Run Kilo Code on Stav

Kilo Code is an open-source coding agent. It runs as a terminal CLI, a VS Code extension and a JetBrains plugin, all on one config file. It can send its model traffic to any OpenAI-compatible endpoint, so it runs on Stav with configuration only.

Point it at Stav and every turn is billed to your Stav credits, shows up in your request log, and follows your workspace policy. Pick a model that is not a Claude or OpenAI model and the work runs where that model's Inference Provider runs — for example in France.

Tested with the Kilo Code CLI 7.8.8 on macOS. The VS Code and JetBrains dialogs write the same config file but were not tested.

Before you start

• A Stav API key. See Authentication.
• Kilo Code installed. CLI: npm install -g @kilocode/cli. See the Kilo Code CLI docs.

Set it up

Put the key in your shell and add a stav provider to the global config file, ~/.config/kilo/kilo.jsonc:

export STAV_API_KEY="sk_stav_live_..."

{
  "$schema": "https://app.kilo.ai/config.json",
  "model": "stav/zai-org/glm-5-2",
  "provider": {
    "stav": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Stav",
      "options": {
        "apiKey": "{env:STAV_API_KEY}",
        "baseURL": "https://api.stav.ai/v1"
      },
      "models": {
        "zai-org/glm-5-2": {
          "name": "GLM-5.2",
          "limit": { "context": 1000000, "output": 16384 }
        },
        "auto": {
          "name": "Smart Router",
          "limit": { "context": 128000, "output": 8192 }
        }
      }
    }
  }
}

• Write the base URL with /v1. Kilo Code adds /chat/completions itself.
• {env:STAV_API_KEY} is only resolved in the global file. A project-level kilo.jsonc ignores it, on purpose, so a repository cannot read your key. Keep the provider in the global file.
• Kilo Code does not fetch Stav's model list for a custom provider. Every model you want in the picker needs its own entry under models, with its limit. The model page in the Stav catalogue shows the context window and max output.
• A model is addressed as stav/<model>, so stav/zai-org/glm-5-2 on the command line and in the picker.

In VS Code or JetBrains: Settings → Providers → Custom provider. Provider ID stav, Provider API OpenAI Compatible, Base URL https://api.stav.ai/v1, your key, then add the models. This writes the same provider.stav block.

Run a turn:

kilo run "Reply with the words: hello from stav"

Pick a model

Any chat model in the Stav catalogue works, by its full slug. Add it under models first.

A named model can be served by more than one Inference Provider, and they are not all in the EU. In our test one turn on zai-org/glm-5-2 was served from Finland or France by a different provider than the rest. To keep every request in Europe, set a sovereignty floor or a compute region for your workspace. See Keep traffic sovereign.

Switch models for one run with kilo run -m stav/auto "...", or in the TUI with /models.

What works

Tested on zai-org/glm-5-2:

A Claude model gets the request translated to Anthropic's Messages API. We tested one streamed turn on anthropic/claude-haiku-5-5.

Tips

• Set limit to the model's real numbers. Kilo Code uses limit.output as max_tokens on every request and limit.context to decide when to compact. The model page in the catalogue shows both.
• auto has no fixed limits. Give it a limit that fits the smallest model your Router may pick. Stav lowers max_tokens to the serving endpoint's cap and names it in X-Stav-Clamped-Params.
• Kilo Code sends content: null on assistant turns that only call tools. Some strict endpoints reject that. Stav accepts it, so the agent loop runs.
• Some traffic does not go through Stav. Kilo Code checks for updates and can send telemetry to Kilo. Only model traffic goes to api.stav.ai.
• The key never goes in a project file. Even {env:...} is ignored there. Use the global file or KILO_CONFIG.

Troubleshooting

For Stav's error codes, see Errors.

Check it worked

Open Your workspace → Monitor → Request log. Each turn is one row, with the model, Inference Provider, country and cost. A short task is usually 3–5 rows. See Request log and prompt capture.