> Connect Atlas Cloud as a native LLM provider for chat, streaming and tool calls

# Atlas Cloud Provider

[Atlas Cloud](https://www.atlascloud.ai/) hosts models from many vendors behind one OpenAI-compatible LLM API. Dewee ships it as the native `atlascloud` provider type, so you only paste an API key: the base URL and a default model are filled in for you.

\| Setting | Value |
\| --- | --- |
\| Provider type | `atlascloud` |
\| API base | `https://api.atlascloud.ai/v1` |
\| Authentication | API key, sent as a Bearer token |
\| Default model | `deepseek-ai/deepseek-v4.1-flash` |

## Connect it

* **Runtime Gateway Dashboard:** Providers → Add Provider → **Atlas Cloud**, then paste the API key.
* **Customer Control Plane:** open the workspace **Providers** page, add **Atlas Cloud**, paste the key, test the connection and pick a model.
* **CLI:**

```bash
printf '%s' "$ATLASCLOUD_API_KEY" | dewee providers create --type atlascloud --name atlas-cloud --api-key-stdin
dewee providers verify <provider-id> --model deepseek-ai/deepseek-v4.1-flash
```

Operators can also set `GOCLAW_ATLASCLOUD_API_KEY` (and optionally `GOCLAW_ATLASCLOUD_API_BASE`) on the gateway.

The API key is stored encrypted and is masked in every response.

## Models

Model IDs keep their vendor prefix, for example `zai-org/glm-5.3` or `moonshotai/kimi-k2.6`. The model picker shows Atlas Cloud’s public model list, which Atlas returns without checking your key, so a filled picker does not mean the key works. Only a model check with the key, such as **Verify** on the Customer Control Plane Providers page or `dewee providers verify --model`, proves the key works. You can still type any model ID Atlas Cloud publishes.

## What works today

* Chat completions, streaming responses, tool calls on models that support them, and token usage reporting.
* Using Atlas Cloud models as agent defaults, when switching models, and as fallbacks.

## Current limits

* Requests use the OpenAI Chat Completions protocol. A model that Atlas Cloud serves only through its Responses, Messages or Gemini-format APIs returns the upstream error.
* Tool calling, reasoning and vision depend on the model you pick. Dewee does not check those capabilities for each Atlas Cloud model.
* The model picker hides models Atlas Cloud marks as not ready. A listed model can still fail if Atlas Cloud stops serving it.
* Atlas Cloud has no embeddings API, so it cannot be the memory embedding provider.
* Atlas Cloud image, video and audio generation runs on a separate asynchronous API that Dewee does not support yet.