Run Your Own Coding Agent Locally (For Free)

In this guide, you’ll learn how to download and run a coding LLM directly on your laptop using Ollama, and connect it to Claude Code or Codex with a single command. The result: a fully functional coding agent that runs locally, costs nothing per token, and keeps your code private.

Who This Is For

• Solo developers and indie founders are spending heavily on Claude Code Max or Codex Pro for tasks that don’t require top-tier reasoning
• Students, hobbyists, and beginner programmers who want a hands-on agentic setup without paying monthly fees
• Anyone working with sensitive or proprietary code who prefers everything to stay on their own machine

What You’ll Build

By the end, you’ll have a local coding agent (Claude Code, Codex, or OpenCode) connected to a free Ollama model running on your hardware. You’ll use the same interface and workflow, just without API costs or external dependencies.

Requirements

• Ollama installed (follow the install guide if needed)
• Claude Code set up, or OpenCode as an alternative
• A terminal and an active project folder

Hardware:

• 8 GB RAM (minimum, for smaller models)
• 16 GB RAM (recommended)
• 32 GB+ or GPU (ideal for larger models)

Step 1: Check Your Hardware and Choose the Right Model

Before downloading anything, understand what your machine can handle.

• On Mac: Apple menu and go to About This Mac
• On Windows: Settings, and then System, and lastly About

Take a screenshot of your specs and ask an LLM:

“Which Ollama coding models can I run on this machine?”

It will suggest suitable models based on your RAM and processor.

General guideline:

RAM	Model Size	Example
8 GB	~3B params	qwen3-coder:3b
12 GB	4–7B params	gemma4:e2b
16 GB	7–12B params	qwen3-coder:7b
32 GB+ / GPU	20B+	gpt-oss:20b

Pro tip: Use the largest model your system can comfortably handle. Bigger models are more reliable when executing tool-based coding tasks.

Step 2: Choose a Model That Supports Agent Workflows

Go to Ollama’s model directory and look for coding models.

Before selecting one, check:

• The Applications section on the model page
• Ensure it supports tools like Claude Code, Codex, or OpenCode

If it doesn’t, skip it, agentic workflows won’t function properly.

Top-performing options:

• qwen3-coder: strong code generation for its size
• gemma4: better at multi-step reasoning and tool usage
• gpt-oss: powerful open-weights model with solid agent support

Pro tip: If unsure, download two models and compare. Storage is cheap; performance matters more.

Step 3: Launch Your Local Coding Agent

From the model page, copy the launch command. Example:

Then,

1. Open your terminal inside your project folder
2. Paste and run the command
3. Confirm the model download

Once complete, you’ll be inside a Claude Code session powered by your local model.

Run /model to confirm which model is active. From here, every response is free.

Pro tip: Use ollama ps in another terminal to monitor performance (RAM, GPU usage, active model).

Step 4: Increase the Context Window

This is a critical step most people miss.

By default, Ollama sets a 4K context window, which is too small for coding agents. It causes the model to forget earlier parts of your session.

To fix it:

• Open the Ollama app
• Go to Settings
• Increase context to 32K or higher (depending on your system)

Pro tip: Don’t max it out blindly. Ask an LLM what your hardware can safely support to avoid crashes.

Step 5: Try OpenCode for a Simpler Setup

If Claude Code feels too complex, switch to OpenCode, a lighter alternative designed for smaller models.

Install on Mac:

curl -fsSL https://opencode.ai/install | bash

Run it with:

ollama launch opencode --model gemma4:e4b

Pro tip: Smaller models perform better when they “think less.” Use plan mode (Shift + Tab) or reduce reasoning settings if tasks feel overcomplicated.

Going Further

Once everything is working, you can extend your setup:

• Use a hybrid workflow: Let Claude Code or Codex handle planning, and your local model handle repetitive tasks
• Control your agent remotely: Use Claude remote control and connect via mobile
• Run on a dedicated machine: Set up an old PC or Mac mini as a local inference server

Final Takeaway

Local coding models aren’t yet a complete replacement for cloud-based agents, but they’re incredibly useful for:

• Practicing agent workflows
• Protecting sensitive code
• Reducing costs on routine tasks

And with rapid improvements, the gap between local and cloud models continues to shrink.