's not, we shouldn't put verify as local, right? So that's something that we shouldn't do in the HTML, but basically you can see that's basically what you get with the local models. Sometimes they might not be as accurate as, you know, the Claude Opus models or, you know, Sonic models, right? So that's a trade-off that you have to be aware when you're trying to switch to the local models.

So then it got me thinking, well, what are some models in Ollama here has the highest accuracy using Claude Code? So here you can see we have the answer for Deepseek Rail Qwen 3 which has like 30B parameters which I don't think is good enough because this 30B parameters which is really high for VRAM, but here we have our GLM 4.7. So I thought to myself, well why not give it a try, right?

Connecting and Using the Cloud-Based GLM-4 Model

So I'm just going to copy that and simply just run this command on my local machine so it can be able to communicates to the GLM. But I don't think this model lives in my local machine. I think it communicates to a third party cloud providers that has the GLM 4.7. So I basically run that command and try to connect it on my local machine. So, I'm just going to click on connect. And once my device is connected, so now what I'm going to do here is I'm just going to do ollama launch claude with config. And I'm just going to choose the 4.7 cloud right here. And I'm going to say yes, launch with Claude Code. And here you can see we have our GLM 4.7 in the cloud. Right. So I'm just going to say hi. And you can see that we have a much faster response from the GLM4.7 cloud here. In this case, I'm just going to trigger the front end design skill and basically do the same thing for email verification by generating the terminal wireframe before it's going to renovate the email. So, I'm just going to run this and let's take a look at what it does here.

Comparing GLM-4's Redesign and Iterating on It

And just to confirm, yes, based on your documentations, the GLM4.7 cloud does run in the cloud, not on your local machine. So, it doesn't really consumes a lot of VRAM power. It kind of defeats the purpose of like having it to run your local machine, right? But if you want to get a higher results, you kind of have to go through this way. Okay.

So then eventually you can see that it generates the answer. So now after it has a full picture of the brand and the current email templates here is what it creates for the minimalistic premium design. And here you can see I'm not sure what this is. Maybe like a logo or something. But then we have our verify email. One tap to activate your accounts expire in 24 hours. So it's asking us if we wanted to implement it. I'm just going to tap this. Say yes. Let's do this. And I'm really curious to see how the GO 4.7 here is going to generate for this one. So, highly recommend if you can't find the right local models here for the answer that you want to generate, maybe try something with the GLM cloud here. So, we can see that right now it's going to ask us permissions if we want to generate this. So, I'm just going to say yes, allow all edits during the sessions. So, it's going to make those changes on my behalf. So, in this case, let's take a look after it has successfully generated this.

All right. So, now you can see this is what it looks like after it has successfully generate everything. Now obviously the first tries aren't perfect but at least you can see that the style here, the structure, if we were to look at the password resets, right, the welcome page is actually looking at the same structure other than the background has changed, but you can see that yes we have a logo still in the middle, right, still at the top, and also we have the email footers, right, where the email footers are still the same. It does analyze the current structure and just basically add it on top of the current structure that's all. All right, so right here you can see other than the emoji, right, which here you can see we have our email emoji here or email icon which is shifted to the left, but literally you can use that to basically ask it to modify it. Overall, you can see it's much more cleaner designs compared to what we have in the last one, right?

Refining the Design and Discussing GLM-4's Free Tier

So, I can literally just take a picture of this, right? Take a screenshot and say things like, "Uh, based on the current design and I would like to, you know, have the background to be consistent throughout all the templates, right? Um so, make sure to, you know, revert the background color change for the email. And also, the other thing is that if you take a look at this image, there's the email icon is shifted to left. Can you make it in the center, please?" And that's basically my prompt. So, in this case, I'm just going to submit this and basically going to let GLM4.7 cloud tier to fix this.

But then you might be wondering, is GLM4.7 here free? Uh, yes, there is a free tier and a pay tier, right? So, the free tier designed to have a light use and you roughly get 250,000 input tokens per hour before the usage is paused. So, usually 250,000 input tokens per hour is probably just one conversation context, right? If you take a look at the context window here, usually one context session inside of Claude Code is usually 200,000 for the contact window. So, it should be enough.

And of course, in my next video, I'll make sure to do a full review of the GLM latest model and basically see how we can be able to integrate our Claude with GLM, but also still uses all the features that Claude Code offers. So, if you're interested to see a full review of GLM, the latest model like GLM5 in my Claude Code sessions to see if it's capable of doing a lot of those daily tasks like building applications or you know adding further features onto it or using all those Claude skills or other functionality that Claude offers then make sure to comment down below and I'll make sure to try to make a video out of it and show you guys how can you use GO inside of Claude Code without any limitations using the tiers that it offers here. Okay, so make sure to comment below, like this video and I'll make sure to make that in the upcoming videos. But let's continue through the video. GLM4.7 cloud to see what the result look like.

Final Results and Conclusion

All right, so back to the fix here. So now both fix has successfully applied. And now let's say if I were to come back to the email verification, you can see that we do have everything. So I told it to basically revert back to the previous background which is now gray, right? Same color as what we have here. But at least you can see that the icon here is now centered and the overall the verification email here is much more minimal on exactly what user had to do. So pretty much that's how we can be able to use the Ollama model here to get a high results using Claude Code completely for free.

Okay, so pretty much that's all for this video and of course in my next video I'll make sure to compare all the models inside of GLM to see if it's really come up to par with the Claude Opus 4.5 inside of Claude Code and try to see if it's able to use all the functionality that Claude Code offers and see if it delivers a similar value compared to Claude Opus 4.5. So make sure to like my this video and comment down below if you're interested and I'll make sure to plan that in the future video. All right, so pretty much that's it for this video. And if you do find out in this video, please make sure to like this video, consider subscribe for more content like this. But with that being said, I'll see you in the next video.