Kilo Code custom OpenAI-compatible provider: VS Code dialog, kilo.json and limits

[apigoto]

I work on the APIGOTO team. APIGOTO is an independent OpenAI-compatible gateway, not affiliated with OpenAI, Anthropic or Google. The steps work for any OpenAI-compatible endpoint.

0. Check the endpoint

curl https://api.china-aiapi.com/v1/models -H "Authorization: Bearer $APIGOTO_API_KEY"

VS Code extension

1. Settings (gear) → Providers tab → scroll down → Custom provider.
2. Provider ID: apigoto. Display name: anything you like.
3. Provider API: OpenAI Compatible (Chat Completions).
4. Base URL: https://api.china-aiapi.com/v1. API key: your key.
5. Models: once the URL and key are valid, Kilo fetches /models. Pick from that list or type MODEL_ID yourself. Fetch models retries after a fix.
6. Submit.

The dialog doesn't set token limits. Those go in the config file:

CLI / kilo.json (~/.config/kilo/kilo.json or ./kilo.json)

{
  "provider": {
    "apigoto": {
      "npm": "@ai-sdk/openai-compatible",
      "env": ["APIGOTO_API_KEY"],
      "models": {
        "MODEL_ID": {
          "name": "My model",
          "limit": { "context": CONTEXT_TOKENS, "output": OUTPUT_TOKENS }
        }
      },
      "options": { "baseURL": "https://api.china-aiapi.com/v1" }
    }
  },
  "model": "apigoto/MODEL_ID"
}

1. npm: @ai-sdk/openai-compatible = Chat Completions. @ai-sdk/openai = Responses, @ai-sdk/anthropic = Messages.
2. env reads the key from an environment variable instead of the file.
3. limit: replace both placeholders with the model's real numbers. If omitted they default to 0, which limits context management.

Troubleshooting

• No model list: the endpoint must serve /models under the Base URL. Otherwise, enter IDs by hand.
• "Model Not Found": the model key doesn't exactly match the upstream ID.
• Long tasks cut off or compact strangely: limit is missing (0) or wrong.
• The upstream rejects tool-call turns: the docs offer options.kiloOpenAICompatibleToolCallContent: "empty-string" for endpoints that need "" instead of null content. Only use it if you hit that error.

Our gateway gives one key for several model families, with upstream failover (at most 3 retries) and per-request logs. Pay-as-you-go, BYOK supported: www.china-aiapi.com

Do you set limit per model by hand, or have you found a cleaner way?