How to use Ollama with OpenCode (local models, 2026)

OpenCode is the open-source (MIT) terminal coding agent. At about 210,000 GitHub stars on 25 Sep 2026, it's the most-starred of the three I cover in this cluster. It supports local models natively, and Ollama now ships a first-party integration for it. Of the three agents, it's the one I'd point a newcomer to local models at first. Here's why, and how to set it up without the usual "tool calls don't work" dead end. The full comparison with Claude Code and Codex is in our three-way guide.

Step 1: install OpenCode

# macOS / Linux (Ollama's docs)
curl -fsSL https://opencode.ai/install | bash

# Windows (Ollama's docs)
npm install -g opencode-ai

Step 2: fix the context before you start

This step is the one that decides whether OpenCode works. Ollama's page is unambiguous: "OpenCode requires a context length of 64k or higher." Ollama's context-length page lists the defaults: 4k under 24 GiB of VRAM, 32k at 24–48 GiB, 256k at 48 GiB+. The Ollama FAQ still says a flat 4,096. On a typical 8–16GB gaming card you start at 4k, which is 1/16th of what OpenCode needs.

OpenCode's own provider docs describe the symptom: "If tool calls aren't working, try increasing num_ctx in Ollama. Start around 16k - 32k." Their 16–32k starting point is lower than Ollama's 64k requirement. I'd go with Ollama's number, since coding agents pull file contents into the context as they work.

OLLAMA_CONTEXT_LENGTH=64000 ollama serve

# check it took: CONTEXT should be 64000, PROCESSOR 100% GPU
ollama ps

Step 3 (fast path): ollama launch opencode

ollama launch opencode            # pick a model, start OpenCode
ollama launch opencode --config   # configure without starting a session

Ollama's docs are specific about what this touches. ollama launch opencode starts OpenCode with an inline config for the selected model and "does not overwrite ~/.config/opencode/opencode.json". Your existing OpenCode settings still apply. The flip side is that models you've defined only in opencode.json won't show up in the launch model picker.

Step 3 (manual path): opencode.json

If you want the setup committed with your project, add an Ollama provider to opencode.json. This is Ollama's documented block, with the model changed to a coding model:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ollama",
      "options": {
        "baseURL": "http://localhost:11434/v1"
      },
      "models": {
        "qwen3-coder:30b": {
          "name": "qwen3-coder:30b"
        }
      }
    }
  }
}

Then just run opencode. When a global and a project config disagree, OpenCode's config precedence rules decide which one wins. It's worth a read if a setting seems to be ignored.

Which model fits at 64k

OpenCode needs a model with tool calling. Ollama links its tools-capable list. For coding on a home GPU, the Ollama library has Qwen3-Coder 30B (19 GB download, 256K max context) and gpt-oss 20B (14 GB, 128K). At the 64k OpenCode requires, our VRAM engine gives:

The q8_0 column is Ollama's documented OLLAMA_KV_CACHE_TYPE=q8_0 setting, which the FAQ says uses "approximately 1/2 the memory of f16" with a very small precision loss. It's a global setting, so it applies to every model you run. Per-GPU tables: best GPU for Qwen3-Coder, best GPU for gpt-oss 20B. The hardware deep-dive is best GPU for Qwen3-Coder.

Weighing it against the others? Here's the same setup for Claude Code and Codex CLI. If you'd rather work inside VS Code, see running Qwen3-Coder in VS Code.