Skip to main content

opencode

opencode accepts OpenAI-compatible providers, and LiteLLM is one, so it talks to the gateway directly — no intermediary, and the local model is selectable like any other.

Merging the provider​

pi-config/bridges/opencode.json is a fragment, not a whole config. Overwriting yours would drop everything else you have set up.

jq -s '.[0] * {provider: ((.[0].provider // {}) + .[1].provider)}' \
  ~/.config/opencode/opencode.json pi-config/bridges/opencode.json \
  > /tmp/opencode.json && mv /tmp/opencode.json ~/.config/opencode/opencode.json

It registers two providers:

{
  "provider": {
    "lmstack-h2": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "lmstack h2-amd (llama.cpp)",
      "options": {
        "baseURL": "http://127.0.0.1:4000/v1",
        "apiKey": "{env:LMSTACK_H2_KEY}"
      },
      "models": {
        "qwen2.5-coder-7b": {
          "name": "Qwen2.5 Coder 7B (lmstack)",
          "limit": { "context": 16384, "output": 4096 }
        }
      }
    }
  }
}

The key​

{env:LMSTACK_H2_KEY} is substituted from the environment, so the key stays out of the config file:

export LMSTACK_H2_KEY=$(grep ^LITELLM_MASTER_KEY ~/.lmstack/stack.env | cut -d= -f2)

For a remote host, read it over SSH rather than copying it into your shell rc.

Keep limit.context honest​

context and output must match what the model YAML declares. Setting them higher does not give you more context — it gives you requests the server truncates or rejects, and the failure surfaces as the model ignoring the start of your file.

Unlike the pi extensions, nothing validates this fragment against the catalog. make validate covers pi-config/extensions/; this file is a copy-paste starting point, so when you change a host's models, change it here too.

Setting a host up from opencode​

You cannot — the installer ships as a Claude Code plugin, and that is the only supported distribution. Run /lmstack:install from Claude Code once, then point opencode at the endpoint it produced using the fragment above.

The two halves are independent. Nothing about the resulting stack is specific to the agent that installed it: it is an OpenAI-compatible endpoint behind a master key, and opencode talks to it like any other provider.