How to Run OpenCode With Local Models (Ollama)

Steps

1. 01

   Shortcut: let Ollama launch OpenCode for you

   Ollama's own integration page (checked 2026-10-01) documents a one-line path: ollama launch opencode starts OpenCode with an inline config for the Ollama model you select, so you can skip the config file entirely. The same page says OpenCode requires a context length of 64k or higher — pick a model that supports it. Use the manual steps below if you want the provider saved in opencode.json. We did not run this command ourselves (no Ollama on our test machine).

   ollama launch opencode

   Copy
2. 02

   Pull a coding model in Ollama

   Download a local model. A coding-tuned model like qwen2.5-coder works well; smaller tags (e.g. :7b) run on modest hardware. Confirm Ollama is serving on its default port 11434.

   ollama pull qwen2.5-coder
   ollama list            # confirm the model is present
   # Ollama serves an OpenAI-compatible API at http://localhost:11434/v1

   Copy
3. 03

   Add Ollama as a provider in opencode.json

   OpenCode reads config from ~/.config/opencode/opencode.json (global) or ./opencode.json (per-project). Add a provider that uses the OpenAI-compatible AI SDK package and points baseURL at your local Ollama endpoint. List each local model under "models" with the exact Ollama tag as the key.

   {
     "$schema": "https://opencode.ai/config.json",
     "provider": {
       "ollama": {
         "npm": "@ai-sdk/openai-compatible",
         "name": "Ollama (local)",
         "options": { "baseURL": "http://localhost:11434/v1" },
         "models": {
           "qwen2.5-coder": { "name": "Qwen2.5 Coder (local)" }
         }
       }
     }
   }

   Copy
4. 04

   Confirm OpenCode sees the local model

   List models for your new provider. Your local model appears as ollama/<tag> — that is the identifier you use to select it.

   opencode models ollama
   # → ollama/qwen2.5-coder

   Copy
5. 05

   Run a coding task against the local model

   Use the -m / --model flag in provider/model format for a one-shot run, or launch the TUI and pick the provider in the model selector. Everything stays on your machine — no cloud calls.

   # One-shot, headless
   opencode run -m ollama/qwen2.5-coder "Explain what this repo does"
   
   # Or open the TUI and choose Ollama (local) in the model picker
   opencode

   Copy

Troubleshooting

Your model does not appear in opencode models ollama

Check the baseURL includes the /v1 suffix (http://localhost:11434/v1) and that the "models" key exactly matches the Ollama tag from ollama list. Restart OpenCode after editing the config.

Connection refused / fetch failed

Ollama is not serving. Run ollama serve (or open the Ollama app) and confirm http://localhost:11434 responds. If you changed Ollama's port, update baseURL to match.

The model answers, but tool calls (file edits, shell commands) fail

This is usually the context window. OpenCode's provider docs say to try increasing num_ctx in Ollama, starting around 16k–32k, and Ollama's OpenCode page says OpenCode needs 64k or more (both checked 2026-10-01). Raise the context length for the model in Ollama, or choose a model that supports a longer context.

Responses are slow or the model runs out of memory

Local models are bound by your hardware. Use a smaller tag (e.g. qwen2.5-coder:7b), close other heavy apps, and prefer a quantized build. Large agentic tasks are slower locally than on a hosted model.