opencode

opencode reads its providers from a JSON config. Adding LuluTokens means declaring one custom provider that uses the OpenAI-compatible SDK and lists the models you want in the picker.

Configure

Pick a config file

Path	Scope
`~/.config/opencode/opencode.json`	All projects
`opencode.json` in the repo root	This project only (commit it, minus secrets)

Declare the provider

opencode.json

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "lulutokens": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LuluTokens",
      "options": {
        "baseURL": "https://api.lulutokens.ai/v1",
        "apiKey": "{env:LULUTOKENS_API_KEY}"
      },
      "models": {
        "deepseek-v4-pro": { "name": "DeepSeek V4 Pro" },
        "qwen3.7-max": { "name": "Qwen3.7 Max" }
      }
    }
  },
  "model": "lulutokens/deepseek-v4-pro"
}

The {env:...} form keeps the key out of the file, so a project-level opencode.json is safe to commit.

Add one entry per model you want in the picker. The key of each entry must match the model ID exactly as it appears in GET /v1/models; name is only the label.

Start opencode

export LULUTOKENS_API_KEY="sk-..."
opencode

Use /models inside opencode to confirm the LuluTokens models are listed, and switch between them.

Storing the key in opencode instead

If you would rather not manage the environment variable, drop the apiKey line and let opencode hold the credential:

opencode auth login

Choose Other, enter lulutokens as the provider ID, and paste the key.

Per-model overrides

Anything the model needs beyond defaults goes on its entry — for example a smaller context window for a model that is documented as supporting less than opencode assumes:

"models": {
  "deepseek-v4-pro": {
    "name": "DeepSeek V4 Pro",
    "limit": { "context": 64000, "output": 8000 }
  }
}

Troubleshooting

Symptom	Fix
Provider missing from `/models`	JSON syntax error — run `opencode` from a terminal and read the startup warning
`401`	`LULUTOKENS_API_KEY` is not set in the shell that launched opencode
`404`	`baseURL` is missing `/v1`
Model answers but tools never run	The model has no tool-calling support; pick one that does