Using with coding tools

Many coding assistants let you point them at an OpenAI-compatible endpoint. The two settings are always the same:

Environment variables

Most tools read the standard OpenAI variables, which is the quickest thing to try first:

export OPENAI_API_KEY="dvt-..."
export OPENAI_BASE_URL="https://api.devotel.com/v1"

Some older tools read OPENAI_API_BASE instead — set both if you are unsure:

export OPENAI_API_BASE="https://api.devotel.com/v1"

Continue (VS Code / JetBrains)

~/.continue/config.json:

{
  "models": [
    {
      "title": "Kimi K3 (Devotel)",
      "provider": "openai",
      "model": "kimi-k3",
      "apiBase": "https://api.devotel.com/v1",
      "apiKey": "dvt-..."
    }
  ]
}

Aider

export OPENAI_API_KEY="dvt-..."
export OPENAI_API_BASE="https://api.devotel.com/v1"
 
aider --model openai/kimi-k3

Zed

In settings.json, under the assistant’s OpenAI-compatible provider:

{
  "language_models": {
    "openai": {
      "api_url": "https://api.devotel.com/v1",
      "available_models": [{ "name": "kimi-k3", "display_name": "Kimi K3 (Devotel)" }]
    }
  }
}

Put the key in the editor’s credential store rather than the settings file.

Anything else

If a tool asks for “OpenAI-compatible endpoint”, “custom base URL” or “self-hosted OpenAI”, give it the base URL and key above. Confirm the plumbing with a plain curl first — if this returns a completion, the endpoint is fine and any remaining problem is in the tool’s configuration:

curl https://api.devotel.com/v1/chat/completions \
  -H "Authorization: Bearer $DEVOTEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"kimi-k3","messages":[{"role":"user","content":"reply with OK"}]}'

Things that trip tools up

Reasoning output. Kimi K3 returns reasoning_content alongside content. Tools that only read content behave correctly. Tools that concatenate every string field may show you the model’s thinking as if it were code — check the tool’s raw-response handling if output looks odd.

Slow first token. Reasoning happens before any visible output, so a tool with an aggressive timeout may give up before the answer starts. Raise the request timeout if you see empty responses.

max_tokens defaults. Some tools set a low max_tokens by default. Because the cap includes reasoning tokens, a low value can consume the whole budget thinking and return nothing. If a tool returns empty completions, raise or remove its output cap — see Chat Completions.

Unsupported endpoints. A tool that wants embeddings (for local codebase indexing, say) will fail on that feature — embeddings are not implemented yet. Chat features still work.