Kilo Code — custom API setup and model selection

Kilo Code is a VS Code extension with an agent mode: it reads your project, edits files, and runs commands, showing changes right in the editor. Like its relatives Cline and Roo Code, it isn't tied to a single model vendor.

Setup

Connecting comes down to three fields in the extension settings: provider, base address, key.

Pick OpenAI Compatible (or the equivalent "compatible provider" entry), then set:

If you prefer the Anthropic format, choose the Anthropic provider and give the address without /v1:

That suffix is the source of most setup trouble. Anthropic clients append /v1/messages themselves; OpenAI clients expect the path to be in the address already. Mix them up and you get an unexplained 404.

The key comes from your Zukko↗ dashboard and works in both formats.

Which model to use

The current list:

curl https://api.zukko.pro/v1/models -H "Authorization: Bearer fetch_your_key"

Non-working models are dropped automatically, so you can choose without checking.

For editor work it pays to keep two profiles and switch: a flagship like claude-opus-5-5 for tasks that need figuring out, and glm-5.3 for mechanical edits. Three times apart in price ($0.30 against $0.10 per million), and for edits like renaming a variable across a project the cheaper model is often enough — test on your own tasks. Keep in mind that GLM has a 128K context.

Keeping spend in check

Agent mode in an editor is deceptive: the changes look small, while under the hood every step ships the whole conversation plus every file read so far. A fifteen-step task easily runs to a hundred thousand input tokens.

What helps.

Start each task in a clean session. The previous context costs money on every step and also misleads the model, which starts leaning on stale information.

Name files explicitly. If you know where to edit, say so. Otherwise the agent spends several steps searching.

Watch the per-model spend breakdown in your dashboard. Usually one task where the agent went in circles accounts for most of the bill.

Common errors

401 or "invalid api key". Truncated key, or a trailing space.

404. An extra or missing /v1.

The model replies in prose instead of editing. Check that agent mode is actually on rather than plain chat, and that the chosen model supports tool calling.

The reply cuts off mid-sentence. Raise the output token limit in the extension settings.