Use opencode with a Local AMD LLM (ROCm): Config & MCP Tools

opencode can run entirely on your local AMD GPU. Here's a copy-paste opencode.json that wires it to your llama-swap model and the stack's MCP tools.

opencode is a terminal AI coding agent that speaks the OpenAI-compatible API — so it can drive your local model on an AMD GPU instead of a paid cloud API. Point it at llama-swap, add the stack's MCP tool servers, and your coding agent runs entirely on your own hardware.

Drop-in opencode config

Save this as ~/.config/opencode/opencode.json (global) or opencode.json in your project root:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "ai-stack": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "ai-stack (local AMD / ROCm)",
      "options": {
        "baseURL": "http://localhost:8080/v1",
        "apiKey": "sk-local"
      },
      "models": {
        "qwen3.8-27b-heretic": {
          "name": "Qwen3.8-27B (local, ROCm)",
          "limit": { "context": 131072, "output": 32768 }
        }
      }
    }
  },
  "mcp": {
    "web-search": { "type": "remote", "url": "http://localhost:8765/mcp" },
    "browser":    { "type": "remote", "url": "http://localhost:8931/mcp" }
  }
}

No cloud key needed. baseURL points at llama-swap's local OpenAI endpoint; apiKey is just a non-empty placeholder. Swap the model id for whatever your stack serves (curl localhost:8080/v1/models).

What each part does

provider.ai-stack — registers your local llama.cpp/llama-swap server as an OpenAI-compatible provider via @ai-sdk/openai-compatible.
options.baseURL — http://localhost:8080/v1, the llama-swap endpoint (ROCm under the hood).
models — the local model(s) and their context/output limits.
mcp — remote MCP tool servers from the stack (web search on 8765, headless browser on 8931). Add more, e.g. a sandboxed shell on http://localhost:8099/mcp.

Run it

# make sure the stack is up
systemctl --user start ai-stack.service
# then just run opencode in your project
opencode

Pick the ai-stack provider/model in opencode and your coding agent is now local, private, and CUDA-free. MCP tools you listed are available to the agent automatically.

Technical FAQ

Can opencode use a local LLM instead of a cloud API?

Yes. opencode supports any OpenAI-compatible endpoint. Point its provider baseURL at a local llama.cpp/llama-swap server (e.g. http://localhost:8080/v1) and it runs fully offline.

How do I connect opencode to a local model on an AMD GPU?

Add a provider using @ai-sdk/openai-compatible with baseURL set to your llama-swap endpoint, and list your model ids. The model runs on AMD ROCm; opencode doesn't care about the GPU vendor.

Does opencode need an API key for a local model?

No real key is needed, but the field must be non-empty — use a placeholder like sk-local. The request never leaves your machine.

How do I give opencode tools like web search or a browser?

Register MCP servers in the mcp section of opencode.json with type remote and the server URL (e.g. web search on http://localhost:8765/mcp, a headless browser on http://localhost:8931/mcp).

Which model should I use with opencode locally?

Any instruct/coding model your stack serves. A 27B at Q6_K runs well on a 32GB AMD card; a smaller coder model is faster for autocomplete. List whatever curl localhost:8080/v1/models returns.

Get the full stack on GitHub →