DEV Community

Manendra Verma
Manendra Verma

Posted on

How to Run ClawBot Locally Using LM Studio (Free Setup Guide)

#openclaw #ai #agents #setupguide

This article explains how you can set up ClawBot locally using LM Studio and run it on your own machine without relying on paid APIs.

By connecting ClawBot to LM Studio, you can use local AI models to test and perform basic development tasks while keeping everything private and under your control.

Since everything runs on your own machine, you get full control and privacy over your data. I tested this setup myself and it works well for experimentation and learning.

Quick Reminder:

  • This setup is mainly for testing and basic tasks.
  • Running AI models locally requires a decent system with good CPU or GPU performance.
  • You should have at least 16 GB RAM (32 GB recommended for smoother performance).
  • For complex tasks, a system with strong GPU or higher CPU power will work much better.
  • The setup is completely free, but remember that the model runs on your own system resources.

Before starting, it’s important to understand the limitations and requirements of running everything locally.

Minimum Hardware (Basic Setup)

This will run 7B models like Qwen2.5-Coder-7B-Instruct.

  • CPU: modern 4–6 core CPU (Intel i5 / Ryzen 5 or better)
  • RAM: 16 GB recommended
  • Storage: 10–20 GB free space
  • GPU: optional (can run on CPU)

Performance:

  • speed: ~5–10 tokens/sec on CPU
  • usage: good for chat and testing agents

Recommended Hardware (Smooth Experience)

For faster inference and multiple tools.

  • CPU: 8+ core processor (Ryzen 7 / i7)
  • RAM: 32 GB
  • GPU: 8–12 GB VRAM (RTX 3060 / 4060 / 4070)
  • Storage: SSD required

Performance:

  • speed: 20–60 tokens/sec
  • latency: much lower
  • agent tasks: smoother

Step By Step Guide

Phase-1 — Install LM Studio

LM Studio runs the local model.

  1. Go to the official website

    https://lmstudio.ai

  2. Download LM Studio for Windows

  3. Install normally.

  4. Open LM Studio

Phase-2 — Download a model

Inside LM Studio:

Left sidebar -> Model Search (4th icon)
Enter fullscreen mode Exit fullscreen mode

Example search:

Qwen2.5-Coder-7B-Instruct
Enter fullscreen mode Exit fullscreen mode

NOTE: Just above the download button you can select the Quantization of the model

Step-2.1: How to select the Right Model

Selecting the best model in LM Studio depends on several technical factors: hardware capacity, task type, model size, and quantization.

1. Check Your Hardware First:

Your RAM and GPU VRAM determine what models can run efficiently.

Typical guidelines:

Hardware Recommended Model Size
8 GB RAM 3B–4B models
16 GB RAM 7B–8B models
32 GB RAM 13B models
64 GB RAM 30B+ models

Meaning of the “B” 

B = Billion
Enter fullscreen mode Exit fullscreen mode

Examples:

7B  = 7 Billion parameters
6.7B = 6.7 Billion parameters
34B = 34 Billion parameters
70B = 70 Billion parameters
Enter fullscreen mode Exit fullscreen mode

So a 7B model has about 7,000,000,000 parameters.

Example Model Sizes
Small models
1B – 3B
Fast but weaker reasoning

Medium models
7B – 8B
Good balance of speed and intelligence

Large models
13B – 34B
Much smarter but need strong hardware

Very large models
70B+
Need powerful GPUs or servers
Enter fullscreen mode Exit fullscreen mode

2. Choose Model Based on Task

Different models are trained for different tasks.

Coding

Best model types:

Qwen2.5-Coder
DeepSeek-Coder
CodeLlama
Enter fullscreen mode Exit fullscreen mode

Content Writing / Chat 

Llama 3.1
Mistral
Qwen2.5
Enter fullscreen mode Exit fullscreen mode

Reasoning / Agents 

DeepSeek-R1 Distill
Qwen2.5
Ministral
Enter fullscreen mode Exit fullscreen mode

3. Pick the Right Quantization

Quantization controls memory usage vs quality.

Common options:

Quantization Quality RAM Usage
Q2 low very small
Q3 medium-low small
Q4_K_M good moderate
Q5_K_M very good higher
Q6/Q8 near original large

Best default choice:

Q4_K_M
Enter fullscreen mode Exit fullscreen mode
  • quality: very good
  • speed: fast
  • memory: manageable

4. Choose Trusted Model Publishers

The same model can appear from multiple uploaders.

Prefer models published by:

Llama family
Qwen family
Mistral family
DeepSeek family
Enter fullscreen mode Exit fullscreen mode

Phase-3 — Load the Model

In the server panel click:

Load Model
Enter fullscreen mode Exit fullscreen mode

Select your downloaded model.

Once loaded you should see it listed under:

Loaded Models
Enter fullscreen mode Exit fullscreen mode

NOTE: Must enable (Manually choose model load parameters)

Put context length low as possible because it will affect response time.

  • Low context length → Fast response but loose accuracy
  • high context length → Slow response but high accuracy
  • GPU offloading decides how many of those layers run on the GPU instead of the CPU.

If your system has no GPU, set:

GPU Offload = 0
Enter fullscreen mode Exit fullscreen mode

If GPU offload = 0 layers: CPU does all calculations.

LLMs are built with layers (like 32–40 layers depending on the model).

Example:

Total layers in model = 32
Enter fullscreen mode Exit fullscreen mode

If you set:

GPU Offload = 10
Enter fullscreen mode Exit fullscreen mode

Then:

10 layers → GPU
22 layers → CPU
Enter fullscreen mode Exit fullscreen mode

So the workload is shared.

Phase-4 — Start LM Studio Local API Server

1. Open LM Studio

Launch LM Studio on your computer.

2. Open Developer Settings

Look at the left sidebar: usually ****(third icon from the top)

In your sidebar click the icon that looks like:

>_
Enter fullscreen mode Exit fullscreen mode

This is the Developer / API panel.

3. Open the Local Server Panel

first option in in Developer sidebar

4. Keep Server Setting like this

5. Start the Server

Click:

toggle the button before server setting
Enter fullscreen mode Exit fullscreen mode

You should see:

Server running at URL comes next to it check it.
Enter fullscreen mode Exit fullscreen mode

Open browser:

http://localhost:1234/v1/models
or
http://172.25.64.1:1234/v1/models
Enter fullscreen mode Exit fullscreen mode

If you see JSON with your model → LM Studio is working.

Example response:

{
  "data": [
    {
      "id": "qwen2.5-coder-7b-instruct",
      "object": "model",
      "owned_by": "organization_owner"
    },
    {
      "id": "text-embedding-nomic-embed-text-v1.5",
      "object": "model",
      "owned_by": "organization_owner"
    }
  ],
  "object": "list"
}
Enter fullscreen mode Exit fullscreen mode

Phase-5 — Install OpenClaw

OpenClaw requires Node.js 22 as a prerequisite. Don't worry—installation is automated.

Universal Installation (All Platforms):

Open your terminal/command prompt and run this single command:

curl -fsSL https://openclaw.ai/install.sh | bash
Enter fullscreen mode Exit fullscreen mode

If that doesn't work, or you're on Windows PowerShell, use:

iwr -useb https://openclaw.ai/install.ps1 | iex
Enter fullscreen mode Exit fullscreen mode

Open PowerShell as Administrator:

  • If install as admin then it will automatically configure the gateway.
  • If it asks anywhere Remote or Local, then set it as local.

Phase-6 — Configure OpenClaw to Use LM Studio

After running make sure you select like these:

Other things you can select skip for now option, or you can setup later any of those you want to. I’ll setup later.

You get base URL and LLM model from the right sidebar of LM studio in info tab. Use yours not copy paste it. you can only use the API KEY as same as mine.

You will asked about following field: (here what I select)

Default model: Keep current (vllm/qwen2.5-coder-7b-instruct)
Channel status: skip for now
Search provider: Skip for now
Configure skills now?: No
Enable hooks? No
Optional apps: skip for now
Control UI: Web UI
Enter fullscreen mode Exit fullscreen mode

Phase-7 — Run OpenClaw

After completion of above command you get a URL like this:

http://127.0.0.1:18789/#token=YOUR-Token
Enter fullscreen mode Exit fullscreen mode

Now copy the URL and paste it any browser search bar. You get the chat interface like this

Now you can use the claw bot here.

Or, you can run using command:

openclaw gateway
Enter fullscreen mode Exit fullscreen mode

Phase-8 — Now if you want to remove clawbot from the system entirely

For security reasons if you don't want to use claw bot follow following steps to remove clawbot from the system:

Step-1: Stop the Running Server

If OpenClaw is running in the terminal:

Ctrl + C
Enter fullscreen mode Exit fullscreen mode

Step-2: Delete the OpenClaw Project Folder

Go to the following path:

C:\Users\youname
Enter fullscreen mode Exit fullscreen mode

Delete the .openclaw folder.

Now, Work is not done yet. You have to delete packages too.

Step-3: Remove Global Packages

Check global packages:

npm list -g --depth=0
Enter fullscreen mode Exit fullscreen mode

If you see something like:

openclaw
Enter fullscreen mode Exit fullscreen mode

remove it:

npm uninstall -g openclaw
Enter fullscreen mode Exit fullscreen mode

Usually, this step is not required because OpenClaw runs locally.

Step-4: Remove Local Models

If you downloaded models for LM Studio, they are stored separately.

Location (Windows):

C:\Users\<your-user>\.lmstudio\models
Enter fullscreen mode Exit fullscreen mode

Delete models if you want to free space.

Step-5: Clear Node Cache (Optional)

This step is optional but removes leftover npm cache.

npm cache clean --force
Enter fullscreen mode Exit fullscreen mode

Reminder:

  • It is only for testing and perform basic tasks.
  • You need large system which has more GPU or CPU power to perform complex task.
  • This is entirely free so you can use it but keep in mind that it is using your system resources to run.

----- 📄 Document Ends Here. Thanks for Reading!!! 🎉 -----

Top comments (0)