I tried building a local AI coding environment with Ollama and OpenCode
This page has been translated by machine translation. View original
Companies are currently competing fiercely to improve AI accuracy. As a result, the operational costs of AI systems are trending upward, and I believe there will likely be a shift from flat-rate plans to pay-as-you-go billing in the future.
I currently use Claude Pro for personal use, and signs of cost pressure are already beginning to emerge. In March 2026, the rate at which session limits are consumed during peak hours (weekdays PT 5:00–11:00) was increased, and in addition to the per-5-hour session limit, a cap was also placed on weekly computational resource usage. It cannot be denied that flat-rate plan restrictions may tighten further, or that a shift to pay-as-you-go pricing may occur. Rather than scrambling to find an alternative when that happens, I decided to try out local LLMs.
This article introduces the steps to build a fully local AI coding environment by installing Ollama as the local LLM runtime and OpenCode as the AI coding agent.
Test Environment
- MacBook Pro (M1 Pro / 32GB)
- macOS 26.4 (25E246)
- Ollama 0.20.5
- OpenCode 1.4.3
- gemma4:26b (running via Ollama)
What is Ollama
Ollama is an open-source tool for running LLMs in a local environment. It allows you to download and run various open-source models such as Meta Llama and Google Gemma with a single command.
What is OpenCode
OpenCode is an open-source, terminal-based AI coding agent. Similar to Claude Code and GitHub Copilot CLI, it creates and edits code in the terminal based on natural language instructions.
A key feature of OpenCode is its support for over 75 LLM providers. In addition to cloud-based models from OpenAI, Anthropic, and Google, it can also use locally running models via Ollama. This means you can run an AI coding agent entirely locally without relying on external APIs.
Step 1: Install Ollama
First, install Ollama. It can also be installed via Homebrew, but the official GitHub README.md introduces an installation method using the officially recommended install script, so I will use that approach.
curl -fsSL https://ollama.com/install.sh | sh
Once the installation is complete, start Ollama.
ollama
The Ollama menu will appear in the terminal.
Selecting Chat with a model allows you to choose from several models. Since I want to try a local LLM this time, I select gemma4.
Once the download is complete, a chat session begins immediately. If a response is returned correctly, the Ollama installation is successful.
Type /exit to exit Ollama for now.
Step 2: Download the Model
In the setup in Step 1, the model was downloaded through an interactive selection at first launch, but for the model to be used in OpenCode, the parameter size must be explicitly specified during download.
This time I will use gemma4:26b. gemma4 is an open model published by Google based on Gemini, and is a new model released in April 2026. As I will explain later, it was chosen because of its fast response speed and comfortable usability even when running locally.
ollama pull gemma4:26b
Once the download is complete, verify that it works.
ollama run gemma4:26b "こんにちは。あなたはどのモデルですか？"
I also tried qwen3.5:27b, but in my environment the response felt slow and not very convenient. In contrast, gemma4:26b has fast responses and seems to perform reasonably correct inference. Performance in this area will vary depending on the amount of available memory.
qwen3.5 appears to be highly regarded for its coding performance, so I plan to try it again at a later date.
Step 3: Install OpenCode
Next, install OpenCode.
curl -fsSL https://opencode.ai/install | bash
Once the installation is complete, verify the version.
opencode --version
1.4.3
Step 4: Configure OpenCode
To use Ollama models with OpenCode, a configuration file must be created. Write the following content in the project root directory or as a global configuration at ~/.config/opencode/opencode.json.
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"ollama": {
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "http://localhost:11434/v1"
},
"models": {
"gemma4:26b": {
"name": "Gemma 4 26B"
}
}
}
}
}
The key configuration points are as follows.
- For
npm, specify the AI SDK provider package used internally by OpenCode. For Ollama integration, use@ai-sdk/openai-compatible. This package is automatically installed by OpenCode, so manual installation is not required. - For
baseURL, specify Ollama's OpenAI-compatible endpoint (http://localhost:11434/v1). Note that/v1is used rather than Ollama's native endpoint (/api). - For
models, specify the name of the model already downloaded in Ollama.
Ollama's default context window varies depending on available VRAM. In this test environment (32GB unified memory), the default was 32K tokens. If you want to change this, it can be adjusted using Ollama's num_ctx parameter.
Step 5: Verify Operation
After confirming that Ollama is running, launch OpenCode.
opencode
After launching, type /models and if Ollama's gemma4:26b appears in the list of options, the configuration has been applied correctly.
I entered a sample prompt and confirmed that a response was returned from the local model.
Summary
I introduced the steps to build a locally running AI coding environment by combining Ollama and OpenCode.
Local LLMs are less accurate compared to cloud-based models, but they do not require an internet connection and incur no API usage fees. I consider them a sufficient option for casual code generation in personal development or for use in projects where privacy is important.
The gemma4:26b used this time was not as useful for coding purposes as I had hoped, but I find it sufficiently practical for general question answering and research. I would like to make use of it for everyday use going forward. I also plan to try other models, such as qwen3.5:27b, which is highly regarded for its coding performance.
Preparing a local LLM environment in anticipation of future price changes in cloud services will not be a wasted effort. I hope this serves as a helpful reference for others considering the adoption of local LLMs.
References
Job Openings: Classmethod is Hiring iOS Engineers
The Starbucks Digital Technology division is looking for engineers capable of iOS app development. We welcome applications from those who would like to join us and share ideas about new Xcode and iOS features in channels like misc-ios!
We are also hiring iOS/Android engineers in other areas. Let's talk about mobile app development together!