opencode setup with ollama

Docs: https://opencode.ai/docs/config/

Setup OLLAMA_DATA in .env to wherever you need to point it to, then using this repo's docker-compose.yml simply run

docker compose up -d

To download a model, such as qwen3:8b, and get it ready to run on ollama, run the following

docker exec -it ollama ollama pull qwen3:8b

# Or if using docker compose
docker compose exec ollama ollama pull qwen3:8b

This following json in ~/.config/opencode/config.json has worked for me so far I, like many others have had issues getting the tools working in opencode using local models.

I found out why agentic actions such as using opencode's provided tools didn't work. It's because Ollama has set the context window at 4096 for the models.

Even though ollama says the context is larger, for example for qwen3 it says context is around 40k, upon running the model via ollama, it will use a default of 4k context, this has to be setup by yourself. Also make sure the model you are planning to use actually supports agentic tools.

This was fixed by doing the following:

For example, after pulling qwen3:8b

$ docker compose exec ollama bash
$ ollama run qwen3:8b
>>> /set parameter num_ctx 16384
Set parameter 'num_ctx' to '16384'
>>> /save qwen3:8b-16k
Created new model 'qwen3:8b-16k'
>>> /bye

Then using the following config.json

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://127.0.0.1:11434/v1",
        "apiKey": "ollama"
      },
      "models": {
        "qwen3:8b-16k": {
          "tools": true
        }
      }
    }
  }
}

First, to prove the magician's sleeves are empty

$ ls -la
total 8
drwxrwxr-x 2 lemonish lemonish 4096 Jul  1 17:21 .
drwxrwxr-x 3 lemonish lemonish 4096 Jul  1 17:21 ..
$ cat todo.md
cat: todo.md: No such file or directory

Run the command

$ opencode run "/no_think generate a todo.md file with the contents 'hello world, from qwen3' in it" --model ollama/qwen3:8b-16k

█▀▀█ █▀▀█ █▀▀ █▀▀▄ █▀▀ █▀▀█ █▀▀▄ █▀▀
█░░█ █░░█ █▀▀ █░░█ █░░ █░░█ █░░█ █▀▀
▀▀▀▀ █▀▀▀ ▀▀▀ ▀  ▀ ▀▀▀ ▀▀▀▀ ▀▀▀  ▀▀▀

>  /no_think generate a todo.md file with the contents 'hello world, from qwen3' in it

@  ollama/qwen3:8b-16k

|  Write    ollama/test/todo.md

<think>
</think>

<think>
</think>

</think>
</think>
</think>

Okay, I'll generate a todo.md file with the specified content. Let me create it for you

$ ls
todo.md
$ cat todo.md
hello world, from qwen3

ta-da!

Troubleshooting & Common Hanging Issues on Windows

If your OpenCode integration with Ollama hangs or doesn't get any responses back, this is usually caused by one of the following reasons:

1. Localhost resolving to IPv6 (::1) instead of IPv4 (127.0.0.1)

By default, Ollama on Windows binds only to the IPv4 loopback address (127.0.0.1:11434). However, on Windows, localhost often resolves to the IPv6 loopback address (::1).

• The Issue: When OpenCode attempts to connect to http://localhost:11434/v1, it tries the IPv6 address first. If your firewall or system network configuration silently drops these packets instead of immediately rejecting them, the connection will hang until it times out.
• The Fix: Change the "baseURL" in your config file to explicitly use http://127.0.0.1:11434/v1 instead of localhost.

2. Missing apiKey in the OpenAI-compatible configuration

The @ai-sdk/openai-compatible provider used in OpenCode expects an API key by default.

• The Issue: If "apiKey" is not specified in the configuration and you do not have an OPENAI_API_KEY environment variable set, the SDK initialization throws a missing API key error internally. Since the OpenCode UI may not capture or display this promise rejection, it will result in a silent hang.
• The Fix: Add "apiKey": "ollama" (or any dummy placeholder string) to the "options" object in your config.json file.

3. WSL2 / Docker Network Bridging

If you are running OpenCode in WSL2 or a Docker container while running Ollama natively on Windows:

• Make sure you set the environment variable OLLAMA_HOST to 0.0.0.0 on Windows so it listens on all interfaces (remember to exit Ollama from the system tray and restart it).
• Add a Windows Firewall rule to allow port 11434 incoming traffic.
• Use http://host.docker.internal:11434/v1 or the Windows host's IP address instead of localhost in your config.